{"id":3429,"date":"2024-07-30T12:26:19","date_gmt":"2024-07-30T04:26:19","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=3429"},"modified":"2024-08-28T21:29:53","modified_gmt":"2024-08-28T13:29:53","slug":"llama-3yigeduoa","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/llama-3yigeduoa\/","title":{"rendered":"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217"},"content":{"rendered":"<p><strong>\u6458\u8981:<\/strong><\/p>\n<p>\u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u7cfb\u5217\u65b0\u7684\u57fa\u7840\u6a21\u578b\uff0c\u79f0\u4e3a Llama 3\u3002Llama 3 \u662f\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b\u7fa4\u4f53\uff0c\u5929\u751f\u652f\u6301\u591a\u8bed\u8a00\u3001\u4ee3\u7801\u7f16\u5199\u3001\u63a8\u7406\u548c\u5de5\u5177\u4f7f\u7528\u3002\u6211\u4eec\u6700\u5927\u7684\u6a21\u578b\u662f\u4e00\u4e2a\u5177\u6709 4050 \u4ebf\u4e2a\u53c2\u6570\u548c\u9ad8\u8fbe 128,000 \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u5bc6\u96c6\u578b Transformer\u3002\u672c\u6587\u5bf9 Llama 3 \u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u5e7f\u6cdb\u7684\u7ecf\u9a8c\u8bc4\u4f30\u3002\u7ed3\u679c\u8868\u660e\uff0cLlama 3 \u5728\u8bb8\u591a\u4efb\u52a1\u4e0a\u90fd\u80fd\u591f\u8fbe\u5230\u4e0e GPT-4 \u7b49\u9886\u5148\u8bed\u8a00\u6a21\u578b\u76f8\u5f53\u7684\u8d28\u91cf\u3002\u6211\u4eec\u5c06 Llama 3 \u516c\u5f00\u53d1\u5e03\uff0c\u5305\u62ec\u9884\u8bad\u7ec3\u548c\u540e\u8bad\u7ec3\u7684 4050 \u4ebf\u53c2\u6570\u8bed\u8a00\u6a21\u578b\uff0c\u4ee5\u53ca\u7528\u4e8e\u8f93\u5165\u8f93\u51fa\u5b89\u5168\u6027\u7684 Llama Guard 3 \u6a21\u578b\u3002\u672c\u6587\u8fd8\u4ecb\u7ecd\u4e86\u5c06\u56fe\u50cf\u3001\u89c6\u9891\u548c\u8bed\u97f3\u529f\u80fd\u901a\u8fc7\u7ec4\u5408\u5f0f\u65b9\u6cd5\u96c6\u6210\u5230 Llama 3 \u4e2d\u7684\u5b9e\u9a8c\u7ed3\u679c\u3002\u6211\u4eec\u89c2\u5bdf\u5230\u8fd9\u79cd\u65b9\u6cd5\u5728\u56fe\u50cf\u3001\u89c6\u9891\u548c\u8bed\u97f3\u8bc6\u522b\u4efb\u52a1\u4e0a\u4e0e\u6700\u5148\u8fdb\u7684\u65b9\u6cd5\u5177\u6709\u7ade\u4e89\u529b\u3002\u7531\u4e8e\u8fd9\u4e9b\u6a21\u578b\u4ecd\u5904\u4e8e\u5f00\u53d1\u9636\u6bb5\uff0c\u56e0\u6b64\u5c1a\u672a\u5e7f\u6cdb\u53d1\u5e03\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u5168\u6587\u4e0b\u8f7dpdf\uff1a<\/strong><br \/>\n<div class=\"huoduan_hide_box\" style=\"border:1px dashed #F60; padding:10px; margin:10px 0; line-height:200%; color:#F00; background-color:#FFF4FF; overflow:hidden; clear:both;\"><img loading=\"lazy\" decoding=\"async\" class=\"wxpic\" align=\"right\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/d8668ed8023fbe2.jpg\" style=\"width:150px;height:150px;margin-left:20px;display:inline;border:none\" width=\"150\" height=\"150\"  alt=\"AI\u751f\u4ea7\u529b\u5e94\u7528\" \/><span style=\"font-size:18px;\">\u6b64\u5904\u5185\u5bb9\u5df2\u7ecf\u88ab\u4f5c\u8005\u9690\u85cf\uff0c\u8bf7\u8f93\u5165\u9a8c\u8bc1\u7801\u67e5\u770b\u5185\u5bb9<\/span><form method=\"post\" style=\"margin:10px 0;\"><span class=\"yzts\" style=\"font-size:18px;float:left;\">\u9a8c\u8bc1\u7801\uff1a<\/span><input name=\"huoduan_verifycode\" id=\"verifycode\" type=\"text\" value=\"\" style=\"border:none;float:left;width:80px; height:32px; line-height:30px; padding:0 5px; border:1px solid #FF6600;-moz-border-radius: 0px;  -webkit-border-radius: 0px;  border-radius:0px;\" \/><input id=\"verifybtn\" style=\"border:none;float:left;width:80px; height:32px; line-height:32px; padding:0 5px; background-color:#F60; text-align:center; border:none; cursor:pointer; color:#FFF;-moz-border-radius: 0px; font-size:14px;  -webkit-border-radius: 0px;  border-radius:0px;\" name=\"\" type=\"submit\" value=\"\u63d0\u4ea4\u67e5\u770b\" \/><\/form><div style=\"clear:left;\"><\/div><span style=\"color:#00BF30\">\u8bf7\u5173\u6ce8\u672c\u7ad9\u5fae\u4fe1\u516c\u4f17\u53f7\uff0c\u56de\u590d\u201c<span style=\"color:blue\">\u9a8c\u8bc1\u7801<\/span>\u201d\uff0c\u83b7\u53d6\u9a8c\u8bc1\u7801\u3002\u5728\u5fae\u4fe1\u91cc\u641c\u7d22\u201c<span style=\"color:blue\">AI\u751f\u4ea7\u529b\u5e94\u7528<\/span>\u201d\u6216\u8005\u201c<span style=\"color:blue\">Artificial9527<\/span>\u201d\u6216\u8005\u5fae\u4fe1\u626b\u63cf\u53f3\u4fa7\u4e8c\u7ef4\u7801\u90fd\u53ef\u4ee5\u5173\u6ce8\u672c\u7ad9\u5fae\u4fe1\u516c\u4f17\u53f7\u3002<\/span><div class=\"cl\"><\/div><\/div><\/p>\n<div>\n<div class=\"markdown-body\">\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>1 \u5f15\u8a00<\/h2>\n<p>&nbsp;<\/p>\n<p><strong>\u57fa\u7840\u6a21\u578b<\/strong>\u662f\u8bed\u8a00\u3001\u89c6\u89c9\u3001\u8bed\u97f3\u548c\u5176\u4ed6\u6a21\u6001\u7684\u4e00\u822c\u6027\u6a21\u578b\uff0c\u65e8\u5728\u652f\u6301\u5e7f\u6cdb\u7684 AI \u4efb\u52a1\u3002\u5b83\u4eec\u6784\u6210\u4e86\u8bb8\u591a\u73b0\u4ee3 AI \u7cfb\u7edf\u7684\u57fa\u7840\u3002<\/p>\n<p>\u73b0\u4ee3\u57fa\u7840\u6a21\u578b\u7684\u5f00\u53d1\u4e3b\u8981\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a<\/p>\n<p><strong>(1) \u9884\u8bad\u7ec3\u9636\u6bb5:<\/strong>\u00a0\u6a21\u578b\u5728\u6d77\u91cf\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f7f\u7528\u7b80\u5355\u4efb\u52a1\u5982\u5355\u8bcd\u9884\u6d4b\u6216\u56fe\u6ce8\u751f\u6210\uff1b<\/p>\n<p><strong>(2) \u540e\u8bad\u7ec3\u9636\u6bb5:<\/strong>\u00a0\u6a21\u578b\u88ab\u5fae\u8c03\u4ee5\u9075\u5faa\u6307\u4ee4\u3001\u4e0e\u4eba\u7c7b\u504f\u597d\u4fdd\u6301\u4e00\u81f4\uff0c\u5e76\u63d0\u9ad8\u7279\u5b9a\u80fd\u529b\uff08\u4f8b\u5982\uff0c\u7f16\u7801\u548c\u63a8\u7406\uff09\u3002<\/p>\n<p>\u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u5957\u65b0\u7684\u8bed\u8a00\u57fa\u7840\u6a21\u578b\uff0c\u79f0\u4e3a Llama 3\u3002Llama 3 Herd \u6a21\u578b\u7cfb\u5217\u5929\u751f\u652f\u6301\u591a\u8bed\u8a00\u3001\u7f16\u7801\u3001\u63a8\u7406\u548c\u5de5\u5177\u4f7f\u7528\u3002\u6211\u4eec\u6700\u5927\u7684\u6a21\u578b\u662f\u5177\u6709 405B \u53c2\u6570\u7684\u7a20\u5bc6 Transformer\uff0c\u80fd\u591f\u5728\u9ad8\u8fbe 128K \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e2d\u5904\u7406\u4fe1\u606f\u3002<\/p>\n<p>\u8868 1 \u5217\u51fa\u4e86\u7f8a\u7fa4\u4e2d\u7684\u6bcf\u4e2a\u6210\u5458\u3002\u672c\u6587\u4e2d\u5448\u73b0\u7684\u6240\u6709\u7ed3\u679c\u5747\u57fa\u4e8e Llama 3.1 \u6a21\u578b\uff08\u7b80\u79f0 Llama 3\uff09\u3002<\/p>\n<p>\u6211\u4eec\u8ba4\u4e3a\u5f00\u53d1\u9ad8\u8d28\u91cf\u57fa\u7840\u6a21\u578b\u7684\u4e09\u628a\u5173\u952e\u5229\u5668\u662f\u6570\u636e\u3001\u89c4\u6a21\u548c\u590d\u6742\u6027\u7ba1\u7406\u3002\u5728\u6211\u4eec\u7684\u5f00\u53d1\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u52aa\u529b\u4f18\u5316\u8fd9\u4e09\u4e2a\u65b9\u9762\uff1a<\/p>\n<ul>\n<li><strong>\u6570\u636e:<\/strong>\u00a0\u4e0e\u4e4b\u524d\u7684 Llama \u7248\u672c\uff08Touvron \u7b49\u4eba\uff0c2023a, b\uff09\u76f8\u6bd4\uff0c\u6211\u4eec\u5728\u9884\u8bad\u7ec3\u548c\u540e\u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684 Both the quantity and quality of data were improved. These improvements include the development of more careful pre-processing and curation pipelines for pre-training data and the development of more rigorous quality assurance and filtering approaches for post-training data. Llama 3 \u5728\u7ea6 15T \u591a\u8bed\u8a00\u6807\u8bb0\u7684\u8bed\u6599\u5e93\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u800c Llama 2 \u5219\u662f\u5728 1.8T \u6807\u8bb0\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u3002<\/li>\n<li><strong>\u89c4\u6a21\u3002<\/strong>\u00a0\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u6bd4\u4e4b\u524d Llama \u6a21\u578b\u66f4\u5927\u7684\u6a21\u578b\uff1a\u6211\u4eec\u7684\u65d7\u8230\u8bed\u8a00\u6a21\u578b\u4f7f\u7528\u4e86 3.8 \u00d7 10<sup>25<\/sup>\u00a0FLOPs \u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u6bd4 Llama 2 \u6700\u5927\u7684\u7248\u672c\u591a\u8fd1 50 \u500d\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5728 15.6T \u6587\u672c\u6807\u8bb0\u4e0a\u9884\u8bad\u7ec3\u4e86\u4e00\u4e2a\u62e5\u6709 405B \u53ef\u8bad\u7ec3\u53c2\u6570\u7684\u65d7\u8230\u6a21\u578b\u3002\u6b63\u5982\u9884\u671f\u7684\u90a3\u6837\uff0c<\/li>\n<\/ul>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3430\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a453af8672d0e88.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"818\" height=\"320\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a453af8672d0e88.png 931w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a453af8672d0e88-300x117.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a453af8672d0e88-768x300.png 768w\" sizes=\"auto, (max-width: 818px) 100vw, 818px\" \/><\/strong><\/p>\n<ul>\n<li><strong>\u7ba1\u7406\u590d\u6742\u6027\u3002<\/strong>\u00a0\u6211\u4eec\u505a\u51fa\u7684\u8bbe\u8ba1\u9009\u62e9\u65e8\u5728\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u6a21\u578b\u5f00\u53d1\u8fc7\u7a0b\u7684\u53ef\u6269\u5c55\u6027\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u9009\u62e9\u4e86\u6807\u51c6\u5bc6\u96c6 <a href=\"https:\/\/www.kdjingpai.com\/pt\/transformer\/\">Transformer<\/a> \u6a21\u578b\u67b6\u6784\uff08Vaswani \u7b49\u4eba\uff0c2017 \u5e74\uff09\u5e76\u8fdb\u884c\u4e86\u4e00\u4e9b\u5fae\u5c0f\u7684\u8c03\u6574\uff0c\u800c\u4e0d\u662f\u91c7\u7528\u4e13\u5bb6\u6df7\u5408\u6a21\u578b\uff08Shazeer \u7b49\u4eba\uff0c2017 \u5e74\uff09\u4ee5\u6700\u5927\u7a0b\u5ea6\u5730\u63d0\u9ad8\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002\u540c\u6837\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u4e00\u79cd\u76f8\u5bf9\u7b80\u5355\u7684\u57fa\u4e8e\u76d1\u7763\u5fae\u8c03 (SFT)\u3001\u62d2\u7edd\u91c7\u6837 (RS) \u548c\u76f4\u63a5\u504f\u597d\u4f18\u5316 (DPO\uff1bRafailov \u7b49\u4eba (2023)) \u7684\u540e\u5904\u7406\u7a0b\u5e8f\uff0c\u800c\u4e0d\u662f\u66f4\u590d\u6742\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff08Ouyang \u7b49\u4eba\uff0c2022 \u5e74\uff1bSchulman \u7b49\u4eba\uff0c2017 \u5e74\uff09\uff0c\u8fd9\u4e9b\u7b97\u6cd5\u5f80\u5f80\u4e0d\u592a\u7a33\u5b9a\u5e76\u4e14\u96be\u4ee5\u6269\u5c55\u3002<\/li>\n<\/ul>\n<p>\u6211\u4eec\u8fd9\u9879\u5de5\u4f5c\u7684\u6210\u679c\u662f Llama 3\uff1a\u4e00\u4e2a\u5305\u542b 8B\u300170B \u548c 405B \u53c2\u6570\u7684\u4e09\u79cd\u591a\u8bed\u8a00<sup>1<\/sup>\u8bed\u8a00\u6a21\u578b\u7684\u7fa4\u4f53\u3002\u6211\u4eec\u5728\u5927\u91cf\u57fa\u51c6\u6570\u636e\u96c6\u4e0a\u8bc4\u4f30\u4e86 Llama 3 \u7684\u6027\u80fd\uff0c\u8fd9\u4e9b\u6570\u636e\u96c6\u6db5\u76d6\u4e86\u5e7f\u6cdb\u7684\u8bed\u8a00\u7406\u89e3\u4efb\u52a1\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u4eba\u5de5\u8bc4\u4f30\uff0c\u5c06 Llama 3 \u4e0e\u7ade\u4e89\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u8868 2 \u5c55\u793a\u4e86\u65d7\u8230 Llama 3 \u6a21\u578b\u5728\u5173\u952e\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6027\u80fd\u6982\u8ff0\u3002\u6211\u4eec\u7684\u5b9e\u9a8c\u8bc4\u4f30\u8868\u660e\uff0c\u6211\u4eec\u7684\u65d7\u8230\u6a21\u578b\u5728\u5404\u79cd\u4efb\u52a1\u4e0a\u90fd\u4e0e GPT-4\uff08OpenAI\uff0c2023a\uff09\u7b49\u9886\u5148\u8bed\u8a00\u6a21\u578b\u6301\u5e73\uff0c\u5e76\u4e14\u63a5\u8fd1\u4e8e\u6700\u5148\u8fdb\u6c34\u5e73\u3002\u6211\u4eec\u8f83\u5c0f\u7684\u6a21\u578b\u662f\u540c\u7c7b\u4e2d\u6700\u597d\u7684\uff0c\u5176\u6027\u80fd\u4f18\u4e8e\u5177\u6709\u76f8\u4f3c\u53c2\u6570\u6570\u91cf\u7684\u5176\u4ed6\u6a21\u578b\uff08Bai \u7b49\u4eba\uff0c2023 \u5e74\uff1bJiang \u7b49\u4eba\uff0c2023 \u5e74\uff09\u3002Llama 3 \u5728\u5e2e\u52a9\u6027\u548c\u65e0\u5bb3\u6027\u4e4b\u95f4\u4e5f\u53d6\u5f97\u4e86\u6bd4\u5176\u524d\u8eab\uff08Touvron \u7b49\u4eba\uff0c2023b\uff09\u66f4\u597d\u7684\u5e73\u8861\u3002\u6211\u4eec\u5728\u7b2c 5.4 \u8282\u4e2d\u8be6\u7ec6\u5206\u6790\u4e86 Llama 3 \u7684\u5b89\u5168\u6027\u3002<\/p>\n<p>\u6211\u4eec\u5c06\u6240\u6709\u4e09\u4e2a Llama 3 \u6a21\u578b\u516c\u5f00\u53d1\u5e03\uff0c\u4f7f\u7528 Llama 3 \u793e\u533a\u8bb8\u53ef\u8bc1\u7684\u66f4\u65b0\u7248\u672c\uff1b\u8bf7\u53c2\u9605 https:\/\/llama.meta.com\u3002\u8fd9\u5305\u62ec\u6211\u4eec 405B \u53c2\u6570\u8bed\u8a00\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u548c\u540e\u5904\u7406\u7248\u672c\uff0c\u4ee5\u53ca\u4e00\u4e2a\u65b0\u7684 Llama Guard \u6a21\u578b\u7248\u672c\uff08Inan \u7b49\u4eba\uff0c2023 \u5e74\uff09\uff0c\u7528\u4e8e\u8f93\u5165\u548c\u8f93\u51fa\u5b89\u5168\u6027\u3002\u6211\u4eec\u5e0c\u671b\u516c\u5f00\u53d1\u5e03\u4e00\u4e2a\u65d7\u8230\u6a21\u578b\u80fd\u591f\u6fc0\u53d1\u7814\u7a76\u754c\u7684\u521b\u65b0\u6d6a\u6f6e\uff0c\u5e76\u52a0\u901f\u671d\u7740\u4eba\u5de5\u667a\u80fd (AGI) \u7684\u8d1f\u8d23\u4efb\u53d1\u5c55\u65b9\u5411\u524d\u8fdb\u3002<\/p>\n<blockquote><p>\u591a\u8bed\u8a00\uff1a\u6307\u7684\u662f\u6a21\u578b\u80fd\u591f\u7406\u89e3\u548c\u751f\u6210\u591a\u79cd\u8bed\u8a00\u7684\u6587\u672c\u3002<\/p><\/blockquote>\n<\/div>\n<\/div>\n<p>&nbsp;<\/p>\n<p>\u5728 Llama 3 \u7684\u5f00\u53d1\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u8fd8\u5f00\u53d1\u4e86\u6a21\u578b\u7684\u591a\u6a21\u6001\u6269\u5c55\uff0c\u4f7f\u5176\u80fd\u591f\u5177\u5907\u56fe\u50cf\u8bc6\u522b\u3001\u89c6\u9891\u8bc6\u522b\u548c\u8bed\u97f3\u7406\u89e3\u80fd\u529b\u3002\u8fd9\u4e9b\u6a21\u578b\u4ecd\u5904\u4e8e\u79ef\u6781\u5f00\u53d1\u9636\u6bb5\uff0c\u5c1a\u672a\u51c6\u5907\u597d\u53d1\u5e03\u3002\u9664\u4e86\u6211\u4eec\u7684\u8bed\u8a00\u5efa\u6a21\u7ed3\u679c\u4e4b\u5916\uff0c\u672c\u6587\u8fd8\u4ecb\u7ecd\u4e86\u6211\u4eec\u5bf9\u8fd9\u4e9b\u591a\u6a21\u6001\u6a21\u578b\u7684\u521d\u6b65\u5b9e\u9a8c\u7ed3\u679c\u3002<\/p>\n<blockquote><p>Llama 3 8B \u548c 70B \u662f\u5728\u591a\u8bed\u8a00\u6570\u636e\u4e0a\u9884\u8bad\u7ec3\u7684\uff0c\u4f46\u5f53\u65f6\u4e3b\u8981\u7528\u4e8e\u82f1\u8bed\u3002<\/p><\/blockquote>\n<div>\n<div class=\"markdown-body\">\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3431\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/968b5605ffe1a9e.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"854\" height=\"505\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/968b5605ffe1a9e.png 1004w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/968b5605ffe1a9e-300x177.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/968b5605ffe1a9e-768x454.png 768w\" sizes=\"auto, (max-width: 854px) 100vw, 854px\" \/><\/p>\n<\/div>\n<\/div>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>2 \u603b\u8ff0<\/h2>\n<p>Llama 3 \u6a21\u578b\u67b6\u6784\u5982\u56fe 1 \u6240\u793a\u3002\u6211\u4eec Llama 3 \u8bed\u8a00\u6a21\u578b\u7684\u5f00\u53d1\u4e3b\u8981\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a<\/p>\n<ul>\n<li><strong>\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u3002<\/strong>\u6211\u4eec\u9996\u5148\u5c06\u5927\u578b\u591a\u8bed\u8a00\u6587\u672c\u8bed\u6599\u8f6c\u6362\u4e3a\u79bb\u6563\u6807\u8bb0\uff0c\u5e76\u5728\u7531\u6b64\u4ea7\u751f\u7684\u6570\u636e\u4e0a\u9884\u8bad\u7ec3\u4e00\u4e2a\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM) \u8fdb\u884c\u4e0b\u4e00\u4e2a\u6807\u8bb0\u9884\u6d4b\u3002\u5728\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u6a21\u578b\u5b66\u4e60\u8bed\u8a00\u7ed3\u6784\u5e76\u4ece\u5b83\u201c\u9605\u8bfb\u201d\u7684\u6587\u672c\u4e2d\u83b7\u53d6\u5927\u91cf\u5173\u4e8e\u4e16\u754c\u7684\u77e5\u8bc6\u3002\u4e3a\u4e86\u6709\u6548\u5730\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u9884\u8bad\u7ec3\u662f\u5728\u5927\u89c4\u6a21\u4e0b\u8fdb\u884c\u7684\uff1a\u6211\u4eec\u4f7f\u7528 8K \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u5728\u4e00\u4e2a\u5305\u542b 15.6T \u4e2a\u6807\u8bb0\u7684\u6a21\u578b\u4e0a\u9884\u8bad\u7ec3\u4e86\u4e00\u4e2a\u5177\u6709 405B \u4e2a\u53c2\u6570\u7684\u6a21\u578b\u3002\u6b64\u6807\u51c6\u9884\u8bad\u7ec3\u9636\u6bb5\u4e4b\u540e\u662f\u7ee7\u7eed\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u8be5\u9636\u6bb5\u5c06\u652f\u6301\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u589e\u52a0\u5230 128K \u4e2a\u6807\u8bb0\u3002\u6709\u5173\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u89c1\u7b2c 3 \u8282\u3002<\/li>\n<li><strong>\u6a21\u578b\u540e\u8bad\u7ec3\u3002<\/strong>\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5bf9\u8bed\u8a00\u6709\u4e30\u5bcc\u7684\u7406\u89e3\uff0c\u4f46\u5b83\u5c1a\u672a\u9075\u5faa\u6307\u4ee4\u6216\u50cf\u6211\u4eec\u671f\u671b\u7684\u52a9\u624b\u90a3\u6837\u884c\u4e3a\u3002\u6211\u4eec\u5728\u51e0\u8f6e\u4e2d\u901a\u8fc7\u4eba\u7c7b\u53cd\u9988\u5bf9\u6a21\u578b\u8fdb\u884c\u6821\u51c6\uff0c\u6bcf\u8f6e\u90fd\u5305\u62ec\u5728\u6307\u4ee4\u8c03\u6574\u6570\u636e\u4e0a\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff1bRafailov et al., 2024\uff09\u3002\u5728\u8fd9\u4e2a\u540e\u8bad\u7ec3\u9636\u6bb5\uff0c\u6211\u4eec\u8fd8\u96c6\u6210\u4e86\u65b0\u7684\u529f\u80fd\uff0c\u4f8b\u5982\u5de5\u5177\u4f7f\u7528\uff0c\u5e76\u5728\u7f16\u7801\u548c\u63a8\u7406\u7b49\u65b9\u9762\u89c2\u5bdf\u5230\u663e\u8457\u6539\u8fdb\u3002\u6709\u5173\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u9605\u7b2c 4 \u8282\u3002\u6700\u540e\uff0c\u5b89\u5168\u7f13\u89e3\u63aa\u65bd\u4e5f\u5728\u540e\u8bad\u7ec3\u9636\u6bb5\u96c6\u6210\u5230\u6a21\u578b\u4e2d\uff0c\u5176\u7ec6\u8282\u5728\u7b2c 5.4\u8282\u4e2d\u63cf\u8ff0\u3002\u751f\u6210\u7684\u6a21\u578b\u5177\u5907\u4e30\u5bcc\u7684\u529f\u80fd\u3002\u5b83\u4eec\u80fd\u591f\u7528\u81f3\u5c11\u516b\u79cd\u8bed\u8a00\u56de\u7b54\u95ee\u9898\uff0c\u7f16\u5199\u9ad8\u8d28\u91cf\u4ee3\u7801\uff0c\u89e3\u51b3\u590d\u6742\u7684\u63a8\u7406\u95ee\u9898\uff0c\u5e76\u5f00\u7bb1\u5373\u7528\u6216\u96f6\u6837\u672c\u65b9\u5f0f\u4f7f\u7528\u5de5\u5177\u3002<\/li>\n<\/ul>\n<p>\u6211\u4eec\u8fd8\u8fdb\u884c\u5b9e\u9a8c\uff0c\u901a\u8fc7\u7ec4\u5408\u5f0f\u65b9\u6cd5\u5c06\u56fe\u50cf\u3001\u89c6\u9891\u548c\u8bed\u97f3\u80fd\u529b\u6dfb\u52a0\u5230 Llama 3 \u4e2d\u3002\u6211\u4eec\u7814\u7a76\u7684\u65b9\u6cd5\u5305\u62ec\u56fe 28 \u6240\u793a\u7684\u4e09\u4e2a\u989d\u5916\u9636\u6bb5\uff1a<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-3432\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/e9d76e374f68496.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"1166\" height=\"310\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/e9d76e374f68496.png 1166w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/e9d76e374f68496-300x80.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/e9d76e374f68496-1024x272.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/e9d76e374f68496-768x204.png 768w\" sizes=\"auto, (max-width: 1166px) 100vw, 1166px\" \/><\/p>\n<ul>\n<li><strong>\u591a\u6a21\u6001\u7f16\u7801\u5668\u9884\u8bad\u7ec3\u3002<\/strong>\u6211\u4eec\u4e3a\u56fe\u50cf\u548c\u8bed\u97f3\u5206\u522b\u8bad\u7ec3\u7f16\u7801\u5668\u3002\u6211\u4eec\u5c06\u56fe\u50cf\u7f16\u7801\u5668\u8bad\u7ec3\u4e8e\u5927\u91cf\u56fe\u50cf\u6587\u672c\u5bf9\u4e0a\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u5b66\u4e60\u5230\u89c6\u89c9\u5185\u5bb9\u4e0e\u5176\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u6211\u4eec\u7684\u8bed\u97f3\u7f16\u7801\u5668\u4f7f\u7528\u4e00\u79cd\u81ea\u76d1\u7763\u65b9\u6cd5\uff0c\u5b83\u906e\u76d6\u8bed\u97f3\u8f93\u5165\u7684\u4e00\u90e8\u5206\uff0c\u5e76\u5c1d\u8bd5\u901a\u8fc7\u79bb\u6563\u6807\u8bb0\u8868\u793a\u91cd\u5efa\u906e\u76d6\u7684\u90e8\u5206\u3002\u56e0\u6b64\uff0c\u6a21\u578b\u5b66\u4e60\u4e86\u8bed\u97f3\u4fe1\u53f7\u7684\u7ed3\u6784\u3002\u6709\u5173\u56fe\u50cf\u7f16\u7801\u5668\u7684\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u9605\u7b2c 7 \u8282\uff1b\u6709\u5173\u8bed\u97f3\u7f16\u7801\u5668\u7684\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u9605\u7b2c 8 \u8282\u3002<\/li>\n<li><strong>\u89c6\u89c9\u9002\u914d\u5668\u8bad\u7ec3\u3002<\/strong>\u6211\u4eec\u8bad\u7ec3\u4e00\u4e2a\u9002\u914d\u5668\uff0c\u5c06\u9884\u8bad\u7ec3\u7684\u56fe\u50cf\u7f16\u7801\u5668\u4e0e\u9884\u8bad\u7ec3\u7684\u8bed\u8a00\u6a21\u578b\u96c6\u6210\u3002\u8be5\u9002\u914d\u5668\u7531\u4e00\u7cfb\u5217\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u7ec4\u6210\uff0c\u5c06\u56fe\u50cf\u7f16\u7801\u5668\u8868\u793a\u8f93\u5165\u5230\u8bed\u8a00\u6a21\u578b\u4e2d\u3002\u9002\u914d\u5668\u5728\u6587\u672c-\u56fe\u50cf\u5bf9\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u8fd9\u4f7f\u56fe\u50cf\u8868\u793a\u4e0e\u8bed\u8a00\u8868\u793a\u76f8\u4e00\u81f4\u3002\u5728\u9002\u914d\u5668\u8bad\u7ec3\u671f\u95f4\uff0c\u6211\u4eec\u4e5f\u66f4\u65b0\u4e86\u56fe\u50cf\u7f16\u7801\u5668\u7684\u53c2\u6570\uff0c\u4f46\u6709\u610f\u4e0d\u66f4\u65b0\u8bed\u8a00\u6a21\u578b\u7684\u53c2\u6570\u3002\u6211\u4eec\u8fd8\u5728\u56fe\u50cf\u9002\u914d\u5668\u4e4b\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u89c6\u9891\u9002\u914d\u5668\uff0c\u4f7f\u7528\u914d\u5bf9\u7684\u89c6\u9891-\u6587\u672c\u6570\u636e\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u8de8\u5e27\u805a\u5408\u4fe1\u606f\u3002\u6709\u5173\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u89c1\u7b2c 7 \u8282\u3002<\/li>\n<li>\u6700\u540e\uff0c\u6211\u4eec\u5c06\u8bed\u97f3\u7f16\u7801\u5668\u901a\u8fc7\u4e00\u4e2a\u9002\u914d\u5668\u96c6\u6210\u5230\u6a21\u578b\u4e2d\uff0c\u8be5\u9002\u914d\u5668\u5c06\u8bed\u97f3\u7f16\u7801\u8f6c\u6362\u4e3a\u53ef\u4ee5\u76f4\u63a5\u8f93\u5165\u5fae\u8c03\u8bed\u8a00\u6a21\u578b\u7684\u6807\u8bb0\u8868\u793a\u3002\u5728\u76d1\u7763\u5f0f\u5fae\u8c03\u9636\u6bb5\uff0c\u8054\u5408\u66f4\u65b0\u9002\u914d\u5668\u548c\u7f16\u7801\u5668\u7684\u53c2\u6570\uff0c\u4ee5\u5b9e\u73b0\u9ad8\u8d28\u91cf\u7684\u8bed\u97f3\u7406\u89e3\u3002\u6211\u4eec\u5728\u8bed\u97f3\u9002\u914d\u5668\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e0d\u6539\u53d8\u8bed\u8a00\u6a21\u578b\u3002\u6211\u4eec\u8fd8\u96c6\u6210\u4e86\u4e00\u4e2a\u6587\u672c\u5230\u8bed\u97f3\u7cfb\u7edf\u3002\u8be6\u60c5\u8bf7\u53c2\u89c1\u7b2c 8 \u8282\u3002<\/li>\n<\/ul>\n<p>\u6211\u4eec\u7684\u591a\u6a21\u6001\u5b9e\u9a8c\u5bfc\u81f4\u6a21\u578b\u80fd\u591f\u8bc6\u522b\u56fe\u50cf\u548c\u89c6\u9891\u7684\u5185\u5bb9\uff0c\u5e76\u901a\u8fc7\u8bed\u97f3\u63a5\u53e3\u652f\u6301\u4ea4\u4e92\u3002\u8fd9\u4e9b\u6a21\u578b\u4ecd\u5728\u5f00\u53d1\u4e2d\uff0c\u5c1a\u672a\u51c6\u5907\u597d\u53d1\u5e03\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>3 \u9884\u8bad\u7ec3<\/h2>\n<p>&nbsp;<\/p>\n<p>\u8bed\u8a00\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u6d89\u53ca\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\uff1a<\/p>\n<p>(1) \u6536\u96c6\u548c\u8fc7\u6ee4\u5927\u89c4\u6a21\u8bad\u7ec3\u8bed\u6599\u5e93\uff1b<\/p>\n<p>(2) \u5f00\u53d1\u6a21\u578b\u67b6\u6784\u548c\u76f8\u5e94\u7684\u7f29\u653e\u5b9a\u5f8b\u4ee5\u786e\u5b9a\u6a21\u578b\u5927\u5c0f\uff1b<\/p>\n<p>(3) \u5f00\u53d1\u9ad8\u6548\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u7684\u6280\u672f\uff1b<\/p>\n<p>(4) \u5f00\u53d1\u9884\u8bad\u7ec3\u65b9\u6848\u3002\u6211\u4eec\u5c06\u5728\u4e0b\u9762\u5206\u522b\u4ecb\u7ecd\u8fd9\u4e9b\u7ec4\u6210\u90e8\u5206\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3.1 \u9884\u8bad\u7ec3\u6570\u636e<\/h3>\n<p>\u6211\u4eec\u4ece\u5404\u79cd\u6570\u636e\u6e90\u521b\u5efa\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u6570\u636e\u96c6\uff0c\u8fd9\u4e9b\u6570\u636e\u6e90\u5305\u542b\u622a\u81f3 2023 \u5e74\u5e95\u7684\u77e5\u8bc6\u3002\u6211\u4eec\u5bf9\u6bcf\u4e2a\u6570\u636e\u6e90\u5e94\u7528\u4e86\u51e0\u79cd\u53bb\u91cd\u65b9\u6cd5\u548c\u6570\u636e\u6e05\u6d17\u673a\u5236\uff0c\u4ee5\u83b7\u5f97\u9ad8\u8d28\u91cf\u7684\u6807\u8bb0\u3002\u6211\u4eec\u5220\u9664\u5305\u542b\u5927\u91cf\u4e2a\u4eba\u8eab\u4efd\u4fe1\u606f (PII) \u7684\u9886\u57df\uff0c\u4ee5\u53ca\u5df2\u77e5\u542b\u6709\u6210\u4eba\u5185\u5bb9\u7684\u9886\u57df\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>3.11 Web \u6570\u636e\u6e05\u6d17<\/h4>\n<p>\u6211\u4eec\u5229\u7528\u7684\u5927\u90e8\u5206\u6570\u636e\u6765\u81ea\u7f51\u7edc\uff0c\u4e0b\u9762\u6211\u4eec\u5c06\u63cf\u8ff0\u6211\u4eec\u7684\u6e05\u6d17\u6d41\u7a0b\u3002<\/p>\n<p><strong>PII \u548c\u5b89\u5168\u8fc7\u6ee4:<\/strong>\u00a0\u9664\u4e86\u5176\u4ed6\u63aa\u65bd\u5916\uff0c\u6211\u4eec\u8fd8\u5b9e\u65bd\u4e86\u65e8\u5728\u53bb\u9664\u53ef\u80fd\u5305\u542b\u4e0d\u5b89\u5168\u5185\u5bb9\u6216\u5927\u91cf PII \u7684\u7f51\u7ad9\u6570\u636e\u7684\u8fc7\u6ee4\u5668\uff0c\u8fd9\u4e9b\u7f51\u7ad9\u7684\u57df\u540d\u6839\u636e\u5404\u79cd Meta \u5b89\u5168\u6807\u51c6\u88ab\u5217\u4e3a\u6709\u5bb3\u7f51\u7ad9\uff0c\u4ee5\u53ca\u5df2\u77e5\u5305\u542b\u6210\u4eba\u5185\u5bb9\u7684\u57df\u540d\u3002<\/p>\n<p><strong>\u6587\u672c\u63d0\u53d6\u548c\u6e05\u6d17:<\/strong>\u00a0\u6211\u4eec\u5904\u7406\u539f\u59cb HTML \u5185\u5bb9\u4ee5\u63d0\u53d6\u9ad8\u8d28\u91cf\u3001\u591a\u6837\u5316\u7684\u6587\u672c\uff0c\u5e76\u5c06\u975e\u622a\u65ad\u7684\u7f51\u9875\u6587\u6863\u7528\u4e8e\u6b64\u76ee\u7684\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u81ea\u5b9a\u4e49\u89e3\u6790\u5668\uff0c\u8be5\u89e3\u6790\u5668\u63d0\u53d6 HTML \u5185\u5bb9\u5e76\u4f18\u5316\u4e86\u6a21\u677f\u79fb\u9664\u548c\u5185\u5bb9\u53ec\u56de\u7684\u7cbe\u5ea6\u3002\u6211\u4eec\u901a\u8fc7\u4eba\u5de5\u8bc4\u4f30\u6765\u8bc4\u4f30\u89e3\u6790\u5668\u7684\u8d28\u91cf\uff0c\u5e76\u5c06\u5176\u4e0e\u9488\u5bf9\u7c7b\u4f3c\u6587\u7ae0\u7684\u5185\u5bb9\u8fdb\u884c\u4f18\u5316\u7684\u6d41\u884c\u7b2c\u4e09\u65b9 HTML \u89e3\u6790\u5668\u8fdb\u884c\u4e86\u6bd4\u8f83\uff0c\u53d1\u73b0\u5176\u8868\u73b0\u826f\u597d\u3002\u6211\u4eec\u4f1a\u8c28\u614e\u5904\u7406\u5305\u542b\u6570\u5b66\u548c\u4ee3\u7801\u5185\u5bb9\u7684 HTML \u9875\u9762\uff0c\u4ee5\u4fdd\u7559\u8fd9\u4e9b\u5185\u5bb9\u7684\u7ed3\u6784\u3002\u6211\u4eec\u4fdd\u7559\u56fe\u50cf alt \u5c5e\u6027\u6587\u672c\uff0c\u56e0\u4e3a\u6570\u5b66\u5185\u5bb9\u901a\u5e38\u4ee5\u9884\u6e32\u67d3\u56fe\u50cf\u7684\u5f62\u5f0f\u8868\u793a\uff0c\u5176\u4e2d\u6570\u5b66\u4e5f\u63d0\u4f9b\u5728 alt \u5c5e\u6027\u4e2d\u3002<\/p>\n<p>\u6211\u4eec\u53d1\u73b0\u4e0e\u7eaf\u6587\u672c\u76f8\u6bd4\uff0cMarkdown \u5bf9\u4e3b\u8981\u8bad\u7ec3\u4e8e Web \u6570\u636e\u7684\u6a21\u578b\u6027\u80fd\u6709\u5bb3\uff0c\u56e0\u6b64\u6211\u4eec\u5220\u9664\u6240\u6709 Markdown \u6807\u8bb0\u3002<\/p>\n<p><strong>\u53bb\u91cd:<\/strong>\u00a0\u6211\u4eec\u5728 URL\u3001\u6587\u6863\u548c\u884c\u7ea7\u522b\u5e94\u7528\u591a\u8f6e\u53bb\u91cd\uff1a<\/p>\n<ul>\n<li><strong>URL \u7ea7\u53bb\u91cd:<\/strong>\u00a0\u6211\u4eec\u5bf9\u6574\u4e2a\u6570\u636e\u96c6\u6267\u884c URL \u7ea7\u53bb\u91cd\u3002\u5bf9\u4e8e\u6bcf\u4e2a URL \u5bf9\u5e94\u7684\u9875\u9762\uff0c\u6211\u4eec\u4fdd\u7559\u6700\u65b0\u7248\u672c\u3002<\/li>\n<li><strong>\u6587\u6863\u7ea7\u53bb\u91cd:<\/strong>\u00a0\u6211\u4eec\u5bf9\u6574\u4e2a\u6570\u636e\u96c6\u6267\u884c\u5168\u5c40 MinHash (Broder, 1997) \u53bb\u91cd\u4ee5\u5220\u9664\u8fd1\u4f3c\u91cd\u590d\u6587\u6863\u3002<\/li>\n<li><strong>\u884c\u7ea7\u53bb\u91cd:<\/strong>\u00a0\u6211\u4eec\u6267\u884c\u7c7b\u4f3c\u4e8e ccNet (Wenzek \u7b49\u4eba\uff0c2019) \u7684\u6fc0\u8fdb\u884c\u7ea7\u53bb\u91cd\u3002\u6211\u4eec\u5220\u9664\u5728\u6bcf\u4e2a\u5305\u542b 3000 \u4e07\u4e2a\u6587\u6863\u7684\u7ec4\u4e2d\u51fa\u73b0\u8d85\u8fc7 6 \u6b21\u7684\u884c\u3002<\/li>\n<\/ul>\n<p>\u5c3d\u7ba1\u6211\u4eec\u7684\u624b\u52a8\u5b9a\u6027\u5206\u6790\u8868\u660e\u884c\u7ea7\u53bb\u91cd\u4e0d\u4ec5\u53ef\u4ee5\u53bb\u9664\u6765\u81ea\u5404\u79cd\u7f51\u7ad9\u7684\u5269\u4f59\u6a21\u677f\u5185\u5bb9\uff08\u4f8b\u5982\u5bfc\u822a\u83dc\u5355\u3001Cookie \u9884\u8b66\uff09\uff0c\u8fd8\u53ef\u4ee5\u53bb\u9664\u9891\u7e41\u7684\u9ad8\u8d28\u91cf\u6587\u672c\uff0c\u4f46\u6211\u4eec\u7684\u7ecf\u9a8c\u8bc4\u4f30\u663e\u793a\u51fa\u660e\u663e\u7684\u6539\u8fdb\u3002<\/p>\n<p><strong>\u542f\u53d1\u5f0f\u8fc7\u6ee4:<\/strong>\u00a0\u6211\u4eec\u5f00\u53d1\u4e86\u542f\u53d1\u5f0f\u65b9\u6cd5\u6765\u5220\u9664\u989d\u5916\u7684\u4f4e\u8d28\u91cf\u6587\u6863\u3001\u5f02\u5e38\u503c\u548c\u91cd\u590d\u6b21\u6570\u8fc7\u591a\u7684\u6587\u6863\u3002\u4e00\u4e9b\u542f\u53d1\u5f0f\u65b9\u6cd5\u7684\u793a\u4f8b\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u6211\u4eec\u4f7f\u7528\u91cd\u590d n \u5143\u7ec4\u8986\u76d6\u7387 (Rae et al., 2021) \u6765\u5220\u9664\u7531\u91cd\u590d\u5185\u5bb9\uff08\u4f8b\u5982\u65e5\u5fd7\u6216\u9519\u8bef\u6d88\u606f\uff09\u7ec4\u6210\u7684\u884c\u3002\u8fd9\u4e9b\u884c\u53ef\u80fd\u975e\u5e38\u957f\u4e14\u552f\u4e00\uff0c\u56e0\u6b64\u4e0d\u80fd\u901a\u8fc7\u884c\u53bb\u91cd\u8fdb\u884c\u8fc7\u6ee4\u3002<\/li>\n<li>\u6211\u4eec\u4f7f\u7528\u201c\u810f\u8bdd\u201d\u8ba1\u6570 (Raffel et al., 2020) \u6765\u8fc7\u6ee4\u6389\u672a\u88ab\u57df\u540d\u9ed1\u540d\u5355\u8986\u76d6\u7684\u6210\u4eba\u7f51\u7ad9\u3002<\/li>\n<li>\u6211\u4eec\u4f7f\u7528\u6807\u8bb0\u5206\u5e03\u7684 Kullback-Leibler \u6563\u5ea6\u6765\u8fc7\u6ee4\u6389\u5305\u542b\u4e0e\u8bad\u7ec3\u8bed\u6599\u5e93\u5206\u5e03\u76f8\u6bd4\u8fc7\u591a\u5f02\u5e38\u6807\u8bb0\u7684\u6587\u6863\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><strong>\u57fa\u4e8e\u6a21\u578b\u7684\u8d28\u91cf\u8fc7\u6ee4:<\/strong><\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5c1d\u8bd5\u5c06\u5404\u79cd\u57fa\u4e8e\u6a21\u578b\u7684\u8d28\u91cf\u5206\u7c7b\u5668\u7528\u4e8e\u9009\u62e9\u9ad8\u8d28\u91cf\u6807\u8bb0\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u4f7f\u7528 fasttext (Joulin et al., 2017) \u7b49\u5feb\u901f\u5206\u7c7b\u5668\uff0c\u8fd9\u4e9b\u5206\u7c7b\u5668\u7ecf\u8fc7\u8bad\u7ec3\u53ef\u4ee5\u8bc6\u522b\u7ed9\u5b9a\u7684\u6587\u672c\u662f\u5426\u4f1a\u88ab\u7ef4\u57fa\u767e\u79d1\u5f15\u7528 (Touvron et al., 2023a)\u3002<\/li>\n<li>\u4f7f\u7528\u66f4\u8017\u8d39\u8ba1\u7b97\u8d44\u6e90\u7684 Roberta \u6a21\u578b\u5206\u7c7b\u5668 (Liu et al., 2019a)\uff0c\u8be5\u5206\u7c7b\u5668\u5728 Llama 2 \u7684\u9884\u6d4b\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u8bad\u7ec3\u57fa\u4e8e Llama 2 \u7684\u8d28\u91cf\u5206\u7c7b\u5668\uff0c\u6211\u4eec\u521b\u5efa\u4e86\u4e00\u7ec4\u7ecf\u8fc7\u6e05\u6d17\u7684 Web \u6587\u6863\uff0c\u63cf\u8ff0\u4e86\u8d28\u91cf\u8981\u6c42\uff0c\u5e76\u6307\u793a Llama 2 \u7684\u804a\u5929\u6a21\u578b\u786e\u5b9a\u6587\u6863\u662f\u5426\u6ee1\u8db3\u8fd9\u4e9b\u8981\u6c42\u3002 \u4e3a\u63d0\u9ad8\u6548\u7387\uff0c\u6211\u4eec\u4f7f\u7528 DistilRoberta (Sanh et al., 2019) \u4e3a\u6bcf\u4e2a\u6587\u6863\u751f\u6210\u8d28\u91cf\u5206\u6570\u3002 \u6211\u4eec\u5c06\u5b9e\u9a8c\u8bc4\u4f30\u5404\u79cd\u8d28\u91cf\u8fc7\u6ee4\u914d\u7f6e\u7684\u6548\u679c\u3002<\/p>\n<p><strong>\u4ee3\u7801\u548c\u63a8\u7406\u6570\u636e:<\/strong><\/p>\n<p>\u7c7b\u4f3c\u4e8e DeepSeek-AI \u7b49\u4eba (2024)\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u9886\u57df\u7279\u5b9a\u7684\u7ba1\u9053\u6765\u63d0\u53d6\u5305\u542b\u4ee3\u7801\u548c\u6570\u5b66\u76f8\u5173\u7f51\u9875\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u8fd9\u4e24\u4e2a\u4ee3\u7801\u548c\u63a8\u7406\u5206\u7c7b\u5668\u90fd\u662f\u4f7f\u7528 Llama 2 \u6ce8\u89e3\u7684 Web \u6570\u636e\u8bad\u7ec3\u7684 DistilledRoberta \u6a21\u578b\u3002 \u4e0e\u4e0a\u9762\u63d0\u5230\u7684\u901a\u7528\u8d28\u91cf\u5206\u7c7b\u5668\u4e0d\u540c\uff0c\u6211\u4eec\u8fdb\u884c\u4e86\u63d0\u793a\u8c03\u6574\u4ee5\u9488\u5bf9\u5305\u542b\u6570\u5b66\u63a8\u8bba\u3001STEM \u9886\u57df\u7684\u63a8\u7406\u4ee5\u53ca\u5d4c\u5165\u81ea\u7136\u8bed\u8a00\u7684\u4ee3\u7801\u7684\u7f51\u9875\u3002\u7531\u4e8e\u4ee3\u7801\u548c\u6570\u5b66\u7684\u6807\u8bb0\u5206\u5e03\u4e0e\u81ea\u7136\u8bed\u8a00\u7684\u6807\u8bb0\u5206\u5e03\u622a\u7136\u4e0d\u540c\uff0c\u56e0\u6b64\u8fd9\u4e9b\u7ba1\u9053\u5b9e\u73b0\u4e86\u9886\u57df\u7279\u5b9a\u7684 HTML \u63d0\u53d6\u3001\u81ea\u5b9a\u4e49\u6587\u672c\u7279\u5f81\u548c\u7528\u4e8e\u8fc7\u6ee4\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\u3002<\/p>\n<p><strong>\u591a\u8bed\u8a00\u6570\u636e:<\/strong><\/p>\n<p>\u4e0e\u4e0a\u9762\u63cf\u8ff0\u7684\u82f1\u6587\u5904\u7406\u7ba1\u9053\u7c7b\u4f3c\uff0c\u6211\u4eec\u5b9e\u65bd\u8fc7\u6ee4\u5668\u4ee5\u5220\u9664\u53ef\u80fd\u5305\u542b\u4e2a\u4eba\u8eab\u4efd\u4fe1\u606f (PII) \u6216\u4e0d\u5b89\u5168\u5185\u5bb9\u7684\u7f51\u7ad9\u6570\u636e\u3002\u6211\u4eec\u7684\u591a\u8bed\u8a00\u6587\u672c\u5904\u7406\u7ba1\u9053\u5177\u6709\u4ee5\u4e0b\u72ec\u7279\u529f\u80fd\uff1a<\/p>\n<ul>\n<li>\u6211\u4eec\u4f7f\u7528\u57fa\u4e8e fasttext \u7684\u8bed\u8a00\u8bc6\u522b\u6a21\u578b\u5c06\u6587\u6863\u5206\u7c7b\u4e3a 176 \u79cd\u8bed\u8a00\u3002<\/li>\n<li>\u6211\u4eec\u5bf9\u6bcf\u4e2a\u8bed\u8a00\u7684\u6570\u636e\u8fdb\u884c\u6587\u6863\u7ea7\u548c\u884c\u7ea7\u53bb\u91cd\u3002<\/li>\n<li>\u6211\u4eec\u5e94\u7528\u8bed\u8a00\u7279\u5b9a\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\u548c\u57fa\u4e8e\u6a21\u578b\u7684\u8fc7\u6ee4\u5668\u6765\u53bb\u9664\u4f4e\u8d28\u91cf\u6587\u6863\u3002<\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u4f7f\u7528\u57fa\u4e8e\u591a\u8bed\u8a00 Llama 2 \u7684\u5206\u7c7b\u5668\u5bf9\u591a\u8bed\u8a00\u6587\u6863\u8fdb\u884c\u8d28\u91cf\u6392\u540d\uff0c\u4ee5\u786e\u4fdd\u4f18\u5148\u5904\u7406\u9ad8\u8d28\u91cf\u5185\u5bb9\u3002\u6211\u4eec\u5728\u9884\u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684\u591a\u8bed\u8a00\u6807\u8bb0\u6570\u91cf\u901a\u8fc7\u5b9e\u9a8c\u786e\u5b9a\uff0c\u5728\u82f1\u8bed\u548c\u591a\u8bed\u8a00\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u6a21\u578b\u6027\u80fd\u53d6\u5f97\u5e73\u8861\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>3.12 \u786e\u5b9a\u6570\u636e\u6df7\u5408<\/h4>\n<p id=\"pageNum-24a10669-b2b4-46b7-a239-49a3fcf3efc4\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\">\u4e3a\u4e86\u83b7\u5f97\u9ad8\u8d28\u91cf\u8bed\u8a00\u6a21\u578b\uff0c\u5fc5\u987b\u8c28\u614e\u786e\u5b9a\u9884\u8bad\u7ec3\u6570\u636e\u6df7\u5408\u4e2d\u4e0d\u540c\u6570\u636e\u6e90\u7684\u6bd4\u4f8b\u3002\u6211\u4eec\u4e3b\u8981\u5229\u7528\u77e5\u8bc6\u5206\u7c7b\u548c\u5c3a\u5ea6\u5b9a\u5f8b\u5b9e\u9a8c\u6765\u786e\u5b9a\u8fd9\u4e00\u6570\u636e\u6df7\u5408\u3002<\/span><\/span><\/p>\n<p id=\"pageNum-094759d7-412f-4575-850d-84987fa2140f\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-wrapper\" lang=\"zh-CN\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\"><strong>\u77e5\u8bc6\u5206\u7c7b\u3002<\/strong>\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u5206\u7c7b\u5668\uff0c\u7528\u4e8e\u5bf9\u7f51\u9875\u6570\u636e\u4e2d\u5305\u542b\u7684\u4fe1\u606f\u7c7b\u578b\u8fdb\u884c\u5206\u7c7b\uff0c\u4ee5\u4fbf\u66f4\u6709\u6548\u5730\u786e\u5b9a\u6570\u636e\u7ec4\u5408\u3002\u6211\u4eec\u4f7f\u7528\u8fd9\u4e2a\u5206\u7c7b\u5668\u5bf9\u7f51\u9875\u4e0a\u8fc7\u5ea6\u4ee3\u8868\u7684\u6570\u636e\u7c7b\u522b\uff08\u4f8b\u5982\u827a\u672f\u548c\u5a31\u4e50\uff09\u8fdb\u884c\u4e0b\u91c7\u6837\u3002<\/span><\/span><\/span><\/p>\n<p id=\"pageNum-3f320fdd-15aa-4e6c-a708-68dcb7cd8b2e\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\"><strong>\u4e3a\u4e86\u786e\u5b9a\u6700\u4f73\u6570\u636e\u6df7\u5408\u65b9\u6848\u3002<\/strong>\u6211\u4eec\u8fdb\u884c\u89c4\u6a21\u5b9a\u5f8b\u5b9e\u9a8c\uff0c\u5176\u4e2d\u6211\u4eec\u5c06\u591a\u4e2a\u5c0f\u578b\u6a21\u578b\u8bad\u7ec3\u4e8e\u7279\u5b9a\u6570\u636e\u6df7\u5408\u96c6\u4e0a\uff0c\u5e76\u5229\u7528\u5176\u9884\u6d4b\u5927\u578b\u6a21\u578b\u5728\u8be5\u6df7\u5408\u96c6\u4e0a\u7684\u6027\u80fd\uff08\u53c2\u89c1\u7b2c 3.2.1 \u8282\uff09\u3002\u6211\u4eec\u591a\u6b21\u91cd\u590d\u6b64\u8fc7\u7a0b\uff0c\u9488\u5bf9\u4e0d\u540c\u7684\u6570\u636e\u6df7\u5408\u96c6\u9009\u62e9\u65b0\u7684\u5019\u9009\u6570\u636e\u6df7\u5408\u96c6\u3002\u968f\u540e\uff0c\u6211\u4eec\u5728\u8be5\u5019\u9009\u6570\u636e\u6df7\u5408\u96c6\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u66f4\u5927\u7684\u6a21\u578b\uff0c\u5e76\u5728\u591a\u4e2a\u5173\u952e\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u8bc4\u4f30\u8be5\u6a21\u578b\u7684\u6027\u80fd\u3002<\/span><\/span><\/p>\n<p id=\"pageNum-8591a281-3504-4920-812e-b55061d28681\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-wrapper\" lang=\"zh-CN\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\"><strong>\u6570\u636e\u6df7\u5408\u6458\u8981\u3002<\/strong>\u6211\u4eec\u7684\u6700\u7ec8\u6570\u636e\u6df7\u5408\u5305\u542b\u5927\u7ea6 50% \u7684\u901a\u7528\u77e5\u8bc6\u6807\u8bb0\u300125% \u7684\u6570\u5b66\u548c\u63a8\u7406\u6807\u8bb0\u300117% \u7684\u4ee3\u7801\u6807\u8bb0\u4ee5\u53ca 8% \u7684\u591a\u8bed\u8a00\u6807\u8bb0\u3002<\/span><\/span><\/span><\/p>\n<h4>3.13 \u9000\u706b\u6570\u636e<\/h4>\n<p>\u5b9e\u8bc1\u7ed3\u679c\u8868\u660e\uff0c\u5728\u5c11\u91cf\u9ad8\u8d28\u91cf\u4ee3\u7801\u548c\u6570\u5b66\u6570\u636e\u4e0a\u8fdb\u884c\u9000\u706b\uff08\u89c1\u7b2c 3.4.3 \u8282\uff09\u53ef\u4ee5\u63d0\u5347\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u5173\u952e\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u6027\u80fd\u3002\u7c7b\u4f3c\u4e8e\u674e\u7b49\u4eba\u7684\u7814\u7a76\uff082024b\uff09\uff0c\u6211\u4eec\u4f7f\u7528\u5305\u542b\u9009\u5b9a\u9886\u57df\u9ad8\u54c1\u8d28\u6570\u636e\u7684\u6df7\u5408\u6570\u636e\u96c6\u8fdb\u884c\u9000\u706b\u3002\u6211\u4eec\u7684\u9000\u706b\u6570\u636e\u4e0d\u5305\u542b\u4efb\u4f55\u6765\u81ea\u5e38\u7528\u57fa\u51c6\u6d4b\u8bd5\u7684\u8bad\u7ec3\u96c6\u3002\u8fd9\u4f7f\u6211\u4eec\u80fd\u591f\u8bc4\u4f30 Llama 3 \u7684\u771f\u5b9e\u5c11\u6837\u672c\u5b66\u4e60\u80fd\u529b\u548c\u57df\u5916\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p>\u9075\u5faa OpenAI (2023a)\uff0c\u6211\u4eec\u5728 GSM8k (Cobbe \u7b49\u4eba\uff0c2021) \u548c MATH (Hendrycks \u7b49\u4eba\uff0c2021b) \u8bad\u7ec3\u96c6\u4e0a\u8bc4\u4f30\u4e86\u9000\u706b\u7684\u6548\u679c\u3002\u6211\u4eec\u53d1\u73b0\u9000\u706b\u5206\u522b\u63d0\u9ad8\u4e86\u9884\u8bad\u7ec3 Llama 3 8B \u6a21\u578b\u5728 GSM8k \u548c MATH \u9a8c\u8bc1\u96c6\u4e0a\u7684\u6027\u80fd 24.0% \u548c 6.4%\u3002\u7136\u800c\uff0c\u5bf9\u4e8e 405B \u6a21\u578b\u7684\u6539\u8fdb\u53ef\u4ee5\u5ffd\u7565\u4e0d\u8ba1\uff0c\u8fd9\u8868\u660e\u6211\u4eec\u7684\u65d7\u8230\u6a21\u578b\u5177\u6709\u5f3a\u5927\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60\u548c\u63a8\u7406\u80fd\u529b\uff0c\u4e0d\u9700\u8981\u7279\u5b9a\u9886\u57df\u7684\u8bad\u7ec3\u6837\u672c\u5c31\u80fd\u83b7\u5f97\u5f3a\u52b2\u7684\u6027\u80fd\u3002<\/p>\n<p><strong>\u4f7f\u7528\u9000\u706b\u8bc4\u4f30\u6570\u636e\u8d28\u91cf\u3002<\/strong>\u4e0e Blakeney \u7b49\u4eba (2024) \u4e00\u6837\uff0c\u6211\u4eec\u53d1\u73b0\u9000\u706b\u4f7f\u6211\u4eec\u80fd\u591f\u5224\u65ad\u5c0f\u9886\u57df\u7279\u5b9a\u6570\u636e\u96c6\u7684\u4ef7\u503c\u3002\u6211\u4eec\u901a\u8fc7\u5c06\u5df2\u8bad\u7ec3 50% \u7684 Llama 3 8B \u6a21\u578b\u7684\u5b66\u4e60\u7387\u5728 400 \u4ebf\u4e2a\u6807\u8bb0\u4e0a\u7ebf\u6027\u9000\u706b\u5230 0 \u6765\u8861\u91cf\u8fd9\u4e9b\u6570\u636e\u96c6\u7684\u4ef7\u503c\u3002\u5728\u8fd9\u4e9b\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u4e3a\u65b0\u6570\u636e\u96c6\u5206\u914d 30% \u7684\u6743\u91cd\uff0c\u5176\u4f59 70% \u7684\u6743\u91cd\u5206\u914d\u7ed9\u9ed8\u8ba4\u6570\u636e\u6df7\u5408\u3002\u4e0e\u5bf9\u6bcf\u4e2a\u5c0f\u6570\u636e\u96c6\u6267\u884c\u6807\u5ea6\u5f8b\u5b9e\u9a8c\u76f8\u6bd4\uff0c\u4f7f\u7528\u9000\u706b\u6765\u8bc4\u4f30\u65b0\u7684\u6570\u636e\u6e90\u66f4\u52a0\u9ad8\u6548\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>3.2 \u6a21\u578b\u67b6\u6784<\/h2>\n<p>&nbsp;<\/p>\n<p>Llama 3 \u4f7f\u7528\u6807\u51c6\u7684\u7a20\u5bc6 Transformer \u67b6\u6784 (Vaswani \u7b49\u4eba\uff0c2017)\u3002\u5b83\u7684\u6a21\u578b\u67b6\u6784\u4e0e Llama \u548c Llama 2\uff08Touvron \u7b49\u4eba\uff0c2023a, b\uff09\u6ca1\u6709\u663e\u8457\u533a\u522b\uff1b\u6211\u4eec\u7684\u6027\u80fd\u63d0\u5347\u4e3b\u8981\u6765\u81ea\u4e8e\u6570\u636e\u8d28\u91cf\u548c\u591a\u6837\u6027\u7684\u6539\u8fdb\u4ee5\u53ca\u8bad\u7ec3\u89c4\u6a21\u7684\u6269\u5927\u3002<\/p>\n<p>\u6211\u4eec\u505a\u4e86\u51e0\u4e2a\u5c0f\u7684\u4fee\u6539\uff1a<\/p>\n<ul>\n<li>\u6211\u4eec\u4f7f\u7528\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b (GQA; Ainslie \u7b49\u4eba (2023))\uff0c\u5176\u4e2d 8 \u4e2a\u5173\u952e\u503c\u5934\u7528\u6765\u63d0\u9ad8\u63a8\u7406\u901f\u5ea6\uff0c\u5e76\u51cf\u5c11\u89e3\u7801\u8fc7\u7a0b\u4e2d\u5173\u952e\u503c\u7f13\u5b58\u7684\u5927\u5c0f\u3002<\/li>\n<li>\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a\u6ce8\u610f\u529b\u63a9\u7801\u6765\u9632\u6b62\u5e8f\u5217\u4e2d\u4e0d\u540c\u6587\u6863\u4e4b\u95f4\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u8fd9\u79cd\u6539\u53d8\u5728\u6807\u51c6\u9884\u8bad\u7ec3\u671f\u95f4\u5f71\u54cd\u6709\u9650\uff0c\u4f46\u5728\u975e\u5e38\u957f\u5e8f\u5217\u7684\u6301\u7eed\u9884\u8bad\u7ec3\u4e2d\u5f88\u91cd\u8981\u3002<\/li>\n<li>\u6211\u4eec\u4f7f\u7528 128K \u4e2a\u6807\u8bb0\u7684\u8bcd\u6c47\u8868\u3002\u6211\u4eec\u7684\u6807\u8bb0\u8bcd\u6c47\u8868\u5c06 tiktoken3 \u8bcd\u6c47\u8868\u7684 100K \u4e2a\u6807\u8bb0\u4e0e 28K \u4e2a\u989d\u5916\u7684\u6807\u8bb0\u76f8\u7ed3\u5408\uff0c\u4ee5\u66f4\u597d\u5730\u652f\u6301\u975e\u82f1\u8bed\u8bed\u8a00\u3002\u4e0e Llama 2 \u8bcd\u6c47\u8868\u76f8\u6bd4\uff0c\u6211\u4eec\u7684\u65b0\u8bcd\u6c47\u8868\u5c06\u82f1\u8bed\u6570\u636e\u6837\u672c\u7684\u538b\u7f29\u7387\u4ece 3.17 \u63d0\u9ad8\u5230 3.94 \u4e2a\u5b57\u7b26\/\u6807\u8bb0\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5728\u76f8\u540c\u7684\u8bad\u7ec3\u8ba1\u7b97\u91cf\u4e0b\u201c\u9605\u8bfb\u201d\u66f4\u591a\u6587\u672c\u3002\u6211\u4eec\u8fd8\u53d1\u73b0\uff0c\u6dfb\u52a0\u6765\u81ea\u7279\u5b9a\u975e\u82f1\u8bed\u8bed\u8a00\u7684 28K \u4e2a\u6807\u8bb0\u6539\u5584\u4e86\u538b\u7f29\u7387\u548c\u4e0b\u6e38\u6027\u80fd\uff0c\u800c\u5bf9\u82f1\u8bed\u6807\u8bb0\u5316\u6ca1\u6709\u5f71\u54cd\u3002<\/li>\n<li>\u6211\u4eec\u5c06 RoPE \u57fa\u7840\u9891\u7387\u8d85\u53c2\u6570\u63d0\u9ad8\u5230 500,000\u3002\u8fd9\u4f7f\u6211\u4eec\u80fd\u591f\u66f4\u597d\u5730\u652f\u6301\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\uff1bXiong \u7b49\u4eba\uff082023\uff09\u8868\u660e\u8be5\u503c\u5bf9\u4e8e\u4e0a\u4e0b\u6587\u957f\u5ea6\u9ad8\u8fbe 32,768 \u7684\u60c5\u51b5\u6709\u6548\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-3434\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a35620bfce08fdb.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"1156\" height=\"409\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a35620bfce08fdb.png 1156w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a35620bfce08fdb-300x106.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a35620bfce08fdb-1024x362.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a35620bfce08fdb-768x272.png 768w\" sizes=\"auto, (max-width: 1156px) 100vw, 1156px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>Llama 3 405B \u4f7f\u7528\u4e86\u4e00\u4e2a\u5177\u6709 126 \u5c42\u300116,384 \u4e2a\u6807\u8bb0\u8868\u793a\u7ef4\u5ea6\u548c 128 \u4e2a\u6ce8\u610f\u529b\u5934\u7684\u67b6\u6784\uff1b\u6709\u5173\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u9605\u8868 3\u3002\u8fd9\u5bfc\u81f4\u6a21\u578b\u5c3a\u5bf8\u7ea6\u4e3a\u6839\u636e\u6211\u4eec\u7684\u6570\u636e\u548c 3.8 \u00d7 10^25 FLOPs \u7684\u8bad\u7ec3\u9884\u7b97\u8ba1\u7b97\u51fa\u7684\u8ba1\u7b97\u6700\u4f18\u503c\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3.2.1 \u5c3a\u5ea6\u5b9a\u5f8b<\/h3>\n<p>\u6211\u4eec\u5229\u7528Scaling Laws\uff08Hoffmann \u7b49\u4eba\uff0c2022 \u5e74\uff1bKaplan \u7b49\u4eba\uff0c2020 \u5e74\uff09\u6765\u786e\u5b9a\u5728\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0c\u65d7\u8230\u6a21\u578b\u7684\u6700\u4f73\u89c4\u6a21\u3002\u9664\u4e86\u786e\u5b9a\u6700\u4f73\u6a21\u578b\u89c4\u6a21\u4e4b\u5916\uff0c\u9884\u6d4b\u65d7\u8230\u6a21\u578b\u5728\u4e0b\u6e38\u57fa\u51c6\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u4e5f\u9762\u4e34\u7740\u91cd\u5927\u6311\u6218\uff0c\u539f\u56e0\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u73b0\u6709\u7684 Scaling Laws \u901a\u5e38\u53ea\u9884\u6d4b\u4e0b\u4e00\u4e2a\u6807\u8bb0\u9884\u6d4b\u635f\u5931\uff0c\u800c\u4e0d\u662f\u7279\u5b9a\u7684\u57fa\u51c6\u6d4b\u8bd5\u6027\u80fd\u3002<\/li>\n<li>Scaling Laws \u53ef\u80fd\u5b58\u5728\u566a\u58f0\u4e14\u4e0d\u53ef\u9760\uff0c\u56e0\u4e3a\u5b83\u4eec\u662f\u57fa\u4e8e\u4f7f\u7528\u8f83\u5c0f\u8ba1\u7b97\u9884\u7b97\u8fdb\u884c\u7684\u9884\u8bad\u7ec3\u8fd0\u884c\u5f00\u53d1\u800c\u6765\u7684\uff08Wei \u7b49\u4eba\uff0c2022b\uff09\u3002<\/li>\n<\/ol>\n<p>\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e9b\u6311\u6218\uff0c\u6211\u4eec\u5b9e\u65bd\u4e86\u4e00\u79cd\u4e24\u9636\u6bb5\u65b9\u6cd5\u6765\u5f00\u53d1\u80fd\u591f\u51c6\u786e\u9884\u6d4b\u4e0b\u6e38\u57fa\u51c6\u6d4b\u8bd5\u6027\u80fd\u7684 Scaling Laws\uff1a<\/p>\n<ol>\n<li>\u6211\u4eec\u9996\u5148\u5efa\u7acb\u9884\u8bad\u7ec3 FLOPs \u4e0e\u8ba1\u7b97\u6700\u4f73\u6a21\u578b\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u7684\u8d1f\u5bf9\u6570\u4f3c\u7136\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u3002<\/li>\n<li>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5229\u7528 Scaling Laws \u6a21\u578b\u548c\u4e4b\u524d\u4f7f\u7528\u66f4\u9ad8\u8ba1\u7b97 FLOPs \u8bad\u7ec3\u7684\u65e7\u6a21\u578b\uff0c\u5c06\u4e0b\u6e38\u4efb\u52a1\u4e0a\u7684\u8d1f\u5bf9\u6570\u4f3c\u7136\u4e0e\u4efb\u52a1\u51c6\u786e\u7387\u76f8\u5173\u8054\u3002\u5728\u6b64\u6b65\u9aa4\u4e2d\uff0c\u6211\u4eec\u4e13\u95e8\u5229\u7528 Llama 2 \u7cfb\u5217\u6a21\u578b\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u79cd\u65b9\u6cd5\u4f7f\u6211\u4eec\u80fd\u591f\u6839\u636e\u7279\u5b9a\u6570\u91cf\u7684\u9884\u8bad\u7ec3 FLOPs \u6765\u9884\u6d4b\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\uff08\u5bf9\u4e8e\u8ba1\u7b97\u6700\u4f73\u6a21\u578b\uff09\u3002\u6211\u4eec\u4f7f\u7528\u7c7b\u4f3c\u7684\u65b9\u6cd5\u6765\u9009\u62e9\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u6570\u636e\u7ec4\u5408\uff08\u8bf7\u53c2\u89c1\u7b2c 3.4 \u8282\uff09\u3002<\/p>\n<p>Scaling<strong> Law \u5b9e\u9a8c\u3002<\/strong>\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u901a\u8fc7\u4f7f\u7528\u5728 6 \u00d7 10^18 FLOPs \u548c 10^22 FLOPs \u4e4b\u95f4\u7684\u8ba1\u7b97\u9884\u7b97\u9884\u8bad\u7ec3\u6a21\u578b\u6765\u6784\u5efa Scaling Laws\u3002\u5728\u6bcf\u4e2a\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0c\u6211\u4eec\u9884\u8bad\u7ec3\u89c4\u6a21\u5728 40M \u5230 16B \u53c2\u6570\u4e4b\u95f4\u7684\u6a21\u578b\uff0c\u5e76\u5728\u6bcf\u4e2a\u8ba1\u7b97\u9884\u7b97\u4e0b\u4f7f\u7528\u4e00\u90e8\u5206\u6a21\u578b\u89c4\u6a21\u3002\u5728\u8fd9\u4e9b\u8bad\u7ec3\u8fd0\u884c\u4e2d\uff0c\u6211\u4eec\u5728 2,000 \u6b21\u8bad\u7ec3\u6b65\u9aa4\u5185\u4f7f\u7528\u4f59\u5f26\u5b66\u4e60\u7387\u8c03\u5ea6\u548c\u7ebf\u6027\u9884\u70ed\u3002\u5cf0\u503c\u5b66\u4e60\u7387\u6839\u636e\u6a21\u578b\u7684\u5927\u5c0f\u8bbe\u7f6e\u4e3a 2 \u00d7 10^-4 \u5230 4 \u00d7 10^-4 \u4e4b\u95f4\u3002\u6211\u4eec\u5c06\u4f59\u5f26\u8870\u51cf\u8bbe\u7f6e\u4e3a\u5cf0\u503c\u7684 0.1 \u500d\u3002\u6bcf\u4e00\u6b65\u7684\u6743\u91cd\u8870\u51cf\u90fd\u8bbe\u7f6e\u4e3a\u8be5\u6b65\u5b66\u4e60\u7387\u7684 0.1 \u500d\u3002\u6211\u4eec\u5bf9\u6bcf\u4e2a\u8ba1\u7b97\u89c4\u6a21\u4f7f\u7528\u56fa\u5b9a\u7684\u6279\u91cf\u5927\u5c0f\uff0c\u8303\u56f4\u5728 250K \u5230 4M \u4e4b\u95f4\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3435\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0ec0ddc0ca3c642.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"805\" height=\"362\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0ec0ddc0ca3c642.png 1131w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0ec0ddc0ca3c642-300x135.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0ec0ddc0ca3c642-1024x461.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0ec0ddc0ca3c642-768x346.png 768w\" sizes=\"auto, (max-width: 805px) 100vw, 805px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>\u8fd9\u4e9b\u5b9e\u9a8c\u4ea7\u751f\u4e86\u56fe 2 \u4e2d\u7684 IsoFLOPs \u66f2\u7ebf\u3002\u8fd9\u4e9b\u66f2\u7ebf\u4e2d\u7684\u635f\u5931\u662f\u5728\u5355\u72ec\u7684\u9a8c\u8bc1\u96c6\u4e0a\u6d4b\u91cf\u7684\u3002\u6211\u4eec\u4f7f\u7528\u4e8c\u9636\u591a\u9879\u5f0f\u62df\u5408\u6d4b\u5f97\u7684\u635f\u5931\u503c\uff0c\u5e76\u786e\u5b9a\u6bcf\u4e2a\u629b\u7269\u7ebf\u7684\u6700\u5c0f\u503c\u3002\u6211\u4eec\u5c06\u629b\u7269\u7ebf\u7684\u6700\u5c0f\u503c\u79f0\u4e3a\u76f8\u5e94\u9884\u8bad\u7ec3\u8ba1\u7b97\u9884\u7b97\u4e0b\u7684\u8ba1\u7b97\u6700\u4f18\u6a21\u578b\u3002<\/p>\n<p>\u6211\u4eec\u4f7f\u7528\u4ee5\u8fd9\u79cd\u65b9\u5f0f\u8bc6\u522b\u7684\u8ba1\u7b97\u6700\u4f18\u6a21\u578b\u6765\u9884\u6d4b\u7279\u5b9a\u8ba1\u7b97\u9884\u7b97\u4e0b\u7684\u6700\u4f73\u8bad\u7ec3\u6807\u8bb0\u6570\u91cf\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u5047\u8bbe\u8ba1\u7b97\u9884\u7b97 C \u4e0e\u6700\u4f73\u8bad\u7ec3\u6807\u8bb0\u6570\u91cf N (C) \u4e4b\u95f4\u5b58\u5728\u5e42\u5f8b\u5173\u7cfb\uff1a<\/p>\n<p>N (C) = AC \u03b1 \u3002<\/p>\n<p>\u6211\u4eec\u4f7f\u7528\u56fe 2 \u4e2d\u7684\u6570\u636e\u62df\u5408 A \u548c \u03b1\u3002\u6211\u4eec\u53d1\u73b0 (\u03b1, A) = (0.53, 0.29)\uff1b\u76f8\u5e94\u7684\u62df\u5408\u5982\u56fe 3 \u6240\u793a\u3002\u5c06\u6240\u5f97\u7f29\u653e\u5b9a\u5f8b\u5916\u63a8\u5230 3.8 \u00d7 10 25 FLOPs \u5efa\u8bae\u8bad\u7ec3\u4e00\u4e2a\u5305\u542b 402B \u53c2\u6570\u7684\u6a21\u578b\uff0c\u5e76\u4f7f\u7528 16.55T \u4e2a\u6807\u8bb0\u3002<\/p>\n<p>\u4e00\u4e2a\u91cd\u8981\u7684\u89c2\u5bdf\u7ed3\u679c\u662f\uff0c\u968f\u7740\u8ba1\u7b97\u9884\u7b97\u7684\u589e\u52a0\uff0cIsoFLOPs \u66f2\u7ebf\u5728\u6700\u5c0f\u503c\u5468\u56f4\u53d8\u5f97\u66f4\u5e73\u5766\u3002\u8fd9\u610f\u5473\u7740\u65d7\u8230\u6a21\u578b\u7684\u6027\u80fd\u5bf9\u6a21\u578b\u5927\u5c0f\u548c\u8bad\u7ec3\u6807\u8bb0\u4e4b\u95f4\u6743\u8861\u7684\u5c0f\u53d8\u5316\u76f8\u5bf9\u7a33\u5b9a\u3002\u57fa\u4e8e\u8fd9\u4e2a\u89c2\u5bdf\u7ed3\u679c\uff0c\u6211\u4eec\u6700\u7ec8\u51b3\u5b9a\u8bad\u7ec3\u4e00\u4e2a\u5305\u542b 405B \u53c2\u6570\u7684\u65d7\u8230\u6a21\u578b\u3002<\/p>\n<p><strong>\u9884\u6d4b\u4e0b\u6e38\u4efb\u52a1\u7684\u6027\u80fd\u3002<\/strong>\u6211\u4eec\u4f7f\u7528\u751f\u6210\u7684\u8ba1\u7b97\u6700\u4f18\u6a21\u578b\u6765\u9884\u6d4b\u65d7\u8230 Llama 3 \u6a21\u578b\u5728\u57fa\u51c6\u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\u3002\u9996\u5148\uff0c\u6211\u4eec\u5c06\u57fa\u51c6\u4e2d\u6b63\u786e\u7b54\u6848\u7684\uff08\u5f52\u4e00\u5316\uff09\u8d1f\u5bf9\u6570\u4f3c\u7136\u4e0e\u8bad\u7ec3 FLOPs \u7ebf\u6027\u76f8\u5173\u8054\u3002\u5728\u6b64\u5206\u6790\u4e2d\uff0c\u6211\u4eec\u4ec5\u4f7f\u7528\u4e86\u5728\u4e0a\u8ff0\u6570\u636e\u6df7\u5408\u4e0a\u8bad\u7ec3\u81f3 10^22 FLOPs \u7684\u7f29\u653e\u5b9a\u5f8b\u6a21\u578b\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u4f7f\u7528\u7f29\u653e\u5b9a\u5f8b\u6a21\u578b\u548c Llama 2 \u6a21\u578b\u5efa\u7acb\u5bf9\u6570\u4f3c\u7136\u548c\u51c6\u786e\u7387\u4e4b\u95f4\u7684 S \u578b\u5173\u7cfb\uff0cLlama 2 \u6a21\u578b\u4f7f\u7528 Llama 2 \u6570\u636e\u6df7\u5408\u548c\u6807\u8bb0\u5668\u8fdb\u884c\u8bad\u7ec3\u3002\u6211\u4eec\u5728\u56fe 4 \u4e2d\u5c55\u793a\u4e86\u8fd9\u4e2a\u5b9e\u9a8c\u5728 ARC Challenge \u57fa\u51c6\u4e0a\u7684\u7ed3\u679c\uff09\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u79cd\u4e24\u6b65\u7f29\u653e\u5b9a\u5f8b\u9884\u6d4b\uff08\u5916\u63a8\u8d85\u8fc7\u56db\u4e2a\u6570\u91cf\u7ea7\uff09\u76f8\u5f53\u51c6\u786e\uff1a\u5b83\u4ec5\u7565\u5fae\u4f4e\u4f30\u4e86\u65d7\u8230 Llama 3 \u6a21\u578b\u7684\u6700\u7ec8\u6027\u80fd\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>3.3 \u57fa\u7840\u8bbe\u65bd\u3001\u6269\u5c55\u548c\u6548\u7387<\/h2>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u63cf\u8ff0\u4e86\u652f\u6301 Llama 3 405B \u9884\u8bad\u7ec3\u7684\u786c\u4ef6\u548c\u57fa\u7840\u67b6\u6784\uff0c\u5e76\u8ba8\u8bba\u4e86\u51e0\u9879\u4f18\u5316\u63aa\u65bd\uff0c\u8fd9\u4e9b\u4f18\u5316\u63aa\u65bd\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3.3.1 \u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd<\/h3>\n<p>Llama 1 \u548c Llama 2 \u6a21\u578b\u5728 Meta \u7684 AI \u7814\u7a76\u8d85\u7ea7\u96c6\u7fa4\uff08Lee \u548c Sengupta\uff0c2022\uff09\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u968f\u7740\u6211\u4eec\u8fdb\u4e00\u6b65\u6269\u5927\u89c4\u6a21\uff0cLlama 3 \u7684\u8bad\u7ec3\u88ab\u8fc1\u79fb\u5230 Meta \u7684\u751f\u4ea7\u96c6\u7fa4\uff08Lee \u7b49\u4eba\uff0c2024\uff09\u3002\u8fd9\u79cd\u8bbe\u7f6e\u4f18\u5316\u4e86\u751f\u4ea7\u7ea7\u53ef\u9760\u6027\uff0c\u8fd9\u662f\u6211\u4eec\u5728\u6269\u5927\u8bad\u7ec3\u89c4\u6a21\u65f6\u81f3\u5173\u91cd\u8981\u7684\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3436\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/f35903d4cdc8e57.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"738\" height=\"317\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/f35903d4cdc8e57.png 1041w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/f35903d4cdc8e57-300x129.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/f35903d4cdc8e57-1024x440.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/f35903d4cdc8e57-768x330.png 768w\" sizes=\"auto, (max-width: 738px) 100vw, 738px\" \/><\/strong><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u8ba1\u7b97\u8d44\u6e90\uff1a<\/strong>\u00a0Llama 3 405B \u5728\u9ad8\u8fbe 16,000 \u5757 H100 GPU \u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6bcf\u5757 GPU \u4ee5 700W TDP \u8fd0\u884c\uff0c\u62e5\u6709 80GB HBM3\uff0c\u4f7f\u7528 Meta \u7684 Grand Teton AI \u670d\u52a1\u5668\u5e73\u53f0\uff08Matt Bowman\uff0c2022\uff09\u3002\u6bcf\u4e2a\u670d\u52a1\u5668\u914d\u5907\u516b\u5757 GPU \u548c\u4e24\u5757 CPU\u3002\u5728\u670d\u52a1\u5668\u5185\u90e8\uff0c\u516b\u5757 GPU \u901a\u8fc7 NVLink \u8fde\u63a5\u3002\u8bad\u7ec3\u4f5c\u4e1a\u4f7f\u7528 MAST\uff08Choudhury \u7b49\u4eba\uff0c2024\uff09\u8fdb\u884c\u8c03\u5ea6\uff0c\u8fd9\u662f Meta \u7684\u5168\u7403\u89c4\u6a21\u8bad\u7ec3\u8c03\u5ea6\u7a0b\u5e8f\u3002<\/p>\n<p><strong>\u5b58\u50a8\uff1a<\/strong>\u00a0Tectonic\uff08Pan \u7b49\u4eba\uff0c2021\uff09\uff0cMeta \u7684\u901a\u7528\u5206\u5e03\u5f0f\u6587\u4ef6\u7cfb\u7edf\uff0c\u7528\u4e8e\u6784\u5efa Llama 3 \u9884\u8bad\u7ec3\u7684\u5b58\u50a8\u7ed3\u6784\uff08Battey \u548c Gupta\uff0c2024\uff09\u3002\u5b83\u63d0\u4f9b 240PB \u7684\u5b58\u50a8\u7a7a\u95f4\uff0c\u7531 7,500 \u53f0\u914d\u5907 SSD \u7684\u670d\u52a1\u5668\u7ec4\u6210\uff0c\u652f\u6301\u53ef\u6301\u7eed\u7684\u541e\u5410\u91cf\u4e3a 2TB\/s\uff0c\u5cf0\u503c\u541e\u5410\u91cf\u4e3a 7TB\/s\u3002\u4e00\u4e2a\u4e3b\u8981\u6311\u6218\u662f\u652f\u6301\u9ad8\u7a81\u53d1\u6027\u68c0\u67e5\u70b9\u5199\u5165\uff0c\u8fd9\u4e9b\u5199\u5165\u4f1a\u5728\u77ed\u65f6\u95f4\u5185\u9971\u548c\u5b58\u50a8\u7ed3\u6784\u3002\u68c0\u67e5\u70b9\u4fdd\u5b58\u6bcf\u4e2a GPU \u7684\u6a21\u578b\u72b6\u6001\uff0c\u8303\u56f4\u4ece 1MB \u5230\u6bcf GPU 4GB\uff0c\u7528\u4e8e\u6062\u590d\u548c\u8c03\u8bd5\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\u6700\u5c0f\u5316 GPU \u5728\u68c0\u67e5\u70b9\u671f\u95f4\u7684\u6682\u505c\u65f6\u95f4\uff0c\u5e76\u589e\u52a0\u68c0\u67e5\u70b9\u9891\u7387\u4ee5\u51cf\u5c11\u5728\u6062\u590d\u540e\u4e22\u5931\u7684\u5de5\u4f5c\u91cf\u3002<\/p>\n<p><strong>\u7f51\u7edc\uff1a<\/strong>\u00a0Llama 3 405B \u4f7f\u7528\u57fa\u4e8e Arista 7800 \u548c Minipack2 \u5f00\u653e\u8ba1\u7b97\u9879\u76ee\uff08OCP\uff09\u673a\u67b6\u4ea4\u6362\u673a\u7684 RDMA over Converged Ethernet\uff08RoCE\uff09\u67b6\u6784\u3002Llama 3 \u7cfb\u5217\u4e2d\u8f83\u5c0f\u6a21\u578b\u4f7f\u7528 Nvidia Quantum2 Infiniband \u7f51\u7edc\u8fdb\u884c\u8bad\u7ec3\u3002RoCE \u548c Infiniband \u96c6\u7fa4\u90fd\u5229\u7528 GPU \u4e4b\u95f4\u7684 400 Gbps \u94fe\u8def\u8fde\u63a5\u3002\u5c3d\u7ba1\u8fd9\u4e9b\u96c6\u7fa4\u7684\u5e95\u5c42\u7f51\u7edc\u6280\u672f\u5b58\u5728\u5dee\u5f02\uff0c\u4f46\u6211\u4eec\u5bf9\u4e24\u8005\u90fd\u8fdb\u884c\u4e86\u8c03\u6574\uff0c\u4ee5\u63d0\u4f9b\u7b49\u6548\u7684\u6027\u80fd\u6765\u5904\u7406\u8fd9\u4e9b\u5927\u578b\u8bad\u7ec3\u5de5\u4f5c\u91cf\u3002\u6211\u4eec\u5c06\u8fdb\u4e00\u6b65\u8be6\u7ec6\u9610\u8ff0\u6211\u4eec\u7684 RoCE \u7f51\u7edc\uff0c\u56e0\u4e3a\u6211\u4eec\u5b8c\u5168\u638c\u63e1\u4e86\u5b83\u7684\u8bbe\u8ba1\u3002<\/p>\n<ul>\n<li><strong>\u7f51\u7edc\u62d3\u6251\uff1a<\/strong> \u6211\u4eec\u7684\u57fa\u4e8e RoCE \u7684 AI \u96c6\u7fa4\u5305\u542b 24,000 \u5757 GPU\uff08\u6ce8\u811a 5\uff09\uff0c\u901a\u8fc7\u4e09\u5c42 Clos \u7f51\u7edc\u8fde\u63a5\uff08Lee \u7b49\u4eba\uff0c2024\uff09\u3002\u5728\u5e95\u5c42\uff0c\u6bcf\u4e2a\u673a\u67b6\u6258\u7ba1 16 \u5757 GPU\uff0c\u5206\u914d\u5230\u4e24\u53f0\u670d\u52a1\u5668\u4e0a\uff0c\u5e76\u901a\u8fc7\u5355\u4e2a Minipack2 \u673a\u67b6\u9876\u7aef\uff08ToR\uff09\u4ea4\u6362\u673a\u8fde\u63a5\u3002\u5728\u4e2d\u95f4\u5c42\uff0c192 \u4e2a\u8fd9\u6837\u7684\u673a\u67b6\u901a\u8fc7\u96c6\u7fa4\u4ea4\u6362\u673a\u8fde\u63a5\u5f62\u6210\u4e00\u4e2a\u5305\u542b 3,072 \u5757 GPU \u7684 Pod\uff0c\u5177\u6709\u5168\u53cc\u5411\u5e26\u5bbd\uff0c\u786e\u4fdd\u6ca1\u6709\u8fc7\u5ea6\u8ba2\u9605\u3002\u5728\u9876\u5c42\uff0c\u540c\u4e00\u4e2a\u6570\u636e\u4e2d\u5fc3\u5efa\u7b51\u7269\u5185\u7684\u516b\u4e2a\u8fd9\u6837\u7684 Pod \u901a\u8fc7\u805a\u5408\u4ea4\u6362\u673a\u8fde\u63a5\u4ee5\u5f62\u6210\u4e00\u4e2a 24,000 \u5757 GPU \u7684\u96c6\u7fa4\u3002\u7136\u800c\uff0c\u805a\u5408\u5c42\u7684\u7f51\u7edc\u8fde\u63a5\u6ca1\u6709\u4fdd\u6301\u5168\u53cc\u5411\u5e26\u5bbd\uff0c\u800c\u662f\u91c7\u7528\u4e86 1:7 \u7684\u8fc7\u5ea6\u8ba2\u9605\u7387\u3002\u6211\u4eec\u7684\u6a21\u578b\u5e76\u884c\u65b9\u6cd5\uff08\u8bf7\u53c2\u89c1\u7b2c 3.3.2 \u8282\uff09\u548c\u8bad\u7ec3\u4f5c\u4e1a\u8c03\u5ea6\u7a0b\u5e8f\uff08Choudhury \u7b49\u4eba\uff0c2024\uff09\u90fd\u88ab\u4f18\u5316\u4e3a\u77e5\u9053\u7f51\u7edc\u62d3\u6251\u7ed3\u6784\uff0c\u65e8\u5728\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11 Pod \u4e4b\u95f4\u7684\u7f51\u7edc\u901a\u4fe1\u3002<\/li>\n<li><strong>\u8d1f\u8f7d\u5747\u8861\uff1a<\/strong> \u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u8bad\u7ec3\u4f1a\u4ea7\u751f\u96be\u4ee5\u901a\u8fc7\u4f20\u7edf\u65b9\u6cd5\uff08\u5982\u7b49\u6210\u672c\u591a\u8def\u5f84\uff08ECMP\uff09\u8def\u7531\uff09\u5728\u6240\u6709\u53ef\u7528\u7f51\u7edc\u8def\u5f84\u4e0a\u5e73\u8861\u7684\u7e41\u91cd\u7f51\u7edc\u6d41\u91cf\u3002\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u4e24\u79cd\u6280\u672f\u3002\u9996\u5148\uff0c\u6211\u4eec\u7684\u96c6\u5408\u5e93\u5728\u4e24\u4e2a GPU \u4e4b\u95f4\u521b\u5efa 16 \u4e2a\u7f51\u7edc\u6d41\uff0c\u800c\u4e0d\u662f\u4e00\u4e2a\uff0c\u4ece\u800c\u51cf\u5c11\u4e86\u6bcf\u4e2a\u6d41\u7684\u6d41\u91cf\uff0c\u5e76\u4e3a\u8d1f\u8f7d\u5747\u8861\u63d0\u4f9b\u4e86\u66f4\u591a\u6d41\u3002\u5176\u6b21\uff0c\u6211\u4eec\u7684\u589e\u5f3a\u578b ECMP\uff08E-ECMP\uff09\u534f\u8bae\u901a\u8fc7\u5bf9 RoCE \u62a5\u5934\u6570\u636e\u5305\u4e2d\u7684\u5176\u4ed6\u5b57\u6bb5\u8fdb\u884c\u54c8\u5e0c\uff0c\u6709\u6548\u5730\u5728\u4e0d\u540c\u7684\u7f51\u7edc\u8def\u5f84\u4e0a\u5e73\u8861\u8fd9\u4e9b 16 \u4e2a\u6d41\u3002<\/li>\n<\/ul>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3437\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d62ae1ee9a37657.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"801\" height=\"169\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d62ae1ee9a37657.png 1047w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d62ae1ee9a37657-300x63.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d62ae1ee9a37657-1024x216.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d62ae1ee9a37657-768x162.png 768w\" sizes=\"auto, (max-width: 801px) 100vw, 801px\" \/><\/strong><\/p>\n<ul>\n<li><strong>\u62e5\u585e\u63a7\u5236\uff1a<\/strong>\u00a0\u6211\u4eec\u5728\u9aa8\u5e72\u7f51\u4e2d\u4f7f\u7528\u6df1\u7f13\u51b2\u4ea4\u6362\u673a\uff08Gangidi \u7b49\u4eba\uff0c2024\uff09\u6765\u5bb9\u7eb3\u7531\u96c6\u5408\u901a\u4fe1\u6a21\u5f0f\u5f15\u8d77\u7684\u77ac\u6001\u62e5\u585e\u548c\u7f13\u51b2\u3002\u8fd9\u6709\u52a9\u4e8e\u9650\u5236\u6301\u7eed\u62e5\u585e\u548c\u7531\u6162\u901f\u670d\u52a1\u5668\u5f15\u8d77\u7f51\u7edc\u53cd\u538b\u7684\u5f71\u54cd\uff0c\u8fd9\u5728\u8bad\u7ec3\u4e2d\u5f88\u5e38\u89c1\u3002\u6700\u540e\uff0c\u901a\u8fc7 E-ECMP \u5b9e\u73b0\u66f4\u597d\u7684\u8d1f\u8f7d\u5747\u8861\u5927\u5927\u964d\u4f4e\u4e86\u53d1\u751f\u62e5\u585e\u7684\u53ef\u80fd\u6027\u3002\u901a\u8fc7\u8fd9\u4e9b\u4f18\u5316\uff0c\u6211\u4eec\u6210\u529f\u5730\u8fd0\u884c\u4e86\u4e00\u4e2a 24,000 \u5757 GPU \u7684\u96c6\u7fa4\uff0c\u65e0\u9700\u4f7f\u7528\u4f20\u7edf\u7684\u62e5\u585e\u63a7\u5236\u65b9\u6cd5\uff0c\u4f8b\u5982\u6570\u636e\u4e2d\u5fc3\u91cf\u5316\u62e5\u585e\u901a\u77e5\uff08DCQCN\uff09\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>3.3.2 \u6a21\u578b\u89c4\u6a21\u6269\u5c55\u4e2d\u7684\u5e76\u884c\u6027<\/h3>\n<p>\u4e3a\u4e86\u6269\u5927\u6211\u4eec\u6700\u5927\u6a21\u578b\u7684\u8bad\u7ec3\u89c4\u6a21\uff0c\u6211\u4eec\u4f7f\u7528 4D \u5e76\u884c\u6280\u672f &#8211; \u4e00\u79cd\u7ed3\u5408\u56db\u79cd\u4e0d\u540c\u5e76\u884c\u65b9\u6cd5\u7684\u65b9\u6848 &#8211; \u5bf9\u6a21\u578b\u8fdb\u884c\u5206\u7247\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u6548\u5730\u5c06\u8ba1\u7b97\u5206\u5e03\u5230\u8bb8\u591a GPU \u4e0a\uff0c\u5e76\u786e\u4fdd\u6bcf\u4e2a GPU \u7684\u6a21\u578b\u53c2\u6570\u3001\u4f18\u5316\u5668\u72b6\u6001\u3001\u68af\u5ea6\u548c\u6fc0\u6d3b\u503c\u90fd\u80fd\u591f\u5bb9\u7eb3\u5728\u5b83\u7684 HBM \u4e2d\u3002\u6211\u4eec\u7684 4D \u5e76\u884c\u5b9e\u73b0\uff08\u5982 et al. (2020)\uff1bRen et al. (2021)\uff1bZhao et al. (2023b)) \u6240\u793a\uff0c\u5b83\u5bf9\u6a21\u578b\u3001\u4f18\u5316\u5668\u548c\u68af\u5ea6\u8fdb\u884c\u5206\u7247\uff0c\u540c\u65f6\u5b9e\u73b0\u4e86\u6570\u636e\u5e76\u884c\uff0c\u8be5\u5e76\u884c\u65b9\u6cd5\u5728\u591a\u4e2a GPU \u4e0a\u5e76\u884c\u5904\u7406\u6570\u636e\u5e76\u5728\u6bcf\u4e2a\u8bad\u7ec3\u6b65\u9aa4\u540e\u8fdb\u884c\u540c\u6b65\u3002\u6211\u4eec\u4f7f\u7528 FSDP \u5bf9 Llama 3 \u7684\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6\u8fdb\u884c\u5206\u7247\uff0c\u4f46\u5bf9\u4e8e\u6a21\u578b\u5206\u7247\uff0c\u6211\u4eec\u4e0d\u4f1a\u5728\u6b63\u5411\u8ba1\u7b97\u540e\u91cd\u65b0\u5206\u7247\uff0c\u4ee5\u907f\u514d\u5728\u53cd\u5411\u4f20\u9012\u8fc7\u7a0b\u4e2d\u53d1\u751f\u989d\u5916\u7684\u5168\u6536\u96c6\u901a\u4fe1\u3002<\/p>\n<p><strong>GPU \u5229\u7528\u7387\u3002<\/strong>\u901a\u8fc7\u4ed4\u7ec6\u8c03\u6574\u5e76\u884c\u914d\u7f6e\u3001\u786c\u4ef6\u548c\u8f6f\u4ef6\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86 BF16 \u6a21\u578b FLOPs \u5229\u7528\u7387\uff08MFU\uff1bChowdhery et al. (2023)) \u7684 38-43%\u3002\u8868 4 \u4e2d\u663e\u793a\u7684\u914d\u7f6e\u8868\u660e\uff0c\u4e0e 8K GPU \u548c DP=64 \u4e0a\u7684 43% \u76f8\u6bd4\uff0c16K GPU \u548c DP=128 \u4e0a MFU \u964d\u81f3 41% \u662f\u7531\u4e8e\u9700\u8981\u964d\u4f4e\u6bcf\u4e2a DP \u7ec4\u7684\u6279\u91cf\u5927\u5c0f\u4ee5\u4fdd\u6301\u8bad\u7ec3\u671f\u95f4\u5168\u5c40\u6807\u8bb0\u6570\u4e0d\u53d8\u3002<\/p>\n<p><strong>\u6d41\u6c34\u7ebf\u5e76\u884c\u6539\u8fdb\u3002<\/strong>\u6211\u4eec\u5728\u73b0\u6709\u7684\u5b9e\u73b0\u4e2d\u9047\u5230\u4e86\u51e0\u4e2a\u6311\u6218\uff1a<\/p>\n<ul>\n<li><strong>\u6279\u91cf\u5c3a\u5bf8\u9650\u5236\u3002<\/strong>\u5f53\u524d\u7684\u5b9e\u73b0\u5bf9\u6bcf\u4e2a GPU \u652f\u6301\u7684\u6279\u91cf\u5c3a\u5bf8\u6709\u9650\u5236\uff0c\u8981\u6c42\u5b83\u53ef\u88ab\u6d41\u6c34\u7ebf\u9636\u6bb5\u6570\u6574\u9664\u3002\u5bf9\u4e8e\u56fe 6 \u4e2d\u7684\u793a\u4f8b\uff0c\u6df1\u5ea6\u4f18\u5148\u8c03\u5ea6 (DFS) \u7684\u6d41\u6c34\u7ebf\u5e76\u884c\uff08Narayanan et al. (2021)\uff09\u9700\u8981 N = PP = 4\uff0c\u800c\u5e7f\u5ea6\u4f18\u5148\u8c03\u5ea6 (BFS\uff1bLamy-Poirier (2023)) \u9700\u8981 N = M\uff0c\u5176\u4e2d M \u662f\u603b\u5fae\u6279\u91cf\u6570\uff0cN \u662f\u540c\u4e00\u9636\u6bb5\u6b63\u5411\u6216\u53cd\u5411\u7684\u8fde\u7eed\u5fae\u6279\u91cf\u6570\u3002\u7136\u800c\uff0c\u9884\u8bad\u7ec3\u901a\u5e38\u9700\u8981\u7075\u6d3b\u5730\u8c03\u6574\u6279\u91cf\u5927\u5c0f\u3002<\/li>\n<li><strong>\u5185\u5b58\u4e0d\u5e73\u8861\u3002<\/strong>\u73b0\u6709\u7684\u6d41\u6c34\u7ebf\u5e76\u884c\u5b9e\u73b0\u4f1a\u5bfc\u81f4\u8d44\u6e90\u6d88\u8017\u4e0d\u5e73\u8861\u3002\u7b2c\u4e00\u9636\u6bb5\u7531\u4e8e\u5d4c\u5165\u548c\u9884\u70ed\u5fae\u6279\u6b21\u800c\u6d88\u8017\u66f4\u591a\u5185\u5b58\u3002<\/li>\n<li><strong>\u8ba1\u7b97\u4e0d\u5e73\u8861\u3002<\/strong>\u00a0\u5728\u6a21\u578b\u7684\u6700\u540e\u4e00\u5c42\u4e4b\u540e\uff0c\u6211\u4eec\u9700\u8981\u8ba1\u7b97\u8f93\u51fa\u548c\u635f\u5931\uff0c\u4f7f\u5f97\u8fd9\u4e2a\u9636\u6bb5\u6210\u4e3a\u6267\u884c\u5ef6\u8fdf\u7684\u74f6\u9888\u3002\u5176\u4e2d Di \u662f\u7b2c i \u4e2a\u5e76\u884c\u7ef4\u5ea6\u7684\u7d22\u5f15\u3002\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\uff0cGPU0[TP0, CP0, PP0, DP0] \u548c GPU1[TP1, CP0, PP0, DP0] \u5728\u540c\u4e00\u4e2a TP \u7ec4\uff0cGPU0 \u548c GPU2 \u5728\u540c\u4e00\u4e2a CP \u7ec4\uff0cGPU0 \u548c GPU4 \u5728\u540c\u4e00\u4e2a PP \u7ec4\uff0cGPU0 \u548c GPU8 \u5728\u540c\u4e00\u4e2a DP \u7ec4\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3439\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"796\" height=\"478\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621.png 1039w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621-300x180.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621-1024x615.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621-768x461.png 768w\" sizes=\"auto, (max-width: 796px) 100vw, 796px\" \/><\/strong><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u6211\u4eec\u4fee\u6539\u4e86\u7ba1\u9053\u8c03\u5ea6\u65b9\u5f0f\uff0c\u5982\u56fe 6 \u6240\u793a\uff0c\u8fd9\u5141\u8bb8\u7075\u6d3b\u5730\u8bbe\u7f6e N\u2014\u2014\u5728\u672c\u4f8b\u4e2d N = 5\uff0c\u5b83\u53ef\u4ee5\u5728\u6bcf\u4e2a\u6279\u6b21\u4e2d\u8fd0\u884c\u4efb\u610f\u6570\u91cf\u7684\u5fae\u6279\u91cf\u3002<\/strong>\u00a0\u8fd9\u4f7f\u5f97\u6211\u4eec\u53ef\u4ee5\uff1a<\/p>\n<p>(1) \u5f53\u6709\u6279\u91cf\u5927\u5c0f\u9650\u5236\u65f6\uff0c\u8fd0\u884c\u6bd4\u9636\u6bb5\u6570\u66f4\u5c11\u7684\u5fae\u6279\u91cf\uff1b\u6216<\/p>\n<p>(2) \u8fd0\u884c\u66f4\u591a\u5fae\u6279\u91cf\u6765\u9690\u85cf\u70b9\u5bf9\u70b9\u901a\u4fe1\uff0c\u5728\u6df1\u5ea6\u4f18\u5148\u8c03\u5ea6 (DFS) \u548c\u5e7f\u5ea6\u4f18\u5148\u8c03\u5ea6 (BFS) \u4e4b\u95f4\u627e\u5230\u6700\u4f73\u7684\u901a\u4fe1\u548c\u5185\u5b58\u6548\u7387\u3002\u4e3a\u4e86\u5e73\u8861\u7ba1\u9053\uff0c\u6211\u4eec\u5206\u522b\u4ece\u7b2c\u4e00\u9636\u6bb5\u548c\u6700\u540e\u4e00\u9636\u6bb5\u4e2d\u51cf\u5c11\u4e00\u4e2a Transformer \u5c42\u3002\u8fd9\u610f\u5473\u7740\u7b2c\u4e00\u9636\u6bb5\u4e0a\u7684\u7b2c\u4e00\u4e2a\u6a21\u578b\u5757\u53ea\u6709\u5d4c\u5165\u5c42\uff0c\u800c\u6700\u540e\u4e00\u9636\u6bb5\u4e0a\u7684\u6700\u540e\u4e00\u4e2a\u6a21\u578b\u5757\u53ea\u6709\u8f93\u51fa\u6295\u5f71\u548c\u635f\u5931\u8ba1\u7b97\u3002<\/p>\n<p><strong>\u4e3a\u4e86\u51cf\u5c11\u7ba1\u9053\u6c14\u6ce1\uff0c\u6211\u4eec\u5728\u4e00\u4e2a\u7ba1\u9053\u7b49\u7ea7\u4e0a\u4f7f\u7528\u4e86\u4ea4\u9519\u8c03\u5ea6\u65b9\u5f0f (Narayanan \u7b49\u4eba\uff0c2021)\uff0c\u62e5\u6709 V \u4e2a\u7ba1\u9053\u9636\u6bb5\u3002<\/strong>\u00a0\u6574\u4f53\u7ba1\u9053\u6c14\u6ce1\u6bd4\u7387\u4e3a PP-1 V * M \u3002\u6b64\u5916\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u5f02\u6b65\u70b9\u5bf9\u70b9\u901a\u4fe1\uff0c\u8fd9\u663e\u8457\u52a0\u5feb\u4e86\u8bad\u7ec3\u901f\u5ea6\uff0c\u5c24\u5176\u662f\u5728\u6587\u6863\u63a9\u7801\u5f15\u5165\u989d\u5916\u8ba1\u7b97\u4e0d\u5e73\u8861\u7684\u60c5\u51b5\u4e0b\u3002\u6211\u4eec\u542f\u7528 TORCH_NCCL_AVOID_RECORD_STREAMS \u6765\u51cf\u5c11\u6765\u81ea\u5f02\u6b65\u70b9\u5bf9\u70b9\u901a\u4fe1\u7684\u5185\u5b58\u4f7f\u7528\u91cf\u3002\u6700\u540e\uff0c\u4e3a\u4e86\u964d\u4f4e\u5185\u5b58\u6210\u672c\uff0c\u57fa\u4e8e\u8be6\u7ec6\u7684\u5185\u5b58\u5206\u914d\u5206\u6790\uff0c\u6211\u4eec\u4e3b\u52a8\u91ca\u653e\u4e0d\u4f1a\u7528\u4e8e\u672a\u6765\u8ba1\u7b97\u7684\u5f20\u91cf\uff0c\u5305\u62ec\u6bcf\u4e2a\u7ba1\u9053\u9636\u6bb5\u7684\u8f93\u5165\u548c\u8f93\u51fa\u5f20\u91cf\u3002** \u901a\u8fc7\u8fd9\u4e9b\u4f18\u5316\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u4e0d\u4f7f\u7528\u6fc0\u6d3b\u68c0\u67e5\u70b9\u7684\u60c5\u51b5\u4e0b\uff0c\u5bf9 8K \u4e2a <a href=\"https:\/\/www.kdjingpai.com\/pt\/tokenization\/\">token<\/a> \u7684\u5e8f\u5217\u8fdb\u884c Llama 3 \u7684\u9884\u8bad\u7ec3\u3002<\/p>\n<p><strong>\u4e0a\u4e0b\u6587\u5e76\u884c\u5316\u7528\u4e8e\u957f\u5e8f\u5217\u3002<\/strong>\u00a0\u6211\u4eec\u5229\u7528\u4e0a\u4e0b\u6587\u5e76\u884c\u5316 (CP) \u6765\u63d0\u9ad8\u5728\u6269\u5c55 Llama 3 \u4e0a\u4e0b\u6587\u957f\u5ea6\u65f6\u7684\u5185\u5b58\u6548\u7387\uff0c\u5e76\u5141\u8bb8\u5728\u957f\u8fbe 128K \u7684\u6781\u957f\u5e8f\u5217\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u5728 CP \u4e2d\uff0c\u6211\u4eec\u8de8\u5e8f\u5217\u7ef4\u5ea6\u8fdb\u884c\u5206\u533a\uff0c\u5177\u4f53\u6765\u8bf4\u6211\u4eec\u5c06\u8f93\u5165\u5e8f\u5217\u5212\u5206\u4e3a 2 \u00d7 CP \u4e2a\u5757\uff0c\u4ee5\u4fbf\u6bcf\u4e2a CP \u7b49\u7ea7\u63a5\u6536\u4e24\u4e2a\u5757\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u8d1f\u8f7d\u5e73\u8861\u3002\u7b2c i \u4e2a CP \u7b49\u7ea7\u63a5\u6536\u5230\u7b2c i \u4e2a\u548c (2 \u00d7 CP -1 -i) \u4e2a\u5757\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4e0d\u540c\u4e8e\u73b0\u6709\u7684\u5728\u73af\u5f62\u7ed3\u6784\u4e2d\u91cd\u53e0\u901a\u4fe1\u548c\u8ba1\u7b97\u7684 CP \u5b9e\u73b0(Liu et al., 2023a)\uff0c\u6211\u4eec\u7684 CP \u5b9e\u73b0\u91c7\u7528\u4e86\u4e00\u79cd\u57fa\u4e8e all-gather \u7684\u65b9\u6cd5\uff0c\u9996\u5148\u5c06\u952e\u503c (K, V) \u5f20\u91cf\u8fdb\u884c\u5168\u5c40\u805a\u5408\uff0c\u7136\u540e\u8ba1\u7b97\u5c40\u90e8\u67e5\u8be2 (Q) \u5f20\u91cf\u5757\u7684\u6ce8\u610f\u529b\u8f93\u51fa\u3002\u5c3d\u7ba1 all-gather \u901a\u4fe1\u5ef6\u8fdf\u5904\u4e8e\u5173\u952e\u8def\u5f84\u4e0a\uff0c\u4f46\u6211\u4eec\u4ecd\u7136\u91c7\u7528\u8fd9\u79cd\u65b9\u6cd5\u6709\u4e24\u4e2a\u4e3b\u8981\u539f\u56e0\uff1a<\/p>\n<p>(1) \u5b83\u66f4\u5bb9\u6613\u66f4\u7075\u6d3b\u5730\u652f\u6301\u57fa\u4e8e all-gather \u7684 CP \u6ce8\u610f\u529b\u4e2d\u7684\u4e0d\u540c\u7c7b\u578b\u6ce8\u610f\u63a9\u7801\uff0c\u4f8b\u5982\u6587\u6863\u63a9\u7801\uff1b<\/p>\n<p>(2) \u66b4\u9732\u7684 all-gather \u5ef6\u8fdf\u5f88\u5c0f\uff0c\u56e0\u4e3a\u901a\u4fe1\u7684 K \u548c V \u5f20\u91cf\u8fdc\u5c0f\u4e8e Q \u5f20\u91cf\uff0c\u8fd9\u662f\u7531\u4e8e\u4f7f\u7528\u4e86 GQA (Ainslie et al., 2023)\u3002\u56e0\u6b64\uff0c\u6ce8\u610f\u529b\u8ba1\u7b97\u7684\u65f6\u95f4\u590d\u6742\u5ea6\u6bd4 all-gather \u5927\u4e00\u4e2a\u6570\u91cf\u7ea7\uff08O(S\u00b2) \u5bf9 O(S)\uff0c\u5176\u4e2d S \u8868\u793a\u5b8c\u5168\u56e0\u679c\u63a9\u7801\u4e2d\u7684\u5e8f\u5217\u957f\u5ea6\uff09\uff0c\u4f7f\u5f97 all-gather \u5f00\u9500\u53ef\u4ee5\u5ffd\u7565\u4e0d\u8ba1\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3440\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/347e2014d12ad2a.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"845\" height=\"291\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/347e2014d12ad2a.png 1290w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/347e2014d12ad2a-300x103.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/347e2014d12ad2a-1024x352.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/347e2014d12ad2a-768x264.png 768w\" sizes=\"auto, (max-width: 845px) 100vw, 845px\" \/><\/strong><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u7f51\u7edc\u611f\u77e5\u5e76\u884c\u5316\u914d\u7f6e\u3002<\/strong>\u5e76\u884c\u5316\u7ef4\u5ea6\u7684\u987a\u5e8f [TP, CP, PP, DP] \u662f\u9488\u5bf9\u7f51\u7edc\u901a\u4fe1\u8fdb\u884c\u4f18\u5316\u7684\u3002\u6700\u5185\u5c42\u7684\u5e76\u884c\u5316\u9700\u8981\u6700\u9ad8\u7684\u7f51\u7edc\u5e26\u5bbd\u548c\u6700\u4f4e\u7684\u5ef6\u8fdf\uff0c\u56e0\u6b64\u901a\u5e38\u88ab\u9650\u5236\u5728\u540c\u4e00\u670d\u52a1\u5668\u5185\u90e8\u3002\u6700\u5916\u5c42\u7684\u5e76\u884c\u5316\u53ef\u80fd\u8de8\u8d8a\u591a\u8df3\u7f51\u7edc\uff0c\u5e94\u80fd\u591f\u5bb9\u5fcd\u66f4\u9ad8\u7684\u7f51\u7edc\u5ef6\u8fdf\u3002\u56e0\u6b64\uff0c\u57fa\u4e8e\u5bf9\u7f51\u7edc\u5e26\u5bbd\u548c\u5ef6\u8fdf\u7684\u8981\u6c42\uff0c\u6211\u4eec\u5c06\u5e76\u884c\u5316\u7ef4\u5ea6\u6309\u7167 [TP, CP, PP, DP] \u7684\u987a\u5e8f\u6392\u5217\u3002DP (\u5373 FSDP) \u662f\u6700\u5916\u5c42\u5e76\u884c\u5316\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u901a\u8fc7\u5f02\u6b65\u9884\u53d6\u5206\u7247\u6a21\u578b\u6743\u91cd\u548c\u51cf\u5c11\u68af\u5ea6\u6765\u5fcd\u53d7\u66f4\u957f\u7684\u7f51\u7edc\u5ef6\u8fdf\u3002\u786e\u5b9a\u5177\u6709\u6700\u5c0f\u901a\u4fe1\u5f00\u9500\u7684\u6700\u4f73\u5e76\u884c\u5316\u914d\u7f6e\uff0c\u540c\u65f6\u907f\u514d GPU \u5185\u5b58\u6ea2\u51fa\u662f\u4e00\u4e2a\u6311\u6218\u3002\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u5185\u5b58\u6d88\u8017\u4f30\u8ba1\u5668\u548c\u4e00\u4e2a\u6027\u80fd\u6295\u5f71\u5de5\u5177\uff0c\u8fd9\u6709\u52a9\u4e8e\u6211\u4eec\u63a2\u7d22\u5404\u79cd\u5e76\u884c\u5316\u914d\u7f6e\uff0c\u5e76\u9884\u6d4b\u6574\u4f53\u8bad\u7ec3\u6027\u80fd\u548c\u6709\u6548\u5730\u8bc6\u522b\u5185\u5b58\u5dee\u8ddd\u3002<\/p>\n<p><strong>\u6570\u503c\u7a33\u5b9a\u6027\u3002<\/strong>\u901a\u8fc7\u6bd4\u8f83\u4e0d\u540c\u5e76\u884c\u8bbe\u7f6e\u4e4b\u95f4\u7684\u8bad\u7ec3\u635f\u5931\uff0c\u6211\u4eec\u4fee\u590d\u4e86\u4e00\u4e9b\u5f71\u54cd\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u6570\u503c\u95ee\u9898\u3002\u4e3a\u4e86\u786e\u4fdd\u8bad\u7ec3\u6536\u655b\uff0c\u6211\u4eec\u5728\u591a\u4e2a\u5fae\u6279\u6b21\u7684\u53cd\u5411\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u4f7f\u7528 FP32 \u68af\u5ea6\u7d2f\u79ef\uff0c\u5e76\u5728 FSDP \u4e2d\u7684\u6570\u636e\u5e76\u884c\u5de5\u4f5c\u5668\u4e4b\u95f4\u4f7f\u7528 FP32 \u8fdb\u884c reduce-scatter \u68af\u5ea6\u3002\u5bf9\u4e8e\u5728\u6b63\u5411\u8ba1\u7b97\u4e2d\u591a\u6b21\u4f7f\u7528\u7684\u4e2d\u95f4\u5f20\u91cf\uff0c\u4f8b\u5982\u89c6\u89c9\u7f16\u7801\u5668\u8f93\u51fa\uff0c\u53cd\u5411\u68af\u5ea6\u4e5f\u4f1a\u4ee5 FP32 \u7d2f\u79ef\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3.3.3 \u96c6\u4f53\u901a\u4fe1<\/h3>\n<p>Llama 3 \u7684\u96c6\u4f53\u901a\u4fe1\u5e93\u57fa\u4e8e Nvidia NCCL \u5e93\u7684\u5206\u652f\uff0c\u79f0\u4e3a NCCLX\u3002NCCLX \u6781\u5927\u5730\u63d0\u9ad8\u4e86 NCCL \u7684\u6027\u80fd\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u9ad8\u5ef6\u8fdf\u7f51\u7edc\u800c\u8a00\u3002\u56de\u60f3\u4e00\u4e0b\uff0c\u5e76\u884c\u7ef4\u5ea6\u7684\u987a\u5e8f\u662f [TP, CP, PP, DP]\uff0c\u5176\u4e2d DP \u5bf9\u5e94\u4e8e FSDP\u3002\u6700\u5916\u5c42\u7684\u5e76\u884c\u7ef4\u5ea6 PP \u548c DP \u53ef\u80fd\u901a\u8fc7\u591a\u8df3\u7f51\u7edc\u8fdb\u884c\u901a\u4fe1\uff0c\u5ef6\u8fdf\u9ad8\u8fbe\u6570\u5341\u5fae\u79d2\u3002\u539f\u59cb NCCL \u7684\u96c6\u4f53\u901a\u4fe1\u64cd\u4f5c all-gather \u548c reduce-scatter \u5728 FSDP \u4e2d\u4f7f\u7528\uff0c\u800c point-to-point \u901a\u4fe1\u5219\u7528\u4e8e PP\uff0c\u8fd9\u4e9b\u90fd\u9700\u8981\u6570\u636e\u5206\u5757\u548c\u5206\u9636\u6bb5\u7684\u6570\u636e\u590d\u5236\u3002\u8fd9\u79cd\u65b9\u6cd5\u4f1a\u5bfc\u81f4\u4ee5\u4e0b\u4e00\u4e9b\u4f4e\u6548\u7387\u95ee\u9898\uff1a<\/p>\n<ol>\n<li>\u9700\u8981\u901a\u8fc7\u7f51\u7edc\u4ea4\u6362\u5927\u91cf\u5c0f\u63a7\u5236\u6d88\u606f\u4ee5\u4fc3\u8fdb\u6570\u636e\u4f20\u8f93\uff1b<\/li>\n<li>\u989d\u5916\u7684\u5185\u5b58\u590d\u5236\u64cd\u4f5c\uff1b<\/li>\n<li>\u4f7f\u7528\u989d\u5916\u7684 GPU \u5468\u671f\u8fdb\u884c\u901a\u4fe1\u3002<\/li>\n<\/ol>\n<p>\u5bf9\u4e8e Llama 3 \u8bad\u7ec3\uff0c\u6211\u4eec\u901a\u8fc7\u8c03\u6574\u5206\u5757\u548c\u6570\u636e\u4f20\u8f93\u4ee5\u9002\u5e94\u6211\u4eec\u7684\u7f51\u7edc\u5ef6\u8fdf\u6765\u89e3\u51b3\u5176\u4e2d\u4e00\u90e8\u5206\u4f4e\u6548\u7387\u95ee\u9898\uff0c\u8fd9\u79cd\u5ef6\u8fdf\u5728\u5927\u578b\u96c6\u7fa4\u4e2d\u53ef\u80fd\u9ad8\u8fbe\u6570\u5341\u5fae\u79d2\u3002\u6211\u4eec\u8fd8\u5141\u8bb8\u5c0f\u63a7\u5236\u6d88\u606f\u4ee5\u66f4\u9ad8\u7684\u4f18\u5148\u7ea7\u7a7f\u8fc7\u6211\u4eec\u7684\u7f51\u7edc\uff0c\u7279\u522b\u907f\u514d\u5728\u6df1\u5ea6\u7f13\u51b2\u6838\u5fc3\u4ea4\u6362\u673a\u4e2d\u51fa\u73b0\u961f\u9996\u963b\u585e\u3002<\/p>\n<p>\u6211\u4eec\u6b63\u5728\u8fdb\u884c\u7684\u5de5\u4f5c\uff0c\u7528\u4e8e\u672a\u6765\u7684 Llama \u7248\u672c\uff0c\u5305\u62ec\u5bf9 NCCLX \u8fdb\u884c\u66f4\u6df1\u5c42\u6b21\u7684\u4fee\u6539\uff0c\u4ee5\u5168\u9762\u89e3\u51b3\u4e0a\u8ff0\u6240\u6709\u95ee\u9898\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3441\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18558645a5c1e9d.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"848\" height=\"522\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18558645a5c1e9d.png 985w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18558645a5c1e9d-300x185.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18558645a5c1e9d-768x473.png 768w\" sizes=\"auto, (max-width: 848px) 100vw, 848px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>3.3.4 \u53ef\u9760\u6027\u548c\u8fd0\u884c\u6311\u6218<\/h3>\n<p>16K GPU \u8bad\u7ec3\u7684\u590d\u6742\u6027\u548c\u6f5c\u5728\u6545\u969c\u573a\u666f\u8d85\u8fc7\u4e86\u6211\u4eec\u64cd\u4f5c\u8fc7\u7684\u66f4\u5927\u7684 CPU \u96c6\u7fa4\u3002\u6b64\u5916\uff0c\u8bad\u7ec3\u7684\u540c\u6b65\u6027\u8d28\u4f7f\u5176\u5bb9\u9519\u6027\u8f83\u4f4e\u2014\u2014\u5355\u4e2a GPU \u6545\u969c\u53ef\u80fd\u9700\u8981\u91cd\u65b0\u542f\u52a8\u6574\u4e2a\u4f5c\u4e1a\u3002\u5c3d\u7ba1\u9762\u4e34\u8fd9\u4e9b\u6311\u6218\uff0c\u4f46\u5bf9\u4e8e Llama 3\uff0c\u6211\u4eec\u5728\u652f\u6301\u81ea\u52a8\u96c6\u7fa4\u7ef4\u62a4\uff08\u4f8b\u5982\u56fa\u4ef6\u548c Linux \u5185\u6838\u5347\u7ea7 (Vigraham \u548c Leonhardi\uff0c2024)\uff09\u7684\u540c\u65f6\uff0c\u5b9e\u73b0\u4e86\u9ad8\u4e8e 90% \u7684\u6709\u6548\u8bad\u7ec3\u65f6\u95f4\uff0c\u8fd9\u5bfc\u81f4\u4e86\u6bcf\u5929\u81f3\u5c11\u4e00\u6b21\u8bad\u7ec3\u4e2d\u65ad\u3002<\/p>\n<p>\u6709\u6548\u8bad\u7ec3\u65f6\u95f4\u662f\u6307\u5728\u7ecf\u8fc7\u7684\u65f6\u95f4\u5185\u7528\u4e8e\u6709\u6548\u8bad\u7ec3\u7684\u65f6\u95f4\u3002 \u5728\u4e3a\u671f 54 \u5929\u7684\u9884\u8bad\u7ec3\u5feb\u7167\u671f\u95f4\uff0c\u6211\u4eec\u7ecf\u5386\u4e86\u603b\u5171 466 \u6b21\u4f5c\u4e1a\u4e2d\u65ad\u3002\u5176\u4e2d 47 \u6b21\u662f\u7531\u4e8e\u81ea\u52a8\u7ef4\u62a4\u64cd\u4f5c\uff08\u4f8b\u5982\u56fa\u4ef6\u5347\u7ea7\u6216\u64cd\u4f5c\u5458\u542f\u52a8\u7684\u64cd\u4f5c\uff0c\u5982\u914d\u7f6e\u6216\u6570\u636e\u96c6\u66f4\u65b0\uff09\u800c\u8ba1\u5212\u7684\u4e2d\u65ad\u3002\u5176\u4f59 419 \u6b21\u662f\u672a\u9884\u671f\u4e2d\u65ad\uff0c\u8fd9\u4e9b\u4e2d\u65ad\u6309\u8868 5 \u5206\u7c7b\u3002\u5927\u7ea6 78% \u7684\u610f\u5916\u4e2d\u65ad\u5f52\u56e0\u4e8e\u5df2\u786e\u8ba4\u7684\u786c\u4ef6\u95ee\u9898\uff0c\u4f8b\u5982 GPU \u6216\u4e3b\u673a\u7ec4\u4ef6\u6545\u969c\uff0c\u6216\u8005\u6000\u7591\u4e0e\u786c\u4ef6\u76f8\u5173\u7684\u95ee\u9898\uff0c\u4f8b\u5982\u65e0\u58f0\u6570\u636e\u635f\u574f\u548c\u8ba1\u5212\u5916\u4e2a\u522b\u4e3b\u673a\u7ef4\u62a4\u4e8b\u4ef6\u3002GPU \u95ee\u9898\u662f\u6700\u5927\u7c7b\u522b\uff0c\u5360\u6240\u6709\u610f\u5916\u95ee\u9898\u7684 58.7%\u3002\u5c3d\u7ba1\u5b58\u5728\u5927\u91cf\u6545\u969c\uff0c\u4f46\u5728\u8fd9\u6bb5\u65f6\u95f4\u5185\uff0c\u4ec5\u9700\u8981\u8fdb\u884c\u4e09\u6b21\u91cd\u5927\u624b\u52a8\u5e72\u9884\uff0c\u5176\u4f59\u95ee\u9898\u90fd\u901a\u8fc7\u81ea\u52a8\u5316\u5904\u7406\u3002<\/p>\n<p>\u4e3a\u4e86\u63d0\u9ad8\u6709\u6548\u8bad\u7ec3\u65f6\u95f4\uff0c\u6211\u4eec\u51cf\u5c11\u4e86\u4f5c\u4e1a\u542f\u52a8\u548c\u68c0\u67e5\u70b9\u65f6\u95f4\uff0c\u5e76\u5f00\u53d1\u4e86\u7528\u4e8e\u5feb\u901f\u8bca\u65ad\u548c\u89e3\u51b3\u95ee\u9898\u7684\u5de5\u5177\u3002 \u6211\u4eec\u5e7f\u6cdb\u4f7f\u7528\u4e86 PyTorch \u5185\u7f6e\u7684 NCCL Flight Recorder (Ansel \u7b49\u4eba\uff0c2024)\u2014\u2014\u6b64\u529f\u80fd\u5c06\u96c6\u4f53\u5143\u6570\u636e\u548c\u5806\u6808\u8ddf\u8e2a\u6355\u83b7\u5230\u73af\u5f62\u7f13\u51b2\u533a\u4e2d\uff0c\u4ece\u800c\u4f7f\u6211\u4eec\u80fd\u591f\u5feb\u901f\u5728\u5927\u89c4\u6a21\u4e0b\u8bca\u65ad\u6302\u8d77\u548c\u6027\u80fd\u95ee\u9898\uff0c\u7279\u522b\u662f\u5728 NCCLX \u65b9\u9762\u3002\u4f7f\u7528\u5b83\uff0c\u6211\u4eec\u53ef\u4ee5\u6709\u6548\u5730\u8bb0\u5f55\u6bcf\u4e2a\u96c6\u4f53\u64cd\u4f5c\u7684\u901a\u4fe1\u4e8b\u4ef6\u548c\u6301\u7eed\u65f6\u95f4\uff0c\u5e76\u5728 NCCLX \u770b\u95e8\u72d7\u6216\u5fc3\u8df3\u8d85\u65f6\u65f6\u81ea\u52a8\u8f6c\u50a8\u8ddf\u8e2a\u6570\u636e\u3002\u901a\u8fc7\u5728\u7ebf\u914d\u7f6e\u66f4\u6539 (Tang \u7b49\u4eba\uff0c2015)\uff0c\u6211\u4eec\u53ef\u4ee5\u9009\u62e9\u6027\u5730\u542f\u7528\u66f4\u8ba1\u7b97\u5bc6\u96c6\u7684\u8ddf\u8e2a\u64cd\u4f5c\u548c\u5143\u6570\u636e\u6536\u96c6\uff0c\u800c\u65e0\u9700\u4ee3\u7801\u53d1\u5e03\u6216\u4f5c\u4e1a\u91cd\u542f\u3002 \u7531\u4e8e\u6211\u4eec\u7684\u7f51\u7edc\u4e2d\u6df7\u5408\u4f7f\u7528\u4e86 NVLink \u548c RoCE\uff0c\u56e0\u6b64\u5728\u5927\u578b\u8bad\u7ec3\u4e2d\u8c03\u8bd5\u95ee\u9898\u53d8\u5f97\u590d\u6742\u3002\u6570\u636e\u4f20\u8f93\u901a\u5e38\u901a\u8fc7 CUDA \u5185\u6838\u53d1\u51fa\u7684\u52a0\u8f7d\/\u5b58\u50a8\u64cd\u4f5c\u8fdb\u884c\u5728 NVLink \u4e0a\u8fdb\u884c\uff0c\u8fdc\u7a0b GPU \u6216 NVLink \u8fde\u63a5\u5931\u8d25\u7ecf\u5e38\u4f1a\u8868\u73b0\u4e3a CUDA \u5185\u6838\u4e2d\u7684\u52a0\u8f7d\/\u5b58\u50a8\u64cd\u4f5c\u505c\u6ede\u800c\u4e0d\u4f1a\u8fd4\u56de\u660e\u786e\u7684\u9519\u8bef\u4ee3\u7801\u3002NCCLX \u901a\u8fc7\u4e0e PyTorch \u7d27\u5bc6\u7684\u8bbe\u8ba1\uff0c\u63d0\u9ad8\u4e86\u6545\u969c\u68c0\u6d4b\u548c\u5b9a\u4f4d\u7684\u901f\u5ea6\u548c\u51c6\u786e\u6027\uff0c\u5141\u8bb8 PyTorch \u8bbf\u95ee NCCLX \u7684\u5185\u90e8\u72b6\u6001\u5e76\u8ddf\u8e2a\u76f8\u5173\u4fe1\u606f\u3002\u867d\u7136\u65e0\u6cd5\u5b8c\u5168\u9632\u6b62\u7531\u4e8e NVLink \u6545\u969c\u5bfc\u81f4\u7684\u505c\u6ede\uff0c\u4f46\u6211\u4eec\u7684\u7cfb\u7edf\u4f1a\u76d1\u89c6\u901a\u4fe1\u5e93\u7684\u72b6\u6001\uff0c\u5e76\u5728\u68c0\u6d4b\u5230\u6b64\u7c7b\u505c\u6ede\u65f6\u81ea\u52a8\u8d85\u65f6\u3002\u6b64\u5916\uff0cNCCLX \u8fd8\u4f1a\u8ddf\u8e2a\u6bcf\u4e2a NCCLX \u901a\u4fe1\u7684\u5185\u6838\u548c\u7f51\u7edc\u6d3b\u52a8\uff0c\u5e76\u63d0\u4f9b\u6b63\u5728\u5931\u8d25\u7684 NCCLX \u96c6\u4f53\u7684\u5185\u90e8\u72b6\u6001\u5feb\u7167\uff0c\u5305\u62ec\u6240\u6709\u79e9\u4e4b\u95f4\u5df2\u5b8c\u6210\u548c\u672a\u5b8c\u6210\u7684\u6570\u636e\u4f20\u8f93\u3002\u6211\u4eec\u5206\u6790\u8fd9\u4e9b\u6570\u636e\u6765\u8c03\u8bd5 NCCLX \u6269\u5c55\u95ee\u9898\u3002<\/p>\n<p>\u6709\u65f6\uff0c\u786c\u4ef6\u95ee\u9898\u4f1a\u5bfc\u81f4\u4ecd\u7136\u53ef\u4ee5\u8fd0\u884c\u4f46\u901f\u5ea6\u6162\u7684 stragglers\uff0c\u8fd9\u4e9b stragglers \u5f88\u96be\u4ee5\u68c0\u6d4b\u5230\u3002\u5373\u4f7f\u53ea\u6709\u4e00\u4e2a straggler\uff0c\u5b83\u4e5f\u4f1a\u51cf\u6162\u6570\u5343\u4e2a\u5176\u4ed6 GPU \u7684\u901f\u5ea6\uff0c\u901a\u5e38\u8868\u73b0\u4e3a\u6b63\u5e38\u8fd0\u884c\u4f46\u901a\u4fe1\u7f13\u6162\u3002 \u6211\u4eec\u5f00\u53d1\u4e86\u5de5\u5177\u6765\u4f18\u5148\u5904\u7406\u6765\u81ea\u9009\u5b9a\u8fdb\u7a0b\u7ec4\u7684\u6f5c\u5728\u95ee\u9898\u901a\u4fe1\u3002\u901a\u8fc7\u4ec5\u8c03\u67e5\u5c11\u6570\u4e3b\u8981\u5acc\u7591\u72af\uff0c\u6211\u4eec\u901a\u5e38\u80fd\u591f\u6709\u6548\u5730\u8bc6\u522b\u51fa stragglers\u3002<\/p>\n<p>\u4e00\u4e2a\u6709\u8da3\u7684\u89c2\u5bdf\u662f\u73af\u5883\u56e0\u7d20\u5bf9\u5927\u89c4\u6a21\u8bad\u7ec3\u6027\u80fd\u7684\u5f71\u54cd\u3002\u5bf9\u4e8e Llama 3 405B\uff0c\u6211\u4eec\u6ce8\u610f\u5230\u57fa\u4e8e\u65f6\u95f4\u53d8\u5316\u7684 1-2% \u7684\u541e\u5410\u91cf\u6ce2\u52a8\u3002\u8fd9\u79cd\u6ce2\u52a8\u662f\u7531\u4e8e\u66f4\u9ad8\u7684\u4e2d\u5348\u6e29\u5ea6\u5f71\u54cd GPU \u52a8\u6001\u7535\u538b\u548c\u9891\u7387\u7f29\u653e\u9020\u6210\u7684\u3002 \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6570\u4e07\u4e2a GPU \u53ef\u80fd\u4f1a\u540c\u65f6\u589e\u52a0\u6216\u51cf\u5c11\u529f\u8017\uff0c\u4f8b\u5982\u7531\u4e8e\u6240\u6709 GPU \u6b63\u5728\u7b49\u5f85\u68c0\u67e5\u70b9\u6216\u96c6\u4f53\u901a\u4fe1\u7ed3\u675f\uff0c\u6216\u8005\u6574\u4e2a\u8bad\u7ec3\u4f5c\u4e1a\u7684\u542f\u52a8\u6216\u5173\u95ed\u3002\u5f53\u53d1\u751f\u8fd9\u79cd\u60c5\u51b5\u65f6\uff0c\u5b83\u4f1a\u5bfc\u81f4\u6570\u636e\u4e2d\u5fc3\u5185\u7684\u529f\u8017\u77ac\u65f6\u6ce2\u52a8\u8fbe\u5230\u6570\u5341\u5146\u74e6\u7684\u5e45\u5ea6\uff0c\u8fd9\u5c06\u62c9\u4f38\u7535\u7f51\u7684\u6781\u9650\u3002\u968f\u7740\u6211\u4eec\u4e3a\u672a\u6765\u751a\u81f3\u66f4\u5927\u7684 Llama \u6a21\u578b\u6269\u5c55\u8bad\u7ec3\u89c4\u6a21\uff0c\u8fd9\u662f\u4e00\u4e2a\u6301\u7eed\u9762\u4e34\u7684\u6311\u6218\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>3.4 \u8bad\u7ec3\u65b9\u6848<\/h2>\n<p>&nbsp;<\/p>\n<p>Llama 3 405B \u7684\u9884\u8bad\u7ec3\u98df\u8c31\u5305\u542b\u4e09\u4e2a\u4e3b\u8981\u9636\u6bb5\uff1a<\/p>\n<p>(1) \u521d\u59cb\u9884\u8bad\u7ec3\uff0c(2) \u957f\u4e0a\u4e0b\u6587\u9884\u8bad\u7ec3\u548c (3) \u9000\u706b\u3002\u4ee5\u4e0b\u5206\u522b\u63cf\u8ff0\u8fd9\u4e09\u4e2a\u9636\u6bb5\u3002\u6211\u4eec\u4f7f\u7528\u7c7b\u4f3c\u7684\u98df\u8c31\u6765\u9884\u8bad\u7ec3 8B \u548c 70B \u6a21\u578b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3.4.1 \u521d\u59cb\u9884\u8bad\u7ec3<\/h3>\n<p>\u6211\u4eec\u4f7f\u7528\u4f59\u5f26\u5b66\u4e60\u7387\u8ba1\u5212\u9884\u8bad\u7ec3Llama 3 405B\u6a21\u578b\uff0c\u6700\u9ad8\u5b66\u4e60\u7387\u4e3a8 \u00d7 10\u207b\u2075\uff0c\u7ebf\u6027\u9884\u70ed8,000\u6b65\uff0c\u7ecf\u8fc71,200,000\u6b65\u8bad\u7ec3\u540e\u8870\u51cf\u81f38 \u00d7 10\u207b\u2077\u3002\u4e3a\u4e86\u63d0\u9ad8\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\uff0c\u6211\u4eec\u5728\u8bad\u7ec3\u521d\u671f\u4f7f\u7528\u8f83\u5c0f\u7684\u6279\u91cf\u5927\u5c0f\uff0c\u968f\u540e\u589e\u52a0\u6279\u91cf\u5927\u5c0f\u4ee5\u63d0\u9ad8\u6548\u7387\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u6700\u521d\u7684\u6279\u91cf\u5927\u5c0f\u4e3a4M\u4e2atokens\uff0c\u5e8f\u5217\u957f\u5ea6\u4e3a4,096\u3002\u5728\u9884\u8bad\u7ec3\u4e86252M\u4e2atokens\u540e\uff0c\u6211\u4eec\u5c06\u6279\u91cf\u5927\u5c0f\u548c\u5e8f\u5217\u957f\u5ea6\u5206\u522b\u52a0\u500d\u81f38M\u4e2a\u5e8f\u5217\u30018,192\u4e2atokens\u3002\u5728\u9884\u8bad\u7ec3\u4e862.87T\u4e2atokens\u540e\uff0c\u6211\u4eec\u518d\u6b21\u5c06\u6279\u91cf\u5927\u5c0f\u52a0\u500d\u81f316M\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u79cd\u8bad\u7ec3\u65b9\u6cd5\u975e\u5e38\u7a33\u5b9a\uff1a\u5f88\u5c11\u51fa\u73b0\u635f\u5931\u5cf0\u503c\uff0c\u5e76\u4e14\u4e0d\u9700\u8981\u8fdb\u884c\u5e72\u9884\u6765\u7ea0\u6b63\u6a21\u578b\u8bad\u7ec3\u7684\u504f\u79bb\u3002<\/p>\n<p>\u8c03\u6574\u6570\u636e\u7ec4\u5408\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5bf9\u9884\u8bad\u7ec3\u6570\u636e\u7ec4\u5408\u8fdb\u884c\u4e86\u51e0\u6b21\u8c03\u6574\uff0c\u4ee5\u6539\u5584\u6a21\u578b\u5728\u7279\u5b9a\u4e0b\u6e38\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002\u7279\u522b\u662f\uff0c\u6211\u4eec\u5728\u9884\u8bad\u7ec3\u671f\u95f4\u589e\u52a0\u4e86\u975e\u82f1\u8bed\u6570\u636e\u7684\u6bd4\u4f8b\uff0c\u4ee5\u63d0\u9ad8Llama 3\u7684\u591a\u8bed\u8a00\u8868\u73b0\u3002\u6211\u4eec\u8fd8\u4e0a\u8c03\u4e86\u6570\u5b66\u6570\u636e\u7684\u6bd4\u4f8b\uff0c\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u6570\u5b66\u63a8\u7406\u80fd\u529b\uff0c\u5728\u9884\u8bad\u7ec3\u7684\u540e\u671f\u9636\u6bb5\u52a0\u5165\u4e86\u66f4\u591a\u7684\u6700\u65b0\u7f51\u7edc\u6570\u636e\uff0c\u4ee5\u66f4\u65b0\u6a21\u578b\u7684\u77e5\u8bc6\u622a\u6b62\u70b9\uff0c\u540c\u65f6\u4e0b\u8c03\u4e86\u540e\u6765\u88ab\u8bc6\u522b\u4e3a\u8d28\u91cf\u8f83\u4f4e\u7684\u6570\u636e\u5b50\u96c6\u7684\u6bd4\u4f8b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3.4.2 \u957f\u4e0a\u4e0b\u6587\u9884\u8bad\u7ec3<\/h3>\n<p>\u5728\u9884\u8bad\u7ec3\u7684\u6700\u540e\u9636\u6bb5\uff0c\u6211\u4eec\u8bad\u7ec3\u957f\u5e8f\u5217\u4ee5\u652f\u6301\u9ad8\u8fbe 128,000 \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u6211\u4eec\u4e0d\u4f1a\u66f4\u65e9\u5730\u8bad\u7ec3\u957f\u5e8f\u5217\uff0c\u56e0\u4e3a\u81ea\u6ce8\u610f\u529b\u5c42\u4e2d\u7684\u8ba1\u7b97\u968f\u7740\u5e8f\u5217\u957f\u5ea6\u7684\u589e\u52a0\u800c\u4e8c\u6b21\u589e\u957f\u3002\u6211\u4eec\u9010\u6b65\u589e\u52a0\u652f\u6301\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u5e76\u5728\u6a21\u578b\u6210\u529f\u9002\u5e94\u589e\u52a0\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u540e\u8fdb\u884c\u9884\u8bad\u7ec3\u3002\u6211\u4eec\u901a\u8fc7\u6d4b\u91cf\u4ee5\u4e0b\u4e24\u4e2a\u65b9\u9762\u6765\u8bc4\u4f30\u6210\u529f\u7684\u9002\u5e94\uff1a<\/p>\n<p>(1) \u6a21\u578b\u5728\u77ed\u4e0a\u4e0b\u6587\u8bc4\u4f30\u4e2d\u7684\u6027\u80fd\u662f\u5426\u5df2\u5b8c\u5168\u6062\u590d\uff1b<\/p>\n<p>(2) \u6a21\u578b\u662f\u5426\u80fd\u591f\u5b8c\u7f8e\u5730\u89e3\u51b3\u957f\u8fbe\u8be5\u957f\u5ea6\u7684\u201c\u5927\u6d77\u635e\u9488\u201d\u4efb\u52a1\u3002\u5728 Llama 3 405B \u9884\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u9010\u6b65\u589e\u52a0\u4e86\u516d\u4e2a\u9636\u6bb5\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u4ece\u6700\u521d\u7684 8,000 \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u5f00\u59cb\uff0c\u6700\u7ec8\u8fbe\u5230 128,000 \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u8fd9\u4e2a\u957f\u4e0a\u4e0b\u6587\u9884\u8bad\u7ec3\u9636\u6bb5\u4f7f\u7528\u4e86\u5927\u7ea6 8000 \u4ebf\u4e2a\u8bad\u7ec3\u6807\u8bb0\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3442\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9c25bbc5c4f665d.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"764\" height=\"374\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9c25bbc5c4f665d.png 1224w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9c25bbc5c4f665d-300x147.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9c25bbc5c4f665d-1024x501.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9c25bbc5c4f665d-768x376.png 768w\" sizes=\"auto, (max-width: 764px) 100vw, 764px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>3.4.3 \u9000\u706b<\/h3>\n<p>\u5728\u6700\u540e 40M \u4e2a\u6807\u8bb0\u7684\u9884\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u5b66\u4e60\u7387\u7ebf\u6027\u9000\u706b\u81f3 0\uff0c\u540c\u65f6\u4fdd\u6301\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e3a 128K \u4e2a\u6807\u8bb0\u3002\u5728\u6b64\u9000\u706b\u9636\u6bb5\uff0c\u6211\u4eec\u8fd8\u8c03\u6574\u4e86\u6570\u636e\u6df7\u5408\uff0c\u4ee5\u589e\u52a0\u975e\u5e38\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u6e90\u7684\u6837\u672c\u91cf\uff1b\u8bf7\u53c2\u89c1\u7b2c 3.1.3 \u8282\u3002\u6700\u540e\uff0c\u6211\u4eec\u5728\u9000\u706b\u671f\u95f4\u8ba1\u7b97\u6a21\u578b\u68c0\u67e5\u70b9\u7684\u5e73\u5747\u503c\uff08Polyak (1991) \u5e73\u5747\uff09\u4ee5\u751f\u6210\u6700\u7ec8\u9884\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>4 \u540e\u7eed\u8bad\u7ec3<\/h2>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u901a\u8fc7\u5e94\u7528\u591a\u8f6e\u540e\u7eed\u8bad\u7ec3\u6765\u751f\u6210\u4e0e\u9f50 Llama 3 \u6a21\u578b\u3002\u8fd9\u4e9b\u540e\u7eed\u8bad\u7ec3\u57fa\u4e8e\u9884\u8bad\u7ec3\u7684\u68c0\u67e5\u70b9\uff0c\u5e76\u7ed3\u5408\u4eba\u7c7b\u53cd\u9988\u8fdb\u884c\u6a21\u578b\u5bf9\u9f50\uff08Ouyang \u7b49\u4eba\uff0c2022\uff1bRafailov \u7b49\u4eba\uff0c2024\uff09\u3002\u6bcf\u8f6e\u540e\u7eed\u8bad\u7ec3\u90fd\u5305\u62ec\u76d1\u7763\u5fae\u8c03 (SFT)\uff0c\u4e4b\u540e\u662f\u76f4\u63a5\u504f\u597d\u4f18\u5316 (DPO; Rafailov \u7b49\u4eba\uff0c2024)\uff0c\u4f7f\u7528\u901a\u8fc7\u4eba\u5de5\u6807\u6ce8\u6216\u5408\u6210\u751f\u6210\u7684\u793a\u4f8b\u8fdb\u884c\u3002\u6211\u4eec\u5728\u7b2c 4.1 \u8282\u548c\u7b2c 4.2 \u8282\u5206\u522b\u63cf\u8ff0\u4e86\u6211\u4eec\u7684\u540e\u7eed\u8bad\u7ec3\u5efa\u6a21\u548c\u6570\u636e\u65b9\u6cd5\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5c06\u5728\u7b2c 4.3 \u8282\u8fdb\u4e00\u6b65\u8be6\u7ec6\u4ecb\u7ecd\u5b9a\u5236\u7684\u6570\u636e\u6574\u7406\u7b56\u7565\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3001\u7f16\u7a0b\u80fd\u529b\u3001\u4e8b\u5b9e\u6027\u3001\u591a\u8bed\u8a00\u652f\u6301\u3001\u5de5\u5177\u4f7f\u7528\u3001\u957f\u4e0a\u4e0b\u6587\u4ee5\u53ca\u7cbe\u786e\u6307\u4ee4\u9075\u5faa\u7b49\u65b9\u9762\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>4.1 \u5efa\u6a21<\/h3>\n<p>\u6211\u4eec\u540e\u8bad\u7ec3\u7b56\u7565\u7684\u57fa\u7840\u662f\u5956\u52b1\u6a21\u578b\u548c\u8bed\u8a00\u6a21\u578b\u3002 \u6211\u4eec\u9996\u5148\u4f7f\u7528\u4eba\u7c7b\u6807\u6ce8\u7684\u504f\u597d\u6570\u636e\uff08\u89c1\u7b2c4.1.2\u8282\uff09\u5728\u9884\u8bad\u7ec3\u68c0\u67e5\u70b9\u4e4b\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\u3002\u7136\u540e\uff0c\u6211\u4eec\u7528\u76d1\u7763\u5fae\u8c03\uff08SFT\uff1b\u89c1\u7b2c4.1.3\u8282\uff09\u5fae\u8c03\u9884\u8bad\u7ec3\u68c0\u67e5\u70b9\uff0c\u5e76\u8fdb\u4e00\u6b65\u7528\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff1b\u89c1\u7b2c4.1.4\u8282\uff09\u4e0e\u68c0\u67e5\u70b9\u5bf9\u9f50\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u5982\u56fe7\u6240\u793a\u3002\u9664\u975e\u53e6\u6709\u8bf4\u660e\uff0c\u5426\u5219\u6211\u4eec\u7684\u5efa\u6a21\u8fc7\u7a0b\u9002\u7528\u4e8eLlama 3 405B\uff0c\u4e3a\u7b80\u4fbf\u8d77\u89c1\uff0c\u6211\u4eec\u5c06Llama 3 405B\u79f0\u4e3aLlama 3\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.1.1 \u804a\u5929\u5bf9\u8bdd\u683c\u5f0f<\/h4>\n<p>\u4e3a\u4e86\u8c03\u6574\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4ee5\u5b9e\u73b0\u4eba\u673a\u4ea4\u4e92\uff0c\u6211\u4eec\u9700\u8981\u5b9a\u4e49\u4e00\u4e2a\u804a\u5929\u5bf9\u8bdd\u534f\u8bae\uff0c\u8ba9\u6a21\u578b\u80fd\u591f\u7406\u89e3\u4eba\u7c7b\u6307\u4ee4\u5e76\u6267\u884c\u5bf9\u8bdd\u4efb\u52a1\u3002\u4e0e\u524d\u8eab\u76f8\u6bd4\uff0cLlama 3 \u5177\u6709\u65b0\u7684\u529f\u80fd\uff0c\u4f8b\u5982\u5de5\u5177\u4f7f\u7528\uff08\u7b2c4.3.5\u8282\uff09\uff0c\u8fd9\u53ef\u80fd\u9700\u8981\u5728\u5355\u4e2a\u5bf9\u8bdd\u56de\u5408\u4e2d\u751f\u6210\u591a\u6761\u6d88\u606f\u5e76\u5c06\u5b83\u4eec\u53d1\u9001\u5230\u4e0d\u540c\u7684\u4f4d\u7f6e\uff08\u4f8b\u5982\uff0c\u7528\u6237\u3001ipython\uff09\u3002\u4e3a\u4e86\u652f\u6301\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u65b0\u7684\u591a\u6d88\u606f\u804a\u5929\u534f\u8bae\uff0c\u8be5\u534f\u8bae\u4f7f\u7528\u5404\u79cd\u7279\u6b8a\u7684\u5934\u90e8\u548c\u7ec8\u6b62\u6807\u8bb0\u3002\u5934\u90e8\u6807\u8bb0\u7528\u4e8e\u6307\u793a\u5bf9\u8bdd\u4e2d\u6bcf\u6761\u6d88\u606f\u7684\u6765\u6e90\u548c\u76ee\u6807\u3002\u540c\u6837\uff0c\u7ec8\u6b62\u6807\u8bb0\u4e5f\u6307\u793a\u4f55\u65f6\u8f6e\u5230\u4eba\u7c7b\u548cAI\u4ea4\u66ff\u53d1\u8a00\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.1.2 \u5956\u52b1\u6a21\u578b<\/h4>\n<p>\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u8986\u76d6\u4e0d\u540c\u80fd\u529b\u7684\u5956\u52b1\u6a21\u578b (RM)\uff0c\u5e76\u5c06\u5176\u6784\u5efa\u5728\u9884\u8bad\u7ec3\u68c0\u67e5\u70b9\u4e4b\u4e0a\u3002 \u8bad\u7ec3\u76ee\u6807\u4e0e Llama 2 \u76f8\u540c\uff0c\u4f46\u6211\u4eec\u79fb\u9664\u4e86\u635f\u5931\u51fd\u6570\u4e2d\u7684\u8fb9\u9645\u9879\uff0c\u56e0\u4e3a\u6211\u4eec\u5728\u6570\u636e\u89c4\u6a21\u6269\u5927\u540e\u89c2\u5bdf\u5230\u6539\u8fdb\u6548\u679c\u51cf\u5c0f\u3002 \u4e0e Llama 2 \u4e00\u6837\uff0c\u6211\u4eec\u5728\u8fc7\u6ee4\u6389\u5177\u6709\u76f8\u4f3c\u54cd\u5e94\u7684\u6837\u672c\u540e\uff0c\u4f7f\u7528\u6240\u6709\u504f\u597d\u6570\u636e\u8fdb\u884c\u5956\u52b1\u5efa\u6a21\u3002<\/p>\n<p>\u9664\u4e86\u6807\u51c6\u7684 (\u88ab\u9009\u62e9, \u88ab\u62d2\u7edd) \u54cd\u5e94\u504f\u597d\u5bf9\u4e4b\u5916\uff0c\u6807\u6ce8\u8fd8\u4e3a\u4e00\u4e9b\u63d0\u793a\u521b\u5efa\u4e86\u7b2c\u4e09\u4e2a\u201c\u7f16\u8f91\u540e\u7684\u54cd\u5e94\u201d\uff0c\u5176\u4e2d\u4ece\u8be5\u5bf9\u4e2d\u9009\u62e9\u7684\u54cd\u5e94\u4f1a\u88ab\u8fdb\u4e00\u6b65\u7f16\u8f91\u4ee5\u6539\u8fdb (\u8bf7\u53c2\u89c1\u7b2c 4.2.1 \u8282)\u3002\u56e0\u6b64\uff0c\u6bcf\u4e2a\u504f\u597d\u6392\u5e8f\u6837\u672c\u5177\u6709\u4e24\u4e2a\u6216\u4e09\u4e2a\u54cd\u5e94\uff0c\u6392\u540d\u660e\u786e\uff08\u7f16\u8f91\u7248 &gt; \u9009\u62e9\u7248 &gt; \u62d2\u7edd\u7248\uff09\u3002 \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u63d0\u793a\u548c\u591a\u4e2a\u54cd\u5e94\u8fde\u63a5\u6210\u4e00\u884c\uff0c\u5e76\u968f\u673a\u6392\u5217\u54cd\u5e94\u3002\u8fd9\u662f\u4e00\u79cd\u5c06\u54cd\u5e94\u653e\u5728\u5355\u72ec\u7684\u884c\u4e2d\u8ba1\u7b97\u5206\u6570\u7684\u6807\u51c6\u573a\u666f\u7684\u8fd1\u4f3c\uff0c\u4f46\u5728\u6211\u4eec\u7684\u6d88\u878d\u5b9e\u9a8c\u4e2d\uff0c\u8fd9\u79cd\u65b9\u6cd5\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\uff0c\u800c\u4e0d\u4f1a\u635f\u5931\u7cbe\u5ea6\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.1.3 \u76d1\u7763\u5fae\u8c03<\/h4>\n<p>\u9996\u5148\u4f7f\u7528\u5956\u52b1\u6a21\u578b\u5bf9\u4eba\u7c7b\u6807\u6ce8\u63d0\u793a\u8fdb\u884c\u62d2\u7edd\u91c7\u6837\uff0c\u8be6\u7ec6\u65b9\u6cd5\u5c06\u5728\u7b2c4.2\u8282\u4e2d\u63cf\u8ff0\u3002\u6211\u4eec\u7ed3\u5408\u8fd9\u4e9b\u62d2\u7edd\u91c7\u6837\u7684\u6570\u636e\u548c\u5176\u4ed6\u6570\u636e\u6e90\uff08\u5305\u62ec\u5408\u6210\u6570\u636e\uff09\uff0c\u4f7f\u7528\u6807\u51c6\u4ea4\u53c9\u71b5\u635f\u5931\u5bf9\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u76ee\u6807\u662f\u9884\u6d4b\u76ee\u6807\u6807\u8bb0\uff08\u540c\u65f6\u5c4f\u853d\u63d0\u793a\u6807\u8bb0\u7684\u635f\u5931\uff09\u3002\u6709\u5173\u6570\u636e\u6df7\u5408\u7684\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\u8bf7\u53c2\u89c1\u7b2c 4.2 \u8282\u3002\u5c3d\u7ba1\u8bb8\u591a\u8bad\u7ec3\u76ee\u6807\u90fd\u662f\u6a21\u578b\u751f\u6210\u7684\uff0c\u6211\u4eec\u4ecd\u5c06\u6b64\u9636\u6bb5\u79f0\u4e3a\u76d1\u7763\u5fae\u8c03\uff08SFT\uff1b Wei \u7b49\u4eba\uff0c2022a\uff1b Sanh \u7b49\u4eba\uff0c2022\uff1b Wang \u7b49\u4eba\uff0c2022b\uff09\u3002<\/p>\n<p>\u6211\u4eec\u7684\u6700\u5927\u6a21\u578b\u5728 8.5K \u81f3 9K \u6b65\u5185\u4ee5 1e-5 \u7684\u5b66\u4e60\u7387\u8fdb\u884c\u5fae\u8c03\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u4e9b\u8d85\u53c2\u6570\u8bbe\u7f6e\u9002\u7528\u4e8e\u4e0d\u540c\u8f6e\u6b21\u548c\u6570\u636e\u6df7\u5408\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.1.4 \u76f4\u63a5\u504f\u597d\u4f18\u5316<\/h4>\n<p>\u6211\u4eec\u8fdb\u4e00\u6b65\u4f7f\u7528\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff1bRafailov \u7b49\u4eba\uff0c2024\uff09\u8bad\u7ec3\u6211\u4eec\u7684 SFT \u6a21\u578b\uff0c\u4ee5\u5b9e\u73b0\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u3002\u5728\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u4e3b\u8981\u4f7f\u7528\u4e0a\u4e00\u8f6e\u5bf9\u9f50\u4e2d\u8868\u73b0\u6700\u4f73\u7684\u6a21\u578b\u6536\u96c6\u5230\u7684\u6700\u65b0\u504f\u597d\u6570\u636e\u6279\u6b21\u3002\u7ed3\u679c\uff0c\u6211\u4eec\u7684\u8bad\u7ec3\u6570\u636e\u66f4\u7b26\u5408\u6bcf\u4e00\u8f6e\u4f18\u5316\u7684\u7b56\u7565\u6a21\u578b\u5206\u5e03\u3002\u6211\u4eec\u4e5f\u63a2\u7d22\u4e86\u7b56\u7565\u7b97\u6cd5\uff0c\u4f8b\u5982 PPO\uff08Schulman \u7b49\u4eba\uff0c2017\uff09\uff0c\u4f46\u53d1\u73b0 DPO \u5728\u5927\u89c4\u6a21\u6a21\u578b\u4e0a\u9700\u8981\u66f4\u5c11\u7684\u8ba1\u7b97\u91cf\uff0c\u5e76\u4e14\u8868\u73b0\u66f4\u597d\uff0c\u5c24\u5176\u662f\u5728\u6307\u4ee4\u9075\u5faa\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u5982 IFEval\uff08Zhou \u7b49\u4eba\uff0c2023\uff09\u3002<\/p>\n<p>\u5bf9\u4e8e Llama 3\uff0c\u6211\u4eec\u4f7f\u7528 1e-5 \u7684\u5b66\u4e60\u7387\uff0c\u5e76\u5c06 \u03b2 \u8d85\u53c2\u6570\u8bbe\u7f6e\u4e3a 0.1\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5bf9 DPO \u5e94\u7528\u4e86\u4ee5\u4e0b\u7b97\u6cd5\u4fee\u6539\uff1a<\/p>\n<ul>\n<li><strong>\u5728 DPO \u635f\u5931\u4e2d\u5c4f\u853d\u683c\u5f0f\u6807\u8bb0:<\/strong>\u00a0\u6211\u4eec\u5c06\u7279\u6b8a\u683c\u5f0f\u6807\u8bb0\uff08\u5305\u62ec\u7b2c 4.1.1 \u8282\u4e2d\u63cf\u8ff0\u7684\u5934\u90e8\u548c\u7ec8\u6b62\u6807\u8bb0\uff09\u4ece\u9009\u5b9a\u7684\u548c\u62d2\u7edd\u7684\u54cd\u5e94\u4e2d\u5c4f\u853d\u6389\uff0c\u4ee5\u7a33\u5b9a DPO \u8bad\u7ec3\u3002\u6211\u4eec\u6ce8\u610f\u5230\uff0c\u8fd9\u4e9b\u6807\u8bb0\u53c2\u4e0e\u635f\u5931\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4e0d\u5e0c\u671b\u7684\u6a21\u578b\u884c\u4e3a\uff0c\u4f8b\u5982\u5c3e\u90e8\u91cd\u590d\u6216\u7a81\u7136\u751f\u6210\u7ec8\u6b62\u6807\u8bb0\u3002\u6211\u4eec\u5047\u8bbe\u8fd9\u662f\u7531\u4e8e DPO \u635f\u5931\u7684\u5bf9\u6bd4\u6027\u8d28\u2014\u2014\u5728\u9009\u5b9a\u548c\u62d2\u7edd\u7684\u54cd\u5e94\u4e2d\u90fd\u5b58\u5728\u5171\u540c\u6807\u8bb0\u4f1a\u5bfc\u81f4\u76f8\u4e92\u51b2\u7a81\u7684\u5b66\u4e60\u76ee\u6807\uff0c\u56e0\u4e3a\u6a21\u578b\u9700\u8981\u540c\u65f6\u589e\u52a0\u548c\u51cf\u5c11\u8fd9\u4e9b\u6807\u8bb0\u7684\u53ef\u80fd\u6027\u3002<\/li>\n<li><strong>\u4f7f\u7528 NLL \u635f\u5931\u8fdb\u884c\u6b63\u5219\u5316:<\/strong>\u00a0\u6211\u4eec\u5bf9\u9009\u5b9a\u7684\u5e8f\u5217\u6dfb\u52a0\u4e86\u4e00\u4e2a\u989d\u5916\u7684\u8d1f\u5bf9\u6570\u4f3c\u7136\uff08NLL\uff09\u635f\u5931\u9879\uff0c\u5176\u7f29\u653e\u7cfb\u6570\u4e3a 0.2\uff0c\u7c7b\u4f3c\u4e8e Pang \u7b49\u4eba\uff082024\uff09\u3002\u8fd9\u6709\u52a9\u4e8e\u8fdb\u4e00\u6b65\u7a33\u5b9a DPO \u8bad\u7ec3\uff0c\u901a\u8fc7\u4fdd\u6301\u751f\u6210\u6240\u9700\u7684\u683c\u5f0f\u5e76\u9632\u6b62\u9009\u5b9a\u54cd\u5e94\u7684\u5bf9\u6570\u6982\u7387\u964d\u4f4e\uff08Pang \u7b49\u4eba\uff0c2024\uff1bPal \u7b49\u4eba\uff0c2024\uff09\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h4>4.1.5 \u6a21\u578b\u5e73\u5747<\/h4>\n<p>\u6700\u540e\uff0c\u6211\u4eec\u5728\u6bcf\u4e2a RM\u3001SFT \u6216 DPO \u9636\u6bb5\u4f7f\u7528\u5404\u79cd\u6570\u636e\u7248\u672c\u6216\u8d85\u53c2\u6570\u7684\u5b9e\u9a8c\u4e2d\u83b7\u5f97\u7684\u6a21\u578b\u8fdb\u884c\u5e73\u5747\uff08Izmailov \u7b49\u4eba\uff0c2019\uff1bWortsman \u7b49\u4eba\uff0c2022\uff1bLi \u7b49\u4eba\uff0c2022\uff09\u3002\u6211\u4eec\u5217\u51fa\u4e86\u7528\u4e8e Llama 3 \u8c03\u8282\u7684\u5185\u90e8\u6536\u96c6\u7684\u4eba\u7c7b\u504f\u597d\u6570\u636e\u7684\u7edf\u8ba1\u4fe1\u606f\u3002\u6211\u4eec\u8bf7\u8bc4\u4ef7\u8005\u4e0e\u6a21\u578b\u8fdb\u884c\u591a\u8f6e\u5bf9\u8bdd\uff0c\u5e76\u6bd4\u8f83\u6bcf\u8f6e\u7684\u56de\u590d\u3002\u5728\u540e\u5904\u7406\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u6bcf\u4e2a\u5bf9\u8bdd\u62c6\u5206\u6210\u591a\u4e2a\u793a\u4f8b\uff0c\u6bcf\u4e2a\u793a\u4f8b\u5305\u542b\u4e00\u4e2a\u63d0\u793a\uff08\u5305\u62ec\u5982\u679c\u6709\u7684\u8bdd\u524d\u4e00\u6b21\u5bf9\u8bdd\uff09\u548c\u4e00\u4e2a\u56de\u590d\uff08\u4f8b\u5982\uff0c\u88ab\u9009\u62e9\u6216\u62d2\u7edd\u7684\u56de\u590d\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3443\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/047e8292afa69af.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"786\" height=\"286\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/047e8292afa69af.png 1225w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/047e8292afa69af-300x109.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/047e8292afa69af-1024x372.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/047e8292afa69af-768x279.png 768w\" sizes=\"auto, (max-width: 786px) 100vw, 786px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h4>4.1.6 \u8fed\u4ee3\u8f6e\u6b21<\/h4>\n<p>\u7ee7Llama 2\u4e4b\u540e\uff0c\u6211\u4eec\u5e94\u7528\u4e0a\u8ff0\u65b9\u6cd5\u8fdb\u884c\u4e86\u516d\u8f6e\u8fed\u4ee3\u3002\u5728\u6bcf\u4e2a\u5faa\u73af\u4e2d\uff0c\u6211\u4eec\u6536\u96c6\u65b0\u7684\u504f\u597d\u6807\u6ce8\u548c\u5fae\u8c03\uff08SFT\uff09\u6570\u636e\uff0c\u5e76\u4ece\u6700\u65b0\u7684\u6a21\u578b\u4e2d\u91c7\u6837\u5408\u6210\u6570\u636e\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>4.2 \u8bad\u7ec3\u540e\u6570\u636e<\/h3>\n<p>\u540e\u8bad\u7ec3\u6570\u636e\u7684\u7ec4\u6210\u5bf9\u8bed\u8a00\u6a21\u578b\u7684\u5b9e\u7528\u6027\u548c\u884c\u4e3a\u8d77\u7740\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u6211\u4eec\u7684\u6ce8\u91ca\u7a0b\u5e8f\u548c\u504f\u597d\u6570\u636e\u6536\u96c6\uff08\u7b2c 4.2.1 \u8282\uff09\u3001SFT \u6570\u636e\u7684\u7ec4\u6210\uff08\u7b2c 4.2.2 \u8282\uff09\u4ee5\u53ca\u6570\u636e\u8d28\u91cf\u63a7\u5236\u548c\u6e05\u6d17\u7684\u65b9\u6cd5\uff08\u7b2c 4.2.3 \u8282\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.2.1 \u504f\u597d\u6570\u636e<\/h4>\n<p>\u6211\u4eec\u7684\u504f\u597d\u6570\u636e\u6807\u6ce8\u8fc7\u7a0b\u4e0e Llama 2 \u76f8\u4f3c\u3002\u5728\u6bcf\u4e00\u8f6e\u4e4b\u540e\uff0c\u6211\u4eec\u90e8\u7f72\u591a\u4e2a\u6a21\u578b\u8fdb\u884c\u6807\u6ce8\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u7528\u6237\u63d0\u793a\u91c7\u6837\u4e24\u4e2a\u6765\u81ea\u4e0d\u540c\u6a21\u578b\u7684\u54cd\u5e94\u3002\u8fd9\u4e9b\u6a21\u578b\u53ef\u4ee5\u4f7f\u7528\u4e0d\u540c\u7684\u6570\u636e\u6df7\u5408\u548c\u5bf9\u9f50\u65b9\u6848\u8bad\u7ec3\uff0c\u4ece\u800c\u5177\u6709\u4e0d\u540c\u7684\u80fd\u529b\u5f3a\u5ea6\uff08\u4f8b\u5982\uff0c\u4ee3\u7801\u4e13\u4e1a\u77e5\u8bc6\uff09\u548c\u589e\u52a0\u6570\u636e\u591a\u6837\u6027\u3002\u6211\u4eec\u8981\u6c42\u6807\u6ce8\u8005\u6839\u636e\u5176\u504f\u597d\u7a0b\u5ea6\u5c06\u504f\u597d\u8bc4\u5206\u5206\u7c7b\u4e3a\u56db\u4e2a\u7ea7\u522b\u4e4b\u4e00\uff1a\u663e\u8457\u66f4\u597d\u3001\u66f4\u597d\u3001\u7565\u597d\u6216\u7565\u5fae\u66f4\u597d\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u5728\u504f\u597d\u6392\u5e8f\u4e4b\u540e\u52a0\u5165\u4e86\u4e00\u4e2a\u7f16\u8f91\u6b65\u9aa4\uff0c\u4ee5\u9f13\u52b1\u6807\u6ce8\u8005\u8fdb\u4e00\u6b65\u6539\u8fdb\u9996\u9009\u54cd\u5e94\u3002\u6807\u6ce8\u8005\u53ef\u4ee5\u76f4\u63a5\u7f16\u8f91\u6240\u9009\u54cd\u5e94\uff0c\u6216\u8005\u7528\u53cd\u9988\u63d0\u793a\u6a21\u578b\u4ee5\u7ec6\u5316\u5176\u81ea\u8eab\u54cd\u5e94\u3002\u56e0\u6b64\uff0c\u90e8\u5206\u504f\u597d\u6570\u636e\u5177\u6709\u4e09\u4e2a\u6392\u5e8f\u7684\u54cd\u5e94\uff08\u7f16\u8f91 &gt; \u9009\u62e9 &gt; \u62d2\u7edd\uff09\u3002<\/p>\n<p>\u8868 6 \u4e2d\u62a5\u544a\u4e86\u6211\u4eec\u7528\u4e8e Llama 3 \u8bad\u7ec3\u7684\u504f\u597d\u6ce8\u91ca\u7edf\u8ba1\u6570\u636e\u3002\u901a\u7528\u82f1\u8bed\u6db5\u76d6\u591a\u4e2a\u5b50\u7c7b\u522b\uff0c\u4f8b\u5982\u57fa\u4e8e\u77e5\u8bc6\u7684\u95ee\u7b54\u6216\u7cbe\u786e\u7684\u6307\u4ee4\u9075\u5faa\uff0c\u8fd9\u4e9b\u90fd\u8d85\u51fa\u4e86\u7279\u5b9a\u80fd\u529b\u7684\u8303\u56f4\u3002\u4e0e Llama 2 \u76f8\u6bd4\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u63d0\u793a\u548c\u54cd\u5e94\u7684\u5e73\u5747\u957f\u5ea6\u589e\u52a0\u4e86\uff0c\u8fd9\u8868\u660e\u6211\u4eec\u6b63\u5728\u66f4\u590d\u6742\u7684\u4efb\u52a1\u4e0a\u8bad\u7ec3 Llama 3\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5b9e\u65bd\u4e86\u8d28\u91cf\u5206\u6790\u548c\u4eba\u5de5\u8bc4\u4f30\u6d41\u7a0b\u6765\u4e25\u683c\u8bc4\u4f30\u6536\u96c6\u7684\u6570\u636e\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u5b8c\u5584\u63d0\u793a\u5e76\u4e3a\u6807\u6ce8\u8005\u63d0\u4f9b\u7cfb\u7edf\u3001\u53ef\u884c\u7684\u53cd\u9988\u3002\u4f8b\u5982\uff0c\u968f\u7740 Llama 3 \u5728\u6bcf\u4e00\u8f6e\u540e\u90fd\u5f97\u5230\u6539\u8fdb\uff0c\u6211\u4eec\u5c06\u76f8\u5e94\u5730\u589e\u52a0\u63d0\u793a\u7684\u590d\u6742\u6027\u4ee5\u9488\u5bf9\u6a21\u578b\u843d\u540e\u7684\u9886\u57df\u3002<\/p>\n<p>\u5728\u6bcf\u4e00\u8f6e\u540e\u671f\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u4f1a\u4f7f\u7528\u5f53\u65f6\u6240\u6709\u53ef\u7528\u7684\u504f\u597d\u6570\u636e\u8fdb\u884c\u5956\u52b1\u5efa\u6a21\uff0c\u800c\u53ea\u4f7f\u7528\u6765\u81ea\u5404\u79cd\u80fd\u529b\u7684\u6700\u65b0\u6279\u6b21\u8fdb\u884c DPO \u8bad\u7ec3\u3002\u5bf9\u4e8e\u5956\u52b1\u5efa\u6a21\u548c DPO\uff0c\u6211\u4eec\u4f7f\u7528\u6807\u8bb0\u4e3a\u201c\u9009\u62e9\u54cd\u5e94\u663e\u8457\u66f4\u597d\u6216\u66f4\u597d\u201d\u7684\u6837\u672c\u8fdb\u884c\u8bad\u7ec3\uff0c\u5e76\u5c06\u5177\u6709\u76f8\u4f3c\u54cd\u5e94\u7684\u6837\u672c\u4e22\u5f03\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.2.2 SFT \u6570\u636e<\/h4>\n<p>\u6211\u4eec\u7684\u5fae\u8c03\u6570\u636e\u4e3b\u8981\u6765\u81ea\u4e8e\u4ee5\u4e0b\u6765\u6e90\uff1a<\/p>\n<ul>\n<li>\u6765\u81ea\u6211\u4eec\u7684\u4eba\u5de5\u6807\u6ce8\u6536\u96c6\u4e2d\u7684\u63d0\u793a\u53ca\u5176\u62d2\u7edd\u91c7\u6837\u54cd\u5e94<\/li>\n<li>\u9488\u5bf9\u7279\u5b9a\u80fd\u529b\u7684\u5408\u6210\u6570\u636e\uff08\u8be6\u60c5\u89c1\u7b2c 4.3 \u8282\uff09<\/li>\n<li>\u5c11\u91cf\u4eba\u5de5\u6807\u6ce8\u7684\u6570\u636e\uff08\u8be6\u60c5\u89c1\u7b2c 4.3 \u8282\uff09<\/li>\n<\/ul>\n<p>\u968f\u7740\u6211\u4eec\u540e\u8bad\u7ec3\u5468\u671f\u7684\u8fdb\u884c\uff0c\u6211\u4eec\u5f00\u53d1\u51fa\u66f4\u5f3a\u5927\u7684 Llama 3 \u53d8\u4f53\uff0c\u5e76\u4f7f\u7528\u8fd9\u4e9b\u53d8\u4f53\u6536\u96c6\u66f4\u5927\u7684\u6570\u636e\u96c6\uff0c\u4ee5\u8986\u76d6\u5e7f\u6cdb\u7684\u590d\u6742\u80fd\u529b\u8303\u56f4\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u62d2\u7edd\u91c7\u6837\u8fc7\u7a0b\u548c\u6700\u7ec8 SFT \u6570\u636e\u6df7\u5408\u7684\u603b\u4f53\u6784\u6210\u7684\u7ec6\u8282\u3002<\/p>\n<p><strong>\u62d2\u7edd\u91c7\u6837\u3002<\/strong>\u5728\u62d2\u7edd\u91c7\u6837 (RS) \u4e2d\uff0c\u5bf9\u4e8e\u6211\u4eec\u5728\u4eba\u5de5\u6807\u6ce8\u671f\u95f4\u6536\u96c6\u7684\u6bcf\u4e2a\u63d0\u793a\uff08\u7b2c 4.2.1 \u8282\uff09\uff0c\u6211\u4eec\u4ece\u6700\u65b0\u7684\u804a\u5929\u6a21\u578b\u7b56\u7565\u4e2d\u91c7\u6837 K \u4e2a\u8f93\u51fa\uff08\u901a\u5e38\u662f\u524d\u4e00\u4e2a\u540e\u8bad\u7ec3\u8fed\u4ee3\u4e2d\u7684\u6700\u4f73\u6267\u884c\u68c0\u67e5\u70b9\uff0c\u6216\u8005\u7279\u5b9a\u80fd\u529b\u7684\u6700\u4f73\u6267\u884c\u68c0\u67e5\u70b9\uff09\u5e76\u4e14\u4f7f\u7528\u6211\u4eec\u7684\u5956\u52b1\u6a21\u578b\u6765\u9009\u62e9\u6700\u4f73\u5019\u9009\u8005\uff0c\u4e0e Bai \u7b49\u4eba (2022) \u4e00\u81f4\u3002\u5728\u540e\u8bad\u7ec3\u7684\u540e\u671f\u9636\u6bb5\uff0c\u6211\u4eec\u4f1a\u5f15\u5165\u7cfb\u7edf\u63d0\u793a\u6765\u5f15\u5bfc RS \u54cd\u5e94\u7b26\u5408\u671f\u671b\u7684\u8bed\u6c14\u3001\u98ce\u683c\u6216\u683c\u5f0f\uff0c\u8fd9\u5bf9\u4e8e\u4e0d\u540c\u7684\u80fd\u529b\u53ef\u80fd\u4f1a\u6709\u6240\u4e0d\u540c\u3002<\/p>\n<p>\u4e3a\u4e86\u63d0\u9ad8\u62d2\u7edd\u91c7\u6837\u7684\u6548\u7387\uff0c\u6211\u4eec\u91c7\u7528\u4e86 PagedAttention\uff08Kwon \u7b49\u4eba\uff0c2023\uff09\u3002PagedAttention \u901a\u8fc7\u52a8\u6001\u952e\u503c\u7f13\u5b58\u5206\u914d\u6765\u63d0\u9ad8\u5185\u5b58\u6548\u7387\u3002\u5b83\u901a\u8fc7\u6839\u636e\u5f53\u524d\u7f13\u5b58\u5bb9\u91cf\u52a8\u6001\u8c03\u5ea6\u8bf7\u6c42\u6765\u652f\u6301\u4efb\u610f\u8f93\u51fa\u957f\u5ea6\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u8fd9\u4f1a\u5e26\u6765\u5185\u5b58\u8017\u5c3d\u65f6\u7684\u4ea4\u6362\u98ce\u9669\u3002\u4e3a\u4e86\u6d88\u9664\u8fd9\u79cd\u4ea4\u6362\u5f00\u9500\uff0c\u6211\u4eec\u5b9a\u4e49\u4e86\u4e00\u4e2a\u6700\u5927\u8f93\u51fa\u957f\u5ea6\uff0c\u5e76\u4e14\u53ea\u6709\u5728\u6709\u8db3\u591f\u7684\u5185\u5b58\u53ef\u4ee5\u5bb9\u7eb3\u8be5\u957f\u5ea6\u7684\u8f93\u51fa\u65f6\u624d\u6267\u884c\u8bf7\u6c42\u3002PagedAttention \u8fd8\u4f7f\u6211\u4eec\u53ef\u4ee5\u8de8\u6240\u6709\u76f8\u5e94\u8f93\u51fa\u5171\u4eab\u63d0\u793a\u7684\u952e\u503c\u7f13\u5b58\u9875\u9762\u3002\u603b\u800c\u8a00\u4e4b\uff0c\u8fd9\u5bfc\u81f4\u62d2\u7edd\u91c7\u6837\u8fc7\u7a0b\u4e2d\u7684\u541e\u5410\u91cf\u63d0\u9ad8\u4e86 2 \u500d\u4ee5\u4e0a\u3002<\/p>\n<p><strong>\u603b\u4f53\u6570\u636e\u7ec4\u6210\u3002<\/strong>\u8868 7 \u663e\u793a\u4e86\u6211\u4eec\u201c\u6709\u7528\u6027\u201d\u6df7\u5408\u6bcf\u4e2a\u5e7f\u6cdb\u7c7b\u522b\u7684\u6570\u636e\u7edf\u8ba1\u4fe1\u606f\u3002\u867d\u7136 SFT \u548c\u504f\u597d\u6570\u636e\u5305\u542b\u91cd\u53e0\u7684\u9886\u57df\uff0c\u4f46\u5b83\u4eec\u7684\u7b56\u5212\u65b9\u5f0f\u4e0d\u540c\uff0c\u5bfc\u81f4\u4e0d\u540c\u7684\u8ba1\u6570\u7edf\u8ba1\u91cf\u3002\u5728\u7b2c 4.2.3 \u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u63cf\u8ff0\u7528\u4e8e\u5206\u7c7b\u6211\u4eec\u7684\u6570\u636e\u6837\u672c\u7684\u4e3b\u9898\u3001\u590d\u6742\u6027\u548c\u8d28\u91cf\u7684\u6280\u672f\u3002\u5728\u6bcf\u4e00\u8f6e\u540e\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u90fd\u4f1a\u4ed4\u7ec6\u8c03\u6574\u6211\u4eec\u603b\u4f53\u7684\u6570\u636e\u6df7\u5408\uff0c\u4ee5\u5728\u591a\u4e2a\u8f74\u4e0a\u8c03\u6574\u6027\u80fd\uff0c\u4ee5\u9002\u5e94\u5e7f\u6cdb\u7684\u57fa\u51c6\u6d4b\u8bd5\u3002\u6211\u4eec\u7684\u6700\u7ec8\u6570\u636e\u6df7\u5408\u4f1a\u5bf9\u67d0\u4e9b\u9ad8\u8d28\u91cf\u6765\u6e90\u8fdb\u884c\u591a\u6b21\u8fed\u4ee3\uff0c\u5e76\u5c06\u5176\u4ed6\u6765\u6e90\u8fdb\u884c\u964d\u91c7\u6837\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3444\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18a6459d74583e2.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"793\" height=\"290\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18a6459d74583e2.png 1219w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18a6459d74583e2-300x110.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18a6459d74583e2-1024x375.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/18a6459d74583e2-768x281.png 768w\" sizes=\"auto, (max-width: 793px) 100vw, 793px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h4>4.2.3 \u6570\u636e\u5904\u7406\u4e0e\u8d28\u91cf\u63a7\u5236<\/h4>\n<p>\u8003\u8651\u5230\u6211\u4eec\u7684\u5927\u90e8\u5206\u8bad\u7ec3\u6570\u636e\u90fd\u662f\u6a21\u578b\u751f\u6210\u7684\uff0c\u56e0\u6b64\u9700\u8981\u4ed4\u7ec6\u6e05\u6d17\u548c\u8d28\u91cf\u63a7\u5236\u3002<\/p>\n<p><strong>\u6570\u636e\u6e05\u6d17\uff1a<\/strong>\u00a0\u5728\u65e9\u671f\u9636\u6bb5\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u6570\u636e\u4e2d\u5b58\u5728\u8bb8\u591a\u4e0d\u5e0c\u671b\u51fa\u73b0\u7684\u6a21\u5f0f\uff0c\u4f8b\u5982\u8fc7\u91cf\u4f7f\u7528\u8868\u60c5\u7b26\u53f7\u6216\u611f\u53f9\u53f7\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5b9e\u65bd\u4e86\u4e00\u7cfb\u5217\u57fa\u4e8e\u89c4\u5219\u7684\u6570\u636e\u5220\u9664\u548c\u4fee\u6539\u7b56\u7565\u6765\u8fc7\u6ee4\u6216\u6e05\u9664\u6709\u95ee\u9898\u7684\u6578\u64da\u3002\u4f8b\u5982\uff0c\u4e3a\u4e86\u51cf\u8f7b\u8fc7\u5ea6\u9053\u6b49\u7684\u8bed\u8c03\u95ee\u9898\uff0c\u6211\u4eec\u5c06\u8bc6\u522b\u8fc7\u5ea6\u4f7f\u7528\u7684\u77ed\u8bed\uff08\u4f8b\u5982\u201c\u5bf9\u4e0d\u8d77\u201d\u6216\u201c\u6211\u9053\u6b49\u201d\uff09\uff0c\u5e76\u4ed4\u7ec6\u5e73\u8861\u6570\u636e\u96c6\u4e2d\u6b64\u7c7b\u6837\u672c\u7684\u6bd4\u4f8b\u3002<\/p>\n<p><strong>\u6570\u636e\u4fee\u526a\uff1a<\/strong>\u00a0\u6211\u4eec\u8fd8\u5e94\u7528\u4e86\u4e00\u7cfb\u5217\u57fa\u4e8e\u6a21\u578b\u7684\u6280\u672f\u6765\u5220\u9664\u4f4e\u8d28\u91cf\u7684\u8bad\u7ec3\u6837\u672c\u5e76\u63d0\u9ad8\u6574\u4f53\u6a21\u578b\u6027\u80fd\uff1a<\/p>\n<ul>\n<li><strong>\u4e3b\u9898\u5206\u7c7b\uff1a<\/strong>\u00a0\u6211\u4eec\u9996\u5148\u5c06 Llama 3 8B \u5fae\u8c03\u6210\u4e00\u4e2a\u4e3b\u9898\u5206\u7c7b\u5668\uff0c\u5e76\u5bf9\u6240\u6709\u6570\u636e\u8fdb\u884c\u63a8\u7406\u4ee5\u5c06\u5176\u5206\u7c7b\u4e3a\u7c97\u7c92\u5ea6\u7c7b\u522b\uff08\u201c\u6570\u5b66\u63a8\u7406\u201d\uff09\u548c\u7ec6\u7c92\u5ea6\u7c7b\u522b\uff08\u201c\u51e0\u4f55\u4e0e\u4e09\u89d2\u51fd\u6570\u201d\uff09\u3002<\/li>\n<li><strong>\u8d28\u91cf\u8bc4\u5206\uff1a<\/strong> \u6211\u4eec\u4f7f\u7528\u5956\u52b1\u6a21\u578b\u548c\u57fa\u4e8e Llama \u7684\u4fe1\u53f7\u6765\u83b7\u5f97\u6bcf\u4e2a\u6837\u672c\u7684\u8d28\u91cf\u5206\u6570\u3002\u5bf9\u4e8e\u57fa\u4e8e RM \u7684\u5f97\u5206\uff0c\u6211\u4eec\u5c06\u5f97\u5206\u5904\u4e8e\u6700\u9ad8\u56db\u5206\u4f4d\u6570\u7684\u6570\u636e\u89c6\u4e3a\u9ad8\u8d28\u91cf\u6570\u636e\u3002\u5bf9\u4e8e\u57fa\u4e8e Llama \u7684\u5f97\u5206\uff0c\u6211\u4eec\u63d0\u793a Llama 3 \u68c0\u67e5\u70b9\u4ee5\u4e09\u4e2a\u7b49\u7ea7\uff08\u51c6\u786e\u6027\u3001\u6307\u4ee4\u9075\u5faa\u6027\u548c\u8bed\u6c14\/\u5448\u73b0\uff09\u5bf9\u901a\u7528\u82f1\u8bed\u6570\u636e\u8fdb\u884c\u8bc4\u5206\uff0c\u5e76\u4ee5\u4e24\u4e2a\u7b49\u7ea7\uff08\u9519\u8bef\u8bc6\u522b\u548c\u7528\u6237\u610f\u56fe\uff09\u5bf9\u4ee3\u7801\u6570\u636e\u8fdb\u884c\u8bc4\u5206\uff0c\u5e76\u5c06\u83b7\u5f97\u6700\u9ad8\u5206\u7684\u6837\u672c\u89c6\u4e3a\u9ad8\u8d28\u91cf\u6570\u636e\u3002 RM \u548c\u57fa\u4e8e Llama \u7684\u5206\u6570\u7684\u51b2\u7a81\u7387\u5f88\u9ad8\uff0c\u6211\u4eec\u53d1\u73b0\u5c06\u8fd9\u4e9b\u4fe1\u53f7\u7ed3\u5408\u8d77\u6765\u53ef\u4ee5\u83b7\u5f97\u6700\u4f73\u7684\u5185\u90e8\u6d4b\u8bd5\u96c6\u53ec\u56de\u7387\u3002\u6700\u7ec8\uff0c\u6211\u4eec\u9009\u62e9\u90a3\u4e9b\u88ab RM \u6216\u57fa\u4e8e Llama \u8fc7\u6ee4\u5668\u6807\u8bb0\u4e3a\u9ad8\u8d28\u91cf\u7684\u793a\u4f8b\u3002<\/li>\n<li><strong>\u96be\u5ea6\u8bc4\u5206\uff1a<\/strong>\u00a0\u7531\u4e8e\u6211\u4eec\u8fd8\u5bf9\u4f18\u5148\u8003\u8651\u66f4\u590d\u6742\u7684\u6a21\u578b\u793a\u4f8b\u611f\u5174\u8da3\uff0c\u6211\u4eec\u4f7f\u7528\u4e24\u4e2a\u96be\u5ea6\u5ea6\u91cf\u6765\u8bc4\u5206\u6570\u636e\uff1aInstag (Lu et al., 2023) \u548c\u57fa\u4e8e Llama \u7684\u8bc4\u5206\u3002\u5bf9\u4e8e Instag\uff0c\u6211\u4eec\u63d0\u793a Llama 3 70B \u5bf9 SFT \u63d0\u793a\u6267\u884c\u610f\u56fe\u6807\u8bb0\uff0c\u5176\u4e2d\u66f4\u591a\u7684\u610f\u56fe\u610f\u5473\u7740\u66f4\u9ad8\u7684\u590d\u6742\u6027\u3002\u6211\u4eec\u8fd8\u63d0\u793a Llama 3 \u4ee5\u4e09\u4e2a\u7b49\u7ea7\uff08Liu et al., 2024c\uff09\u6d4b\u91cf\u5bf9\u8bdd\u7684\u96be\u5ea6\u3002<\/li>\n<\/ul>\n<ul>\n<li><strong>\u8bed\u4e49\u53bb\u91cd\uff1a<\/strong>\u00a0\u6700\u540e\uff0c\u6211\u4eec\u6267\u884c\u8bed\u4e49\u53bb\u91cd (Abbas et al., 2023; Liu et al., 2024c)\u3002 \u6211\u4eec\u9996\u5148\u4f7f\u7528 RoBERTa (Liu et al., 2019b) \u5bf9\u5b8c\u6574\u7684\u5bf9\u8bdd\u8fdb\u884c\u805a\u7c7b\uff0c\u5e76\u5728\u6bcf\u4e2a\u805a\u7c7b\u4e2d\u6309\u8d28\u91cf\u5206\u6570 \u00d7 \u96be\u5ea6\u5206\u6570\u6392\u5e8f\u3002\u7136\u540e\uff0c\u6211\u4eec\u901a\u8fc7\u8fed\u4ee3\u6240\u6709\u6392\u5e8f\u793a\u4f8b\u8fdb\u884c\u8d2a\u5a6a\u9009\u62e9\uff0c\u53ea\u4fdd\u7559\u4e0e\u5230\u76ee\u524d\u4e3a\u6b62\u5728\u805a\u7c7b\u4e2d\u770b\u5230\u7684\u793a\u4f8b\u7684\u6700\u5927\u4f59\u5f26\u76f8\u4f3c\u5ea6\u5c0f\u4e8e\u9608\u503c\u7684\u793a\u4f8b\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>4.3 \u80fd\u529b<\/h3>\n<p>\u6211\u4eec\u7279\u522b\u5f3a\u8c03\u4e86\u4e00\u4e9b\u4e3a\u4e86\u63d0\u5347\u7279\u5b9a\u80fd\u529b\u6240\u505a\u7684\u52aa\u529b\uff0c\u4f8b\u5982\u4ee3\u7801\u5904\u7406\uff08\u7b2c4.3.1\u8282\uff09\u3001\u591a\u8bed\u8a00\u6027\uff08\u7b2c4.3.2\u8282\uff09\u3001\u6570\u5b66\u548c\u63a8\u7406\u80fd\u529b\uff08\u7b2c4.3.3\u8282\uff09\u3001\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u80fd\u529b\uff08\u7b2c4.3.4\u8282\uff09\u3001\u5de5\u5177\u4f7f\u7528\uff08\u7b2c4.3.5\u8282\uff09\u3001\u4e8b\u5b9e\u6027\uff08\u7b2c4.3.6\u8282\uff09\u4ee5\u53ca\u53ef\u63a7\u6027\uff08\u7b2c4.3.7\u8282\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.3.1 \u4ee3\u7801<\/h4>\n<p>\u81ea\u4ece <a href=\"https:\/\/www.kdjingpai.com\/pt\/microsoft-copilo\/\">Copilot<\/a> \u548c Codex (Chen \u7b49\u4eba\uff0c2021 \u5e74) \u53d1\u5e03\u4ee5\u6765\uff0c\u7528\u4e8e\u4ee3\u7801\u7684 LLM \u5df2\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\u3002\u5f00\u53d1\u4eba\u5458\u73b0\u5728\u5e7f\u6cdb\u4f7f\u7528\u8fd9\u4e9b\u6a21\u578b\u6765\u751f\u6210\u4ee3\u7801\u7247\u6bb5\u3001\u8c03\u8bd5\u3001\u81ea\u52a8\u5316\u4efb\u52a1\u548c\u63d0\u9ad8\u4ee3\u7801\u8d28\u91cf\u3002\u5bf9\u4e8e Llama 3\uff0c\u6211\u4eec\u7684\u76ee\u6807\u662f\u6539\u8fdb\u5e76\u8bc4\u4f30\u4ee5\u4e0b\u4f18\u5148\u7ea7\u7f16\u7a0b\u8bed\u8a00\u7684\u4ee3\u7801\u751f\u6210\u3001\u6587\u6863\u3001\u8c03\u8bd5\u548c\u5ba1\u67e5\u529f\u80fd\uff1aPython\u3001Java\u3001JavaScript\u3001C\/C++\u3001TypeScript\u3001Rust\u3001PHP\u3001HTML\/CSS\u3001SQL\u3001bash\/shell\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86\u901a\u8fc7\u8bad\u7ec3\u4ee3\u7801\u4e13\u5bb6\u3001\u4e3a SFT \u751f\u6210\u5408\u6210\u6570\u636e\u3001\u901a\u8fc7\u7cfb\u7edf\u63d0\u793a\u8f6c\u5411\u6539\u8fdb\u683c\u5f0f\u4ee5\u53ca\u521b\u5efa\u8d28\u91cf\u8fc7\u6ee4\u5668\u4ee5\u4ece\u8bad\u7ec3\u6570\u636e\u4e2d\u5220\u9664\u4e0d\u826f\u6837\u672c\u6765\u6539\u8fdb\u8fd9\u4e9b\u7f16\u7801\u529f\u80fd\u7684\u5de5\u4f5c\u3002<\/p>\n<p><strong>\u4e13\u5bb6\u8bad\u7ec3\u3002<\/strong>\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u4ee3\u7801\u4e13\u5bb6\uff0c\u5e76\u5728\u968f\u540e\u7684\u591a\u8f6e\u540e\u8bad\u7ec3\u4e2d\u4f7f\u7528\u5b83\u6765\u6536\u96c6\u9ad8\u8d28\u91cf\u7684\u4eba\u7c7b\u4ee3\u7801\u6ce8\u91ca\u3002\u8fd9\u662f\u901a\u8fc7\u5206\u652f\u4e3b\u8981\u9884\u8bad\u7ec3\u8fd0\u884c\u5e76\u7ee7\u7eed\u5bf9\u4e3b\u8981 (&gt;85%) \u4e3a\u4ee3\u7801\u6570\u636e\u7684 1T \u4ee4\u724c\u6df7\u5408\u8fdb\u884c\u9884\u8bad\u7ec3\u6765\u5b9e\u73b0\u7684\u3002\u5df2\u8bc1\u660e\u5728\u7279\u5b9a\u9886\u57df\u6301\u7eed\u8fdb\u884c\u9886\u57df\u7279\u5b9a\u6570\u636e\u9884\u8bad\u7ec3\u53ef\u4ee5\u6709\u6548\u5730\u63d0\u9ad8\u6027\u80fd (Gururangan \u7b49\u4eba\uff0c2020 \u5e74)\u3002\u6211\u4eec\u9075\u5faa\u4e0e CodeLlama (Rozi\u00e8re \u7b49\u4eba\uff0c2023 \u5e74) \u76f8\u4f3c\u7684\u914d\u65b9\u3002\u5728\u8bad\u7ec3\u7684\u6700\u540e\u51e0\u5343\u6b65\u4e2d\uff0c\u6211\u4eec\u5728\u9ad8\u8d28\u91cf\u7684\u4ed3\u5e93\u7ea7\u522b\u4ee3\u7801\u6570\u636e\u6df7\u5408\u4e0a\u6267\u884c\u957f\u4e0a\u4e0b\u6587\u5fae\u8c03 (LCFT)\uff0c\u5c06\u4e13\u5bb6\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u5230 16K \u4ee4\u724c\u3002\u6700\u540e\uff0c\u6211\u4eec\u9075\u5faa\u7b2c 4.1 \u8282\u4e2d\u63cf\u8ff0\u7684\u7c7b\u4f3c\u7684\u540e\u8bad\u7ec3\u5efa\u6a21\u914d\u65b9\u6765\u5bf9\u9f50\u8be5\u6a21\u578b\uff0c\u4f46\u4f7f\u7528\u4e3b\u8981\u9488\u5bf9\u4ee3\u7801\u7684 SFT \u548c DPO \u6570\u636e\u6df7\u5408\u3002\u8be5\u6a21\u578b\u4e5f\u7528\u4e8e\u7f16\u7801\u63d0\u793a\u7684\u62d2\u7edd\u91c7\u6837 (\u7b2c 4.2.2 \u8282)\u3002<\/p>\n<p><strong>\u5408\u6210\u6570\u636e\u751f\u6210\u3002<\/strong>\u5728\u5f00\u53d1\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0\u4ee3\u7801\u751f\u6210\u5b58\u5728\u5173\u952e\u95ee\u9898\uff0c\u5305\u62ec\u96be\u4ee5\u9075\u5faa\u6307\u4ee4\u3001\u4ee3\u7801\u8bed\u6cd5\u9519\u8bef\u3001\u4ee3\u7801\u751f\u6210\u4e0d\u6b63\u786e\u4ee5\u53ca\u96be\u4ee5\u4fee\u590d\u9519\u8bef\u3002\u5c3d\u7ba1\u7406\u8bba\u4e0a\u5bc6\u96c6\u7684\u4eba\u7c7b\u6ce8\u91ca\u53ef\u4ee5\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u4f46\u5408\u6210\u6570\u636e\u751f\u6210\u63d0\u4f9b\u4e86\u4e00\u79cd\u8865\u5145\u65b9\u6cd5\uff0c\u6210\u672c\u66f4\u4f4e\uff0c\u89c4\u6a21\u66f4\u5927\uff0c\u4e0d\u53d7\u6807\u6ce8\u4eba\u5458\u4e13\u4e1a\u6c34\u5e73\u7684\u9650\u5236\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u4f7f\u7528 Llama 3 \u548c\u4ee3\u7801\u4e13\u5bb6\u6765\u751f\u6210\u5927\u91cf\u5408\u6210\u7684 SFT \u5bf9\u8bdd\u3002 \u6211\u4eec\u63cf\u8ff0\u4e86\u751f\u6210\u5408\u6210\u4ee3\u7801\u6570\u636e\u7684\u4e09\u4e2a\u9ad8\u7ea7\u65b9\u6cd5\u3002\u603b\u7684\u6765\u8bf4\uff0c\u6211\u4eec\u5728 SFT \u671f\u95f4\u4f7f\u7528\u4e86\u8d85\u8fc7 270 \u4e07\u4e2a\u5408\u6210\u793a\u4f8b\u3002<\/p>\n<p>1. <strong>\u5408\u6210\u6570\u636e\u751f\u6210\uff1a\u6267\u884c\u53cd\u9988\u3002<\/strong>8B \u548c 70B \u6a21\u578b\u5728\u7531\u66f4\u5927\u7684\u3001\u66f4\u80fd\u80dc\u4efb\u7684\u6a21\u578b\u751f\u6210\u7684\u8bad\u7ec3\u6570\u636e\u4e0a\u8868\u73b0\u51fa\u663e\u8457\u7684\u6027\u80fd\u6539\u8fdb\u3002\u7136\u800c\uff0c\u6211\u4eec\u7684\u521d\u6b65\u5b9e\u9a8c\u8868\u660e\uff0c\u4ec5\u7528 Llama 3 405B \u8bad\u7ec3\u5176\u81ea\u8eab\u751f\u6210\u7684\u6578\u64da\u6ca1\u6709\u5e2e\u52a9\uff08\u751a\u81f3\u4f1a\u964d\u4f4e\u6027\u80fd\uff09\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u9650\u5236\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u6267\u884c\u53cd\u9988\u4f5c\u4e3a\u771f\u7406\u4e4b\u6e90\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u4ece\u9519\u8bef\u4e2d\u5b66\u4e60\u5e76\u4fdd\u6301\u5728\u6b63\u8f68\u4e0a\u3002\u7279\u522b\u662f\uff0c\u6211\u4eec\u4f7f\u7528\u4ee5\u4e0b\u8fc7\u7a0b\u751f\u6210\u5927\u7ea6\u4e00\u767e\u4e07\u4e2a\u5408\u6210\u7684\u4ee3\u7801\u5bf9\u8bdd\u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li><strong>\u95ee\u9898\u63cf\u8ff0\u751f\u6210\uff1a<\/strong>\u9996\u5148\uff0c\u6211\u4eec\u751f\u6210\u4e86\u4e00\u5927\u6279\u6db5\u76d6\u5404\u79cd\u4e3b\u9898\uff08\u5305\u62ec\u957f\u5c3e\u5206\u5e03\uff09\u7684\u7f16\u7a0b\u95ee\u9898\u63cf\u8ff0\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u79cd\u591a\u6837\u6027\uff0c\u6211\u4eec\u4ece\u5404\u79cd\u6765\u6e90\u968f\u673a\u91c7\u6837\u4ee3\u7801\u7247\u6bb5\u5e76\u63d0\u793a\u6a21\u578b\u6839\u636e\u8fd9\u4e9b\u793a\u4f8b\u751f\u6210\u7f16\u7a0b\u95ee\u9898\u3002\u8fd9\u4f7f\u6211\u4eec\u80fd\u591f\u5229\u7528\u5e7f\u6cdb\u7684\u4e3b\u9898\u5e76\u521b\u5efa\u4e00\u4e2a\u5168\u9762\u7684\u95ee\u9898\u63cf\u8ff0\u96c6 (Wei \u7b49\u4eba\uff0c2024 \u5e74)\u3002<\/li>\n<li><strong>\u89e3\u51b3\u65b9\u6848\u751f\u6210\uff1a<\/strong>\u7136\u540e\uff0c\u6211\u4eec\u63d0\u793a Llama 3 \u7528\u7ed9\u5b9a\u7684\u7f16\u7a0b\u8bed\u8a00\u89e3\u51b3\u6bcf\u4e2a\u95ee\u9898\u3002\u6211\u4eec\u89c2\u5bdf\u5230\u5728\u63d0\u793a\u4e2d\u6dfb\u52a0\u826f\u597d\u7684\u7f16\u7a0b\u89c4\u5219\u53ef\u4ee5\u63d0\u9ad8\u751f\u6210\u7684\u89e3\u51b3\u65b9\u6848\u8d28\u91cf\u3002\u6b64\u5916\uff0c\u6211\u4eec\u53d1\u73b0\u8981\u6c42\u6a21\u578b\u7528\u6ce8\u91ca\u89e3\u91ca\u5176\u601d\u7ef4\u8fc7\u7a0b\u4e5f\u5f88\u6709\u5e2e\u52a9\u3002<\/li>\n<li><strong>\u6b63\u786e\u6027\u5206\u6790\uff1a<\/strong>\u00a0\u5728\u751f\u6210\u89e3\u51b3\u65b9\u6848\u540e\uff0c\u8bc6\u522b\u5176\u6b63\u786e\u6027\u5e76\u975e\u4fdd\u8bc1\uff0c\u5e76\u5c06\u4e0d\u6b63\u786e\u7684\u89e3\u51b3\u65b9\u6848\u5305\u542b\u5728\u5fae\u8c03\u6570\u636e\u96c6\u53ef\u80fd\u4f1a\u635f\u5bb3\u6a21\u578b\u7684\u8d28\u91cf\u81f3\u5173\u91cd\u8981\u3002\u867d\u7136\u6211\u4eec\u4e0d\u80fd\u786e\u4fdd\u5b8c\u5168\u6b63\u786e\uff0c\u4f46\u6211\u4eec\u5f00\u53d1\u4e86\u8fd1\u4f3c\u6b63\u786e\u6027\u7684\u65b9\u6cd5\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u5c06\u63d0\u53d6\u7684\u6e90\u4ee3\u7801\u4ece\u751f\u6210\u7684\u89e3\u51b3\u65b9\u6848\u4e2d\uff0c\u5e76\u5e94\u7528\u9759\u6001\u548c\u52a8\u6001\u5206\u6790\u6280\u672f\u7ec4\u5408\u8fdb\u884c\u6d4b\u8bd5\u5176\u6b63\u786e\u6027\uff0c\u5305\u62ec\uff1a\n<ul>\n<li><strong>\u9759\u6001\u5206\u6790\uff1a<\/strong>\u00a0\u6211\u4eec\u5c06\u6240\u6709\u751f\u6210\u7684\u4ee3\u7801\u90fd\u901a\u8fc7\u89e3\u6790\u5668\u548c\u4ee3\u7801\u68c0\u67e5\u5de5\u5177\u8fd0\u884c\uff0c\u4ee5\u786e\u4fdd\u8bed\u6cd5\u6b63\u786e\u6027\uff0c\u6355\u6349\u8bed\u6cd5\u9519\u8bef\u3001\u672a\u521d\u59cb\u5316\u53d8\u91cf\u6216\u672a\u5bfc\u5165\u51fd\u6570\u7684\u4f7f\u7528\u3001\u4ee3\u7801\u98ce\u683c\u95ee\u9898\u3001\u7c7b\u578b\u9519\u8bef\u7b49\u3002<\/li>\n<li><strong>\u5355\u5143\u6d4b\u8bd5\u751f\u6210\u548c\u6267\u884c\uff1a<\/strong>\u00a0\u5bf9\u4e8e\u6bcf\u4e2a\u95ee\u9898\u548c\u89e3\u51b3\u65b9\u6848\uff0c\u6211\u4eec\u63d0\u793a\u6a21\u578b\u751f\u6210\u5355\u5143\u6d4b\u8bd5\uff0c\u5e76\u5728\u5bb9\u5668\u5316\u73af\u5883\u4e2d\u4e0e\u89e3\u51b3\u65b9\u6848\u4e00\u8d77\u6267\u884c\uff0c\u6355\u83b7\u8fd0\u884c\u65f6\u6267\u884c\u9519\u8bef\u548c\u4e00\u4e9b\u8bed\u4e49\u9519\u8bef\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u9519\u8bef\u53cd\u9988\u548c\u8fed\u4ee3\u81ea\u6211\u6821\u6b63\uff1a<\/strong>\u00a0\u5f53\u89e3\u51b3\u65b9\u6848\u5728\u4efb\u4f55\u6b65\u9aa4\u90fd\u5931\u8d25\u65f6\uff0c\u6211\u4eec\u5c06\u63d0\u793a\u6a21\u578b\u5bf9\u5176\u8fdb\u884c\u4fee\u6539\u3002\u63d0\u793a\u5305\u542b\u539f\u59cb\u95ee\u9898\u63cf\u8ff0\u3001\u9519\u8bef\u89e3\u51b3\u65b9\u6848\u4ee5\u53ca\u6765\u81ea\u89e3\u6790\u5668\/\u4ee3\u7801\u68c0\u67e5\u5de5\u5177\/\u6d4b\u8bd5\u7a0b\u5e8f\u7684\u53cd\u9988\uff08\u6807\u51c6\u8f93\u51fa\u3001\u6807\u51c6\u9519\u8bef\u548c\u8fd4\u56de\u7801\uff09\u3002\u5355\u5143\u6d4b\u8bd5\u6267\u884c\u5931\u8d25\u540e\uff0c\u6a21\u578b\u53ef\u4ee5\u4fee\u590d\u4ee3\u7801\u4ee5\u901a\u8fc7\u73b0\u6709\u7684\u6d4b\u8bd5\uff0c\u6216\u8005\u4fee\u6539\u5176\u5355\u5143\u6d4b\u8bd5\u4ee5\u9002\u5e94\u751f\u6210\u7684\u4ee3\u7801\u3002\u53ea\u6709\u901a\u8fc7\u6240\u6709\u68c0\u67e5\u7684\u5bf9\u8bdd\u624d\u4f1a\u88ab\u5305\u542b\u5728\u6700\u7ec8\u7684\u6570\u636e\u96c6\u4e2d\uff0c\u7528\u4e8e\u76d1\u7763\u5f0f\u5fae\u8c03\uff08SFT\uff09\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u5927\u7ea6 20% \u7684\u89e3\u51b3\u65b9\u6848\u6700\u521d\u4e0d\u6b63\u786e\u4f46\u8fdb\u884c\u4e86\u81ea\u6211\u6821\u6b63\uff0c\u8fd9\u8868\u660e\u6a21\u578b\u4ece\u6267\u884c\u53cd\u9988\u4e2d\u5b66\u4e60\u5e76\u63d0\u9ad8\u4e86\u5176\u6027\u80fd\u3002<\/li>\n<li><strong>\u5fae\u8c03\u548c\u8fed\u4ee3\u6539\u8fdb\uff1a<\/strong> \u5fae\u8c03\u8fc7\u7a0b\u4f1a\u8fdb\u884c\u591a\u8f6e\uff0c\u6bcf\u8f6e\u90fd\u5efa\u7acb\u5728\u4e4b\u524d\u7684\u8f6e\u5b50\u4e0a\u3002\u6bcf\u6b21\u5fae\u8c03\u540e\uff0c\u6a21\u578b\u90fd\u4f1a\u5f97\u5230\u6539\u8fdb\uff0c\u4e3a\u4e0b\u4e00\u8f6e\u751f\u6210\u66f4\u9ad8\u8d28\u91cf\u7684\u5408\u6210\u6570\u636e\u3002\u8fd9\u4e2a\u8fed\u4ee3\u8fc7\u7a0b\u5141\u8bb8\u5bf9\u6a21\u578b\u6027\u80fd\u8fdb\u884c\u9010\u6b65\u7ec6\u5316\u548c\u589e\u5f3a\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><strong>2. \u5408\u6210\u6570\u636e\u751f\u6210\uff1a\u7f16\u7a0b\u8bed\u8a00\u7ffb\u8bd1\u3002<\/strong>\u00a0\u6211\u4eec\u89c2\u5bdf\u5230\u4e3b\u8981\u7f16\u7a0b\u8bed\u8a00\uff08\u4f8b\u5982 Python\/C++\uff09\u548c\u4e0d\u592a\u5e38\u89c1\u7684\u7f16\u7a0b\u8bed\u8a00\uff08\u4f8b\u5982 Typescript\/PHP\uff09\u4e4b\u95f4\u5b58\u5728\u6027\u80fd\u5dee\u8ddd\u3002\u8fd9\u5e76\u4e0d\u5947\u602a\uff0c\u56e0\u4e3a\u6211\u4eec\u5bf9\u4e8e\u4e0d\u592a\u5e38\u89c1\u7684\u7f16\u7a0b\u8bed\u8a00\u62e5\u6709\u66f4\u5c11\u8bad\u7ec3\u6570\u636e\u3002\u4e3a\u4e86\u7f13\u89e3\u8fd9\u79cd\u60c5\u51b5\uff0c\u6211\u4eec\u5c06\u901a\u8fc7\u5c06\u5e38\u89c1\u7f16\u7a0b\u8bed\u8a00\u7684\u6570\u636e\u7ffb\u8bd1\u6210\u4e0d\u592a\u5e38\u89c1\u7684\u8bed\u8a00\u6765\u8865\u5145\u73b0\u6709\u6570\u636e\uff08\u7c7b\u4f3c\u4e8e Chen \u7b49\u4eba (2023) \u5728\u63a8\u7406\u9886\u57df\uff09\u3002\u8fd9\u662f\u901a\u8fc7\u63d0\u793a Llama 3 \u5e76\u901a\u8fc7\u8bed\u6cd5\u89e3\u6790\u3001\u7f16\u8bd1\u548c\u6267\u884c\u6765\u786e\u4fdd\u8d28\u91cf\u5b9e\u73b0\u7684\u3002\u56fe 8 \u5c55\u793a\u4e86\u4ece Python \u7ffb\u8bd1\u7684\u5408\u6210 PHP \u4ee3\u7801\u793a\u4f8b\u3002\u8fd9\u663e\u8457\u63d0\u9ad8\u4e86 MultiPL-E (Cassano \u7b49\u4eba\uff0c2023) \u57fa\u51c6\u6d4b\u91cf\u7684\u4e0d\u592a\u5e38\u89c1\u8bed\u8a00\u7684\u6027\u80fd\u3002<\/p>\n<p><strong>3. \u5408\u6210\u6570\u636e\u751f\u6210\uff1a\u53cd\u5411\u7ffb\u8bd1\u3002<\/strong>\u00a0\u4e3a\u4e86\u6539\u8fdb\u67d0\u4e9b\u7f16\u7801\u80fd\u529b\uff08\u4f8b\u5982\u6587\u6863\u3001\u89e3\u91ca\uff09\uff0c\u5728\u6267\u884c\u53cd\u9988\u7684\u4fe1\u606f\u91cf\u4e0d\u8db3\u4ee5\u786e\u5b9a\u8d28\u91cf\u7684\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u91c7\u7528\u53e6\u4e00\u79cd\u591a\u6b65\u9aa4\u65b9\u6cd5\u3002\u4f7f\u7528\u6b64\u8fc7\u7a0b\uff0c\u6211\u4eec\u751f\u6210\u4e86\u7ea6 120 \u4e07\u4e2a\u4e0e\u4ee3\u7801\u89e3\u91ca\u3001\u751f\u6210\u3001\u6587\u6863\u548c\u8c03\u8bd5\u76f8\u5173\u7684\u5408\u6210\u5bf9\u8bdd\u3002\u4ece\u9884\u8bad\u7ec3\u6570\u636e\u4e2d\u5404\u79cd\u8bed\u8a00\u7684\u4ee3\u7801\u7247\u6bb5\u5f00\u59cb\uff1a<\/p>\n<p>&nbsp;<\/p>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3445\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/70859eaf3cc99f8.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"629\" height=\"257\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/70859eaf3cc99f8.png 1209w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/70859eaf3cc99f8-300x123.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/70859eaf3cc99f8-1024x418.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/70859eaf3cc99f8-768x314.png 768w\" sizes=\"auto, (max-width: 629px) 100vw, 629px\" \/><\/strong><\/p>\n<p>&nbsp;<\/p>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3446\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/78c4ecf57f7db6f.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"718\" height=\"283\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/78c4ecf57f7db6f.png 1217w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/78c4ecf57f7db6f-300x118.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/78c4ecf57f7db6f-1024x403.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/78c4ecf57f7db6f-768x302.png 768w\" sizes=\"auto, (max-width: 718px) 100vw, 718px\" \/><\/strong><\/p>\n<p>&nbsp;<\/p>\n<ul>\n<li><strong>\u751f\u6210\uff1a<\/strong>\u00a0\u6211\u4eec\u63d0\u793a Llama 3 \u751f\u6210\u4ee3\u8868\u76ee\u6807\u80fd\u529b\u7684\u6570\u636e\uff08\u4f8b\u5982\uff0c\u4e3a\u4ee3\u7801\u7247\u6bb5\u6dfb\u52a0\u6ce8\u91ca\u548c\u6587\u6863\u5b57\u7b26\u4e32\uff0c\u6216\u8005\u8981\u6c42\u6a21\u578b\u89e3\u91ca\u4e00\u6bb5\u4ee3\u7801\uff09\u3002<\/li>\n<li><strong>\u53cd\u5411\u7ffb\u8bd1:<\/strong>\u00a0\u6211\u4eec\u4f1a\u63d0\u793a\u6a21\u578b\u5c06\u5408\u6210\u751f\u6210\u7684\u6570\u636e\u201c\u53cd\u5411\u7ffb\u8bd1\u201d\u56de\u539f\u59cb\u4ee3\u7801\uff08\u4f8b\u5982\uff0c\u6211\u4eec\u63d0\u793a\u6a21\u578b\u4ec5\u4ece\u5176\u6587\u6863\u4e2d\u751f\u6210\u4ee3\u7801\uff0c\u6216\u8005\u8ba9\u6211\u4eec\u8be2\u95ee\u6a21\u578b\u4ec5\u4ece\u5176\u89e3\u91ca\u4e2d\u751f\u6210\u4ee3\u7801\uff09\u3002<\/li>\n<li><strong>\u8fc7\u6ee4:<\/strong>\u00a0\u4f7f\u7528\u539f\u59cb\u4ee3\u7801\u4f5c\u4e3a\u53c2\u8003\uff0c\u6211\u4eec\u63d0\u793a Llama 3 \u786e\u5b9a\u8f93\u51fa\u7684\u8d28\u91cf\uff08\u4f8b\u5982\uff0c\u6211\u4eec\u8be2\u95ee\u6a21\u578b\u53cd\u5411\u7ffb\u8bd1\u540e\u7684\u4ee3\u7801\u4e0e\u539f\u59cb\u4ee3\u7801\u7684\u5fe0\u5b9e\u5ea6\u5982\u4f55\uff09\u3002\u7136\u540e\uff0c\u6211\u4eec\u5728 SFT \u4e2d\u4f7f\u7528\u5177\u6709\u6700\u9ad8\u81ea\u6211\u9a8c\u8bc1\u5206\u6570\u7684\u751f\u6210\u7684\u793a\u4f8b\u3002<\/li>\n<\/ul>\n<p><strong>\u7cfb\u7edf\u63d0\u793a\u5f15\u5bfc\uff0c\u7528\u4e8e\u62d2\u7edd\u91c7\u6837\u3002<\/strong>\u00a0\u5728\u62d2\u7edd\u91c7\u6837\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u4ee3\u7801\u7279\u5b9a\u7684\u7cfb\u7edf\u63d0\u793a\u6765\u6539\u8fdb\u4ee3\u7801\u7684\u53ef\u8bfb\u6027\u3001\u6587\u6863\u3001\u5b8c\u6574\u6027\u548c\u5177\u4f53\u6027\u3002\u56de\u5fc6\u7b2c 7 \u8282\u4e2d\u63d0\u5230\u7684\u5185\u5bb9\uff0c\u8fd9\u4e9b\u6570\u636e\u7528\u4e8e\u5fae\u8c03\u8bed\u8a00\u6a21\u578b\u3002\u56fe 9 \u663e\u793a\u4e86\u7cfb\u7edf\u63d0\u793a\u5982\u4f55\u5e2e\u52a9\u63d0\u9ad8\u751f\u6210\u4ee3\u7801\u8d28\u91cf\u7684\u793a\u4f8b &#8211; \u5b83\u6dfb\u52a0\u4e86\u5fc5\u8981\u7684\u6ce8\u91ca\uff0c\u4f7f\u7528\u4e86\u66f4\u5177\u4fe1\u606f\u91cf\u7684\u53d8\u91cf\u540d\uff0c\u8282\u7701\u5185\u5b58\u7b49\u3002<\/p>\n<p><strong>\u4f7f\u7528\u6267\u884c\u548c\u6a21\u578b\u4f5c\u4e3a\u8bc4\u5224\u6807\u51c6\u8fc7\u6ee4\u8bad\u7ec3\u6570\u636e\u3002<\/strong>\u00a0\u5982\u7b2c 4.2.3 \u8282\u6240\u8ff0\uff0c\u6211\u4eec\u5076\u5c14\u4f1a\u5728\u62d2\u7edd\u91c7\u6837\u7684\u6570\u636e\u4e2d\u9047\u5230\u8d28\u91cf\u95ee\u9898\uff0c\u4f8b\u5982\u5305\u542b\u9519\u8bef\u7684\u4ee3\u7801\u5757\u3002\u68c0\u6d4b\u62d2\u7edd\u91c7\u6837\u6570\u636e\u4e2d\u7684\u8fd9\u4e9b\u95ee\u9898\u4e0d\u50cf\u68c0\u6d4b\u6211\u4eec\u7684\u5408\u6210\u4ee3\u7801\u6570\u636e\u90a3\u6837\u7b80\u5355\uff0c\u56e0\u4e3a\u62d2\u7edd\u91c7\u6837\u7684\u54cd\u5e94\u901a\u5e38\u5305\u542b\u81ea\u7136\u8bed\u8a00\u548c\u4ee3\u7801\u7684\u6df7\u5408\u4f53\uff0c\u800c\u8fd9\u4e9b\u4ee3\u7801\u53ef\u80fd\u5e76\u4e0d\u603b\u662f\u53ef\u6267\u884c\u7684\u3002\uff08\u4f8b\u5982\uff0c\u7528\u6237\u63d0\u793a\u53ef\u80fd\u4f1a\u660e\u786e\u8981\u6c42\u4f2a\u4ee3\u7801\u6216\u4ec5\u5bf9\u53ef\u6267\u884c\u7a0b\u5e8f\u7684\u975e\u5e38\u5c0f\u7684\u7247\u6bb5\u8fdb\u884c\u7f16\u8f91\u3002\uff09 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5229\u7528\u201c\u6a21\u578b\u4f5c\u4e3a\u8bc4\u5224\u8005\u201d\u65b9\u6cd5\uff0c\u5176\u4e2d\u66f4\u65e9\u7248\u672c\u7684 Llama 3 \u4f1a\u6839\u636e\u4e24\u4e2a\u6807\u51c6\u8bc4\u4f30\u5e76\u5206\u914d\u4e8c\u8fdb\u5236\uff080\/1\uff09\u5206\u6570\uff1a\u4ee3\u7801\u6b63\u786e\u6027\u548c\u4ee3\u7801\u98ce\u683c\u3002\u6211\u4eec\u53ea\u4fdd\u7559\u83b7\u5f97 2 \u5206\u6ee1\u5206\u7684\u6837\u672c\u3002\u6700\u521d\uff0c\u8fd9\u79cd\u4e25\u683c\u7684\u8fc7\u6ee4\u5bfc\u81f4\u4e86\u4e0b\u6e38\u57fa\u51c6\u6027\u80fd\u4e0b\u964d\uff0c\u4e3b\u8981\u662f\u56e0\u4e3a\u5b83\u4e0d\u6210\u6bd4\u4f8b\u5730\u5220\u9664\u4e86\u5177\u6709\u6311\u6218\u6027\u63d0\u793a\u7684\u793a\u4f8b\u3002\u4e3a\u4e86\u62b5\u6d88\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u5c06\u4e00\u4e9b\u5206\u7c7b\u4e3a\u6700\u5177\u6311\u6218\u6027\u7684\u7f16\u7801\u6570\u636e\u7684\u54cd\u5e94\u8fdb\u884c\u6218\u7565\u6027\u4fee\u6539\uff0c\u76f4\u5230\u5b83\u4eec\u6ee1\u8db3\u57fa\u4e8e Llama \u7684\u201c\u6a21\u578b\u4f5c\u4e3a\u8bc4\u5224\u8005\u201d\u6807\u51c6\u3002\u901a\u8fc7\u6539\u8fdb\u8fd9\u4e9b\u5177\u6709\u6311\u6218\u6027\u7684\u95ee\u9898\uff0c\u7f16\u7801\u6570\u636e\u5728\u8d28\u91cf\u548c\u96be\u5ea6\u4e4b\u95f4\u53d6\u5f97\u4e86\u5e73\u8861\uff0c\u4ece\u800c\u5b9e\u73b0\u4e86\u6700\u4f73\u7684\u4e0b\u6e38\u6027\u80fd\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.3.2 \u591a\u8bed\u8a00\u80fd\u529b<\/h4>\n<p>\u672c\u8282\u63cf\u8ff0\u4e86\u6211\u4eec\u5982\u4f55\u6539\u8fdb Llama 3 \u7684\u591a\u8bed\u8a00\u80fd\u529b\uff0c\u5305\u62ec\uff1a\u8bad\u7ec3\u4e00\u4e2a\u5728\u66f4\u591a\u591a\u8bed\u8a00\u6570\u636e\u4e0a\u4e13\u4e1a\u5316\u7684\u4e13\u5bb6\u6a21\u578b\uff1b\u4e3a\u5fb7\u8bed\u3001\u6cd5\u8bed\u3001\u610f\u5927\u5229\u8bed\u3001\u8461\u8404\u7259\u8bed\u3001\u5370\u5730\u8bed\u3001\u897f\u73ed\u7259\u8bed\u548c\u6cf0\u8bed\u91c7\u8d2d\u5e76\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u591a\u8bed\u8a00\u6307\u4ee4\u5fae\u8c03\u6570\u636e\uff1b\u4ee5\u53ca\u89e3\u51b3\u591a\u8bed\u8a00\u8bed\u8a00\u5f15\u5bfc\u7684\u5177\u4f53\u6311\u6218\uff0c\u4ee5\u63d0\u9ad8\u6211\u4eec\u6a21\u578b\u7684\u6574\u4f53\u6027\u80fd\u3002<\/p>\n<p><strong>\u4e13\u5bb6\u8bad\u7ec3\u3002<\/strong>\u6211\u4eec\u7684 Llama 3 \u9884\u8bad\u7ec3\u6570\u636e\u6df7\u5408\u5305\u542b\u7684\u82f1\u8bed\u6807\u8bb0\u8fdc\u8fdc\u591a\u4e8e\u975e\u82f1\u8bed\u6807\u8bb0\u3002\u4e3a\u4e86\u6536\u96c6\u66f4\u9ad8\u8d28\u91cf\u7684\u975e\u82f1\u8bed\u4eba\u5de5\u6ce8\u91ca\uff0c\u6211\u4eec\u901a\u8fc7\u5206\u652f\u9884\u8bad\u7ec3\u8fd0\u884c\u5e76\u7ee7\u7eed\u5728\u5305\u542b 90% \u591a\u8bed\u8a00\u6807\u8bb0\u7684\u6570\u636e\u6df7\u5408\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u6765\u8bad\u7ec3\u4e00\u4e2a\u591a\u8bed\u8a00\u4e13\u5bb6\u6a21\u578b\u3002\u7136\u540e\uff0c\u6211\u4eec\u5728\u7b2c 4.1 \u8282\u4e2d\u6309\u7167\u8bf4\u660e\u5bf9\u8be5\u4e13\u5bb6\u6a21\u578b\u8fdb\u884c\u540e\u8bad\u7ec3\u3002\u7136\u540e\uff0c\u8be5\u4e13\u5bb6\u6a21\u578b\u7528\u4e8e\u6536\u96c6\u66f4\u9ad8\u8d28\u91cf\u7684\u975e\u82f1\u8bed\u4eba\u5de5\u6ce8\u91ca\uff0c\u76f4\u5230\u9884\u8bad\u7ec3\u5b8c\u5168\u5b8c\u6210\u3002<\/p>\n<p><strong>\u591a\u8bed\u8a00\u6570\u636e\u6536\u96c6\u3002<\/strong>\u6211\u4eec\u7684\u591a\u8bed\u8a00 SFT \u6570\u636e\u4e3b\u8981\u6765\u81ea\u4ee5\u4e0b\u6765\u6e90\u3002\u603b\u4f53\u5206\u5e03\u4e3a\uff1a2.4% \u4eba\u5de5\u6ce8\u91ca\u300144.2% \u6765\u81ea\u5176\u4ed6 NLP \u4efb\u52a1\u7684\u6570\u636e\u300118.8% \u62d2\u7edd\u91c7\u6837\u6570\u636e\u548c 34.6% \u7ffb\u8bd1\u63a8\u7406\u6570\u636e\u3002<\/p>\n<ul>\n<li><strong>\u4eba\u5de5\u6ce8\u91ca\uff1a<\/strong>\u6211\u4eec\u4ece\u8bed\u8a00\u5b66\u5bb6\u548c\u6bcd\u8bed\u4eba\u58eb\u90a3\u91cc\u6536\u96c6\u9ad8\u8d28\u91cf\u7684\u3001\u624b\u52a8\u6807\u6ce8\u7684\u6570\u636e\u3002\u8fd9\u4e9b\u6ce8\u91ca\u4e3b\u8981\u7531\u5f00\u653e\u5f0f\u63d0\u793a\u7ec4\u6210\uff0c\u5b83\u4eec\u4ee3\u8868\u4e86\u73b0\u5b9e\u4e16\u754c\u7684\u7528\u4f8b\u3002<\/li>\n<li><strong>\u6765\u81ea\u5176\u4ed6 NLP \u4efb\u52a1\u7684\u6570\u636e\uff1a<\/strong>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u589e\u5f3a\uff0c\u6211\u4eec\u4f7f\u7528\u6765\u81ea\u5176\u4ed6\u4efb\u52a1\u7684\u591a\u8bed\u8a00\u8bad\u7ec3\u6570\u636e\u5e76\u5c06\u5b83\u4eec\u6539\u5199\u6210\u5bf9\u8bdd\u683c\u5f0f\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u4f7f\u7528\u4e86 exams-qa\uff08Hardalov \u7b49\u4eba\uff0c2020 \u5e74\uff09\u548c Conic10k\uff08Wu \u7b49\u4eba\uff0c2023 \u5e74\uff09\u7684\u6570\u636e\u3002\u4e3a\u4e86\u6539\u8fdb\u8bed\u8a00\u5bf9\u9f50\uff0c\u6211\u4eec\u8fd8\u4f7f\u7528\u4e86 GlobalVoices\uff08Prokopidis \u7b49\u4eba\uff0c2016 \u5e74\uff09\u548c Wikimedia\uff08Tiedemann\uff0c2012 \u5e74\uff09\u7684\u5e76\u884c\u6587\u672c\u3002\u6211\u4eec\u4f7f\u7528\u57fa\u4e8e LID \u7684\u8fc7\u6ee4\u548c Blaser2.0\uff08Seamless Communication \u7b49\u4eba\uff0c2023 \u5e74\uff09\u6765\u5220\u9664\u4f4e\u8d28\u91cf\u6570\u636e\u3002\u5bf9\u4e8e\u5e76\u884c\u6587\u672c\u6570\u636e\uff0c\u6211\u4eec\u6ca1\u6709\u76f4\u63a5\u4f7f\u7528\u53cc\u6587\u672c\u5bf9\uff0c\u800c\u662f\u5e94\u7528\u4e86\u4e00\u4e2a\u53d7 Wei \u7b49\u4eba (2022a) \u542f\u53d1\u7684\u591a\u8bed\u8a00\u6a21\u677f\uff0c\u4ee5\u66f4\u597d\u5730\u6a21\u62df\u7ffb\u8bd1\u548c\u8bed\u8a00\u5b66\u4e60\u573a\u666f\u4e2d\u7684\u771f\u5b9e\u5bf9\u8bdd\u3002<\/li>\n<li><strong>\u62d2\u7edd\u91c7\u6837\u6570\u636e\uff1a<\/strong>\u6211\u4eec\u5c06\u62d2\u7edd\u91c7\u6837\u5e94\u7528\u4e8e\u4eba\u7c7b\u6ce8\u91ca\u7684\u63d0\u793a\uff0c\u4ee5\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u6837\u672c\u7528\u4e8e\u5fae\u8c03\uff0c\u4e0e\u82f1\u8bed\u6570\u636e\u7684\u5904\u7406\u8fc7\u7a0b\u76f8\u6bd4\u8fdb\u884c\u4e86\u5f88\u5c11\u7684\u4fee\u6539\uff1a\n<ul>\n<li>\u751f\u6210\uff1a\u6211\u4eec\u5728\u540e\u8bad\u7ec3\u7684\u65e9\u671f\u8f6e\u6b21\u4e2d\u63a2\u7d22\u4e86\u968f\u673a\u9009\u62e9\u6e29\u5ea6\u8d85\u53c2\u6570\u7684\u8303\u56f4\u4e3a 0.2 -1\uff0c\u4ee5\u5b9e\u73b0\u591a\u6837\u5316\u7684\u751f\u6210\u3002\u4f7f\u7528\u9ad8\u6e29\u5ea6\u65f6\uff0c\u591a\u8bed\u8a00\u63d0\u793a\u7684\u54cd\u5e94\u53ef\u80fd\u4f1a\u53d8\u5f97\u5bcc\u6709\u521b\u610f\u548c\u9f13\u821e\u4eba\u5fc3\uff0c\u4f46\u4e5f\u4f1a\u5bb9\u6613\u51fa\u73b0\u4e0d\u5fc5\u8981\u7684\u6216\u4e0d\u81ea\u7136\u7684\u4ee3\u7801\u5207\u6362\u3002\u5728\u540e\u8bad\u7ec3\u7684\u6700\u540e\u9636\u6bb5\uff0c\u6211\u4eec\u4f7f\u7528 0.6 \u7684\u6052\u5b9a\u503c\u6765\u5e73\u8861\u8fd9\u4e00\u6743\u8861\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u4f7f\u7528\u4e86\u4e13\u95e8\u7684\u7cfb\u7edf\u63d0\u793a\u6765\u6539\u8fdb\u54cd\u5e94\u683c\u5f0f\u3001\u7ed3\u6784\u548c\u4e00\u822c\u53ef\u8bfb\u6027\u3002<\/li>\n<li>\u9009\u62e9\uff1a\u5728\u57fa\u4e8e\u5956\u52b1\u6a21\u578b\u7684\u9009\u62e9\u4e4b\u524d\uff0c\u6211\u4eec\u5b9e\u65bd\u4e86\u591a\u8bed\u8a00\u7279\u5b9a\u7684\u68c0\u67e5\uff0c\u4ee5\u786e\u4fdd\u63d0\u793a\u548c\u56de\u590d\u4e4b\u95f4\u7684\u8bed\u8a00\u5339\u914d\u7387\u5f88\u9ad8\uff08\u4f8b\u5982\uff0c\u7f57\u9a6c\u5316\u5370\u5730\u8bed\u63d0\u793a\u4e0d\u5e94\u671f\u671b\u4f7f\u7528\u5370\u5730\u8bed\u68b5\u6587\u811a\u672c\u8fdb\u884c\u56de\u590d\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u7ffb\u8bd1\u6570\u636e\uff1a<\/strong>\u6211\u4eec\u8bd5\u56fe\u907f\u514d\u4f7f\u7528\u673a\u5668\u7ffb\u8bd1\u6570\u636e\u6765\u5fae\u8c03\u6a21\u578b\uff0c\u4ee5\u9632\u6b62\u51fa\u73b0\u7ffb\u8bd1\u5f0f\u82f1\u8bed\uff08Bizzoni \u7b49\u4eba\uff0c2020 \u5e74\uff1bMuennighoff \u7b49\u4eba\uff0c2023 \u5e74\uff09\u6216\u53ef\u80fd\u51fa\u73b0\u7684\u59d3\u540d\u504f\u89c1\uff08Wang \u7b49\u4eba\uff0c2022a\uff09\u3001\u6027\u522b\u504f\u89c1\uff08Savoldi \u7b49\u4eba\uff0c2021 \u5e74\uff09\u6216\u6587\u5316\u504f\u89c1\uff08Ji \u7b49\u4eba\uff0c2023 \u5e74\uff09\u3002\u6b64\u5916\uff0c\u6211\u4eec\u65e8\u5728\u9632\u6b62\u6a21\u578b\u4ec5\u63a5\u89e6\u6839\u690d\u4e8e\u82f1\u8bed\u6587\u5316\u80cc\u666f\u7684\u4efb\u52a1\uff0c\u8fd9\u4e9b\u4efb\u52a1\u53ef\u80fd\u65e0\u6cd5\u4ee3\u8868\u6211\u4eec\u65e8\u5728\u6355\u6349\u7684\u8bed\u8a00\u548c\u6587\u5316\u591a\u6837\u6027\u3002\u6211\u4eec\u5bf9\u6b64\u505a\u4e86\u4e00\u4e2a\u4f8b\u5916\uff0c\u5e76\u5c06\u5408\u6210\u7684\u5b9a\u91cf\u63a8\u7406\u6570\u636e\uff08\u6709\u5173\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u89c1\u7b2c 4.3.3 \u8282\uff09\u7ffb\u8bd1\u6210\u975e\u82f1\u8bed\uff0c\u4ee5\u63d0\u9ad8\u975e\u82f1\u8bed\u8bed\u8a00\u4e2d\u7684\u5b9a\u91cf\u63a8\u7406\u6027\u80fd\u3002\u7531\u4e8e\u8fd9\u4e9b\u6570\u5b66\u95ee\u9898\u7684\u8bed\u8a00\u672c\u8d28\u7b80\u5355\uff0c\u56e0\u6b64\u53d1\u73b0\u7ffb\u8bd1\u540e\u7684\u6837\u672c\u51e0\u4e4e\u6ca1\u6709\u8d28\u91cf\u95ee\u9898\u3002\u6211\u4eec\u89c2\u5bdf\u5230\u4ece\u6dfb\u52a0\u6b64\u7ffb\u8bd1\u6570\u636e\u7684 MGSM\uff08Shi \u7b49\u4eba\uff0c2022 \u5e74\uff09\u83b7\u5f97\u4e86\u663e\u8457\u7684\u6536\u76ca\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h4>4.3.3 \u6570\u5b66\u4e0e\u63a8\u7406<\/h4>\n<p>\u6211\u4eec\u5b9a\u4e49\u63a8\u7406\u4e3a\u6267\u884c\u591a\u6b65\u8ba1\u7b97\u5e76\u5f97\u51fa\u6b63\u786e\u6700\u7ec8\u7b54\u6848\u7684\u80fd\u529b\u3002<\/p>\n<p>\u51e0\u4e2a\u6311\u6218\u6307\u5bfc\u4e86\u6211\u4eec\u8bad\u7ec3\u64c5\u957f\u6570\u5b66\u63a8\u7406\u6a21\u578b\u7684\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li><strong>\u7f3a\u4e4f\u63d0\u793a:<\/strong>\u00a0\u968f\u7740\u95ee\u9898\u590d\u6742\u5ea6\u7684\u589e\u52a0\uff0c\u7528\u4e8e\u76d1\u7763\u5fae\u8c03 (SFT) \u7684\u6709\u6548\u63d0\u793a\u6216\u95ee\u9898\u7684\u6570\u91cf\u51cf\u5c11\u3002\u8fd9\u79cd\u7a00\u7f3a\u6027\u4f7f\u5f97\u521b\u5efa\u591a\u6837\u5316\u4e14\u5177\u6709\u4ee3\u8868\u6027\u7684\u8bad\u7ec3\u6570\u636e\u96c6\u6765\u6559\u6388\u6a21\u578b\u5404\u79cd\u6570\u5b66\u6280\u80fd\u53d8\u5f97\u56f0\u96be\uff08Yu \u7b49\u4eba\uff0c2023\uff1bYue \u7b49\u4eba\uff0c2023\uff1bLuo \u7b49\u4eba\uff0c2023\uff1bMitra \u7b49\u4eba\uff0c2024\uff1bShao \u7b49\u4eba\uff0c2024\uff1bYue \u7b49\u4eba\uff0c2024b\uff09\u3002<\/li>\n<li><strong>\u7f3a\u4e4f\u771f\u5b9e\u63a8\u7406\u8fc7\u7a0b:<\/strong>\u00a0\u6709\u6548\u7684\u63a8\u7406\u9700\u8981\u9010\u6b65\u89e3\u51b3\u65b9\u6848\u6765\u4fc3\u8fdb\u63a8\u7406\u8fc7\u7a0b\uff08Wei \u7b49\u4eba\uff0c2022c\uff09\u3002\u7136\u800c\uff0c\u901a\u5e38\u7f3a\u5c11\u771f\u5b9e\u63a8\u7406\u8fc7\u7a0b\uff0c\u8fd9\u4e9b\u8fc7\u7a0b\u5bf9\u4e8e\u6307\u5bfc\u6a21\u578b\u5982\u4f55\u9010\u6b65\u5206\u89e3\u95ee\u9898\u5e76\u5f97\u51fa\u6700\u7ec8\u7b54\u6848\u81f3\u5173\u91cd\u8981\uff08Zelikman \u7b49\u4eba\uff0c2022\uff09\u3002<\/li>\n<li><strong>\u9519\u8bef\u7684\u4e2d\u95f4\u6b65\u9aa4:<\/strong>\u00a0\u4f7f\u7528\u6a21\u578b\u751f\u6210\u7684\u63a8\u7406\u94fe\u65f6\uff0c\u4e2d\u95f4\u6b65\u9aa4\u53ef\u80fd\u5e76\u4e0d\u603b\u662f\u6b63\u786e\uff08Cobbe \u7b49\u4eba\uff0c2021\uff1bUesato \u7b49\u4eba\uff0c2022\uff1bLightman \u7b49\u4eba\uff0c2023\uff1bWang \u7b49\u4eba\uff0c2023a\uff09\u3002\u8fd9\u79cd\u4e0d\u51c6\u786e\u6027\u4f1a\u5bfc\u81f4\u6700\u7ec8\u7b54\u6848\u9519\u8bef\uff0c\u9700\u8981\u89e3\u51b3\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u6a21\u578b\u4f7f\u7528\u5916\u90e8\u5de5\u5177:<\/strong>\u00a0\u589e\u5f3a\u6a21\u578b\u4ee5\u5229\u7528\u5916\u90e8\u5de5\u5177\uff08\u5982\u4ee3\u7801\u89e3\u91ca\u5668\uff09\u7684\u80fd\u529b\uff0c\u4f7f\u5176\u53ef\u4ee5\u901a\u8fc7\u4ea4\u7ec7\u4ee3\u7801\u548c\u6587\u672c\u8fdb\u884c\u63a8\u7406\uff08Gao \u7b49\u4eba\uff0c2023\uff1bChen \u7b49\u4eba\uff0c2022\uff1bGou \u7b49\u4eba\uff0c2023\uff09\u3002\u8fd9\u79cd\u80fd\u529b\u53ef\u4ee5\u663e\u7740\u63d0\u9ad8\u4ed6\u4eec\u7684\u95ee\u9898\u89e3\u51b3\u80fd\u529b\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u548c\u63a8\u7406\u4e4b\u95f4\u7684\u5dee\u5f02:<\/strong>\u00a0\u6a21\u578b\u5728\u8bad\u7ec3\u671f\u95f4\u8fdb\u884c\u5fae\u8c03\u7684\u65b9\u5f0f\u901a\u5e38\u4e0e\u5728\u63a8\u7406\u671f\u95f4\u4f7f\u7528\u7684\u65b9\u5f0f\u5b58\u5728\u5dee\u5f02\u3002\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u7ecf\u8fc7\u5fae\u8c03\u7684\u6a21\u578b\u53ef\u80fd\u4f1a\u4e0e\u4eba\u7c7b\u6216\u5176\u4ed6\u6a21\u578b\u4e92\u52a8\uff0c\u9700\u8981\u901a\u8fc7\u53cd\u9988\u6765\u6539\u8fdb\u5176\u63a8\u7406\u80fd\u529b\u3002\u786e\u4fdd\u8bad\u7ec3\u548c\u5b9e\u9645\u5e94\u7528\u4e4b\u95f4\u7684\u4e00\u81f4\u6027\u5bf9\u4e8e\u7ef4\u6301\u63a8\u7406\u6027\u80fd\u81f3\u5173\u91cd\u8981\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\uff0c\u6211\u4eec\u5e94\u7528\u4ee5\u4e0b\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li><strong>\u89e3\u51b3\u63d0\u793a\u7f3a\u4e4f\u95ee\u9898:<\/strong>\u00a0\u6211\u4eec\u4ece\u6570\u5b66\u4e0a\u4e0b\u6587\u4e2d\u83b7\u53d6\u76f8\u5173\u9884\u8bad\u7ec3\u6570\u636e\uff0c\u5e76\u5c06\u5176\u8f6c\u6362\u4e3a\u95ee\u9898-\u7b54\u6848\u683c\u5f0f\uff0c\u53ef\u7528\u4e8e\u76d1\u7763\u5fae\u8c03\u3002\u6b64\u5916\uff0c\u6211\u4eec\u4f1a\u8bc6\u522b\u6a21\u578b\u8868\u73b0\u4e0d\u4f73\u7684\u6570\u5b66\u6280\u80fd\uff0c\u5e76\u79ef\u6781\u5730\u4ece\u4eba\u7c7b\u90a3\u91cc\u6536\u96c6\u63d0\u793a\u4ee5\u6559\u6388\u6a21\u578b\u8fd9\u4e9b\u6280\u80fd\u3002\u4e3a\u4e86\u4fc3\u8fdb\u8fd9\u4e00\u8fc7\u7a0b\uff0c\u6211\u4eec\u521b\u5efa\u4e86\u4e00\u4e2a\u6570\u5b66\u6280\u80fd\u5206\u7c7b\u6cd5\uff08Didolkar \u7b49\u4eba\uff0c2024\uff09\uff0c\u5e76\u8981\u6c42\u4eba\u7c7b\u63d0\u4f9b\u76f8\u5e94\u7684\u63d0\u793a\/\u95ee\u9898\u3002<\/li>\n<li><strong>\u7528\u9010\u6b65\u63a8\u7406\u6b65\u9aa4\u589e\u5f3a\u8bad\u7ec3\u6570\u636e:<\/strong>\u00a0\u6211\u4eec\u4f7f\u7528 Llama 3 \u4e3a\u4e00\u7ec4\u63d0\u793a\u751f\u6210\u9010\u6b65\u89e3\u51b3\u65b9\u6848\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u63d0\u793a\uff0c\u6a21\u578b\u90fd\u4f1a\u4ea7\u751f\u53ef\u53d8\u6570\u91cf\u7684\u751f\u6210\u7ed3\u679c\u3002\u7136\u540e\u6839\u636e\u6b63\u786e\u7b54\u6848\u8fc7\u6ee4\u8fd9\u4e9b\u751f\u6210\u7ed3\u679c\uff08Li \u7b49\u4eba\uff0c2024a\uff09\u3002\u6211\u4eec\u8fd8\u8fdb\u884c\u81ea\u6211\u9a8c\u8bc1\uff0c\u5176\u4e2d Llama 3 \u7528\u4e8e\u9a8c\u8bc1\u7279\u5b9a\u9010\u6b65\u89e3\u51b3\u65b9\u6848\u662f\u5426\u5bf9\u7ed9\u5b9a\u7684\u95ee\u9898\u6709\u6548\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u901a\u8fc7\u6d88\u9664\u6a21\u578b\u4e0d\u4ea7\u751f\u6709\u6548\u63a8\u7406\u8f68\u8ff9\u7684\u5b9e\u4f8b\u6765\u63d0\u9ad8\u5fae\u8c03\u6570\u636e\u7684\u8d28\u91cf\u3002<\/li>\n<li><strong>\u8fc7\u6ee4\u9519\u8bef\u7684\u63a8\u7406\u6b65\u9aa4:<\/strong>\u00a0\u6211\u4eec\u8bad\u7ec3\u7ed3\u679c\u548c\u9010\u6b65\u5956\u52b1\u6a21\u578b\uff08Lightman \u7b49\u4eba\uff0c2023\uff1bWang \u7b49\u4eba\uff0c2023a\uff09\u6765\u8fc7\u6ee4\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u4e0d\u6b63\u786e\u7684\u8bad\u7ec3\u6570\u636e\u3002\u8fd9\u4e9b\u5956\u52b1\u6a21\u578b\u7528\u4e8e\u6d88\u9664\u5177\u6709\u65e0\u6548\u9010\u6b65\u63a8\u7406\u7684\u6570\u636e\uff0c\u786e\u4fdd\u5fae\u8c03\u5f97\u5230\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u3002\u5bf9\u4e8e\u66f4\u5177\u6311\u6218\u6027\u7684\u63d0\u793a\uff0c\u6211\u4eec\u4f7f\u7528\u5e26\u6709\u5b66\u4e60\u9010\u6b65\u5956\u52b1\u6a21\u578b\u7684\u8499\u7279\u5361\u7f57\u6811\u641c\u7d22 (MCTS) \u6765\u751f\u6210\u6709\u6548\u7684\u63a8\u7406\u8f68\u8ff9\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u589e\u5f3a\u9ad8\u8d28\u91cf\u63a8\u7406\u6570\u636e\u7684\u6536\u96c6\uff08Xie \u7b49\u4eba\uff0c2024\uff09\u3002<\/li>\n<li><strong>\u4ea4\u7ec7\u4ee3\u7801\u548c\u6587\u672c\u63a8\u7406:<\/strong>\u00a0\u6211\u4eec\u63d0\u793a Llama 3 \u901a\u8fc7\u6587\u672c\u63a8\u7406\u53ca\u5176\u5173\u8054 Python \u4ee3\u7801\u7ec4\u5408\u6765\u89e3\u51b3\u63a8\u7406\u95ee\u9898\uff08Gou \u7b49\u4eba\uff0c2023\uff09\u3002\u4ee3\u7801\u6267\u884c\u7528\u4f5c\u53cd\u9988\u4fe1\u53f7\uff0c\u4ee5\u6d88\u9664\u63a8\u7406\u94fe\u65e0\u6548\u7684\u60c5\u51b5\uff0c\u786e\u4fdd\u63a8\u7406\u8fc7\u7a0b\u7684\u6b63\u786e\u6027\u3002<\/li>\n<li><strong>\u4ece\u53cd\u9988\u548c\u9519\u8bef\u4e2d\u5b66\u4e60:<\/strong>\u00a0\u4e3a\u4e86\u6a21\u62df\u4eba\u7c7b\u53cd\u9988\uff0c\u6211\u4eec\u5229\u7528\u9519\u8bef\u7684\u751f\u6210\u7ed3\u679c\uff08\u5373\u5bfc\u81f4\u9519\u8bef\u63a8\u7406\u8f68\u8ff9\u7684\u751f\u6210\u7ed3\u679c\uff09\uff0c\u5e76\u901a\u8fc7\u63d0\u793a Llama 3 \u751f\u6210\u6b63\u786e\u7684\u751f\u6210\u7ed3\u679c\u6765\u8fdb\u884c\u9519\u8bef\u66f4\u6b63\uff08An \u7b49\u4eba\uff0c2023b\uff1bWelleck \u7b49\u4eba\uff0c2022\uff1bMadaan \u7b49\u4eba\uff0c2024a\uff09\u3002\u4f7f\u7528\u9519\u8bef\u5c1d\u8bd5\u7684\u53cd\u9988\u5e76\u7ea0\u6b63\u5b83\u4eec\u7684\u81ea\u8fed\u4ee3\u8fc7\u7a0b\u6709\u52a9\u4e8e\u63d0\u9ad8\u6a21\u578b\u51c6\u786e\u63a8\u7406\u7684\u80fd\u529b\u5e76\u4ece\u9519\u8bef\u4e2d\u5b66\u4e60\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h4>4.3.4 \u957f\u4e0a\u4e0b\u6587<\/h4>\n<p>\u5728\u6700\u540e\u7684\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u6211\u4eec\u5c06Llama 3\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u4ece8K\u4e2a\u6807\u8bb0\u6269\u5c55\u5230128K\u4e2a\u6807\u8bb0\uff08\u6709\u5173\u8be6\u7ec6\u4fe1\u606f\u8bf7\u53c2\u89c1\u7b2c3.4\u8282\uff09\u3002\u4e0e\u9884\u8bad\u7ec3\u7c7b\u4f3c\uff0c\u6211\u4eec\u53d1\u73b0\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\uff0c\u5fc5\u987b\u4ed4\u7ec6\u8c03\u6574\u914d\u65b9\u4ee5\u5e73\u8861\u77ed\u4e0a\u4e0b\u6587\u548c\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u3002<\/p>\n<p><strong>SFT \u548c\u5408\u6210\u6570\u636e\u751f\u6210\u3002<\/strong>\u00a0\u7b80\u5355\u5730\u5e94\u7528\u6211\u4eec\u7684\u73b0\u6709 SFT \u914d\u65b9\uff0c\u53ea\u4f7f\u7528\u77ed\u4e0a\u4e0b\u6587\u6570\u636e\u5bfc\u81f4\u9884\u8bad\u7ec3\u4e2d\u7684\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u663e\u8457\u4e0b\u964d\uff0c\u7a81\u51fa\u4e86\u5728 SFT \u6570\u636e\u7ec4\u5408\u4e2d\u7eb3\u5165\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u7684\u5fc5\u8981\u6027\u3002\u7136\u800c\uff0c\u5b9e\u9645\u4e0a\uff0c\u8ba9\u4eba\u5de5\u6807\u6ce8\u8fd9\u4e9b\u793a\u4f8b\u5927\u90e8\u5206\u662f\u4e0d\u5207\u5b9e\u9645\u7684\uff0c\u56e0\u4e3a\u9605\u8bfb\u5197\u957f\u7684\u4e0a\u4e0b\u6587\u975e\u5e38\u7e41\u7410\u4e14\u8017\u65f6\uff0c\u56e0\u6b64\u6211\u4eec\u4e3b\u8981\u4f9d\u8d56\u5408\u6210\u6570\u636e\u6765\u5f25\u8865\u8fd9\u4e00\u5dee\u8ddd\u3002 \u6211\u4eec\u4f7f\u7528\u65e9\u671f\u7248\u672c\u7684 Llama 3 \u751f\u6210\u5408\u6210\u6570\u636e\uff0c\u57fa\u4e8e\u5173\u952e\u7684\u957f\u4e0a\u4e0b\u6587\u7528\u4f8b\uff1a\uff08\u53ef\u80fd\u4e3a\u591a\u8f6e\uff09\u95ee\u7b54\u3001\u957f\u6587\u6863\u6458\u8981\u4ee5\u53ca\u4ee3\u7801\u5e93\u4e0a\u7684\u63a8\u7406\uff0c\u5e76\u5728\u4e0b\u9762\u66f4\u8be6\u7ec6\u5730\u63cf\u8ff0\u8fd9\u4e9b\u7528\u4f8b\u3002<\/p>\n<ul>\n<li><strong>\u95ee\u7b54\uff1a<\/strong>\u00a0\u6211\u4eec\u4ece\u9884\u8bad\u7ec3\u6570\u636e\u96c6\u4e2d\u4ed4\u7ec6\u6311\u9009\u4e86\u4e00\u7ec4\u957f\u6587\u6863\u3002\u6211\u4eec\u5c06\u8fd9\u4e9b\u6587\u6863\u62c6\u5206\u62108K\u4e2a\u6807\u8bb0\u7684\u5757\uff0c\u5e76\u63d0\u793a\u65e9\u671f\u7248\u672c\u7684 Llama 3 \u6a21\u578b\u5728\u968f\u673a\u9009\u62e9\u7684\u5757\u4e0a\u751f\u6210 QA \u5bf9\u3002 \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6574\u4e2a\u6587\u6863\u90fd\u88ab\u7528\u4f5c\u4e0a\u4e0b\u6587\u3002<\/li>\n<li><strong>\u6458\u8981\uff1a<\/strong>\u00a0\u6211\u4eec\u901a\u8fc7\u9996\u5148\u4f7f\u7528\u6211\u4eec\u6700\u5f3a\u7684 Llama 3 8K \u4e0a\u4e0b\u6587\u6a21\u578b\u5bf9 8K \u8f93\u5165\u957f\u5ea6\u7684\u5757\u8fdb\u884c\u5c42\u6b21\u5316\u6458\u8981\u6765\u5e94\u7528\u957f\u4e0a\u4e0b\u6587\u6587\u6863\u7684\u5c42\u6b21\u5316\u6458\u8981\u3002\u7136\u540e\u5bf9\u8fd9\u4e9b\u6458\u8981\u8fdb\u884c\u6c47\u603b\u3002 \u5728\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u63d0\u4f9b\u5b8c\u6574\u6587\u6863\u5e76\u63d0\u793a\u6a21\u578b\u603b\u7ed3\u6587\u6863\uff0c\u540c\u65f6\u4fdd\u7559\u6240\u6709\u91cd\u8981\u7ec6\u8282\u3002 \u6211\u4eec\u8fd8\u57fa\u4e8e\u6587\u6863\u7684\u6458\u8981\u751f\u6210 QA \u5bf9\uff0c\u5e76\u7528\u9700\u8981\u5bf9\u6574\u4e2a\u957f\u6587\u6863\u5168\u5c40\u7406\u89e3\u7684\u95ee\u9898\u63d0\u793a\u6a21\u578b\u3002<\/li>\n<li><strong>\u957f\u4e0a\u4e0b\u6587\u4ee3\u7801\u63a8\u7406\uff1a<\/strong>\u00a0\u6211\u4eec\u89e3\u6790 Python \u6587\u4ef6\u4ee5\u8bc6\u522b\u5bfc\u5165\u8bed\u53e5\u5e76\u786e\u5b9a\u5b83\u4eec\u7684\u4f9d\u8d56\u5173\u7cfb\u3002 \u4ece\u8fd9\u91cc\u5f00\u59cb\uff0c\u6211\u4eec\u9009\u62e9\u6700\u5e38\u7528\u7684\u6587\u4ef6\uff0c\u7279\u522b\u662f\u90a3\u4e9b\u88ab\u81f3\u5c11\u5176\u4ed6\u4e94\u4e2a\u6587\u4ef6\u5f15\u7528\u7684\u6587\u4ef6\u3002 \u6211\u4eec\u4ece\u5b58\u50a8\u5e93\u4e2d\u5220\u9664\u8fd9\u4e9b\u5173\u952e\u6587\u4ef6\u4e2d\u7684\u4e00\u4e2a\uff0c\u5e76\u63d0\u793a\u6a21\u578b\u8bc6\u522b\u4f9d\u8d56\u4e8e\u7f3a\u5c11\u7684\u6587\u4ef6\uff0c\u5e76\u751f\u6210\u5fc5\u8981\u7684\u7f3a\u5931\u4ee3\u7801\u3002<\/li>\n<\/ul>\n<p>\u6211\u4eec\u8fdb\u4e00\u6b65\u6839\u636e\u5e8f\u5217\u957f\u5ea6\uff0816K\u300132K\u300164K \u548c 128K\uff09\u5bf9\u8fd9\u4e9b\u5408\u6210\u751f\u6210\u7684\u6837\u672c\u8fdb\u884c\u5206\u7c7b\uff0c\u4ee5\u5b9e\u73b0\u66f4\u7cbe\u7ec6\u7684\u8f93\u5165\u957f\u5ea6\u5b9a\u4f4d\u3002<\/p>\n<p>\u901a\u8fc7\u4ed4\u7ec6\u7684\u6d88\u878d\u5b9e\u9a8c\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u5c06 0.1% \u7684\u5408\u6210\u751f\u6210\u7684\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u4e0e\u539f\u59cb\u77ed\u4e0a\u4e0b\u6587\u6570\u636e\u6df7\u5408\u53ef\u4ee5\u4f18\u5316\u77ed\u4e0a\u4e0b\u6587\u548c\u957f\u4e0a\u4e0b\u6587\u57fa\u51c6\u6d4b\u8bd5\u7684\u6027\u80fd\u3002<\/p>\n<p><strong>DPO\u3002<\/strong>\u00a0\u6211\u4eec\u6ce8\u610f\u5230\u5728 DPO \u4e2d\u4ec5\u4f7f\u7528\u77ed\u4e0a\u4e0b\u6587\u8bad\u7ec3\u6570\u636e\u4e0d\u4f1a\u5bf9\u957f\u4e0a\u4e0b\u6587\u6027\u80fd\u4ea7\u751f\u8d1f\u9762\u5f71\u54cd\uff0c\u53ea\u8981 SFT \u6a21\u578b\u5bf9\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u6548\u679c\u826f\u597d\u3002\u6211\u4eec\u6000\u7591\u8fd9\u662f\u56e0\u4e3a\u6211\u4eec\u7684 DPO \u914d\u65b9\u6bd4 SFT \u7684\u4f18\u5316\u5668\u6b65\u9aa4\u66f4\u5c11\u3002\u8003\u8651\u5230\u8fd9\u4e00\u53d1\u73b0\uff0c\u6211\u4eec\u5728\u957f\u4e0a\u4e0b\u6587 SFT \u68c0\u67e5\u70b9\u4e4b\u4e0a\u4fdd\u7559\u6807\u51c6\u7684\u77ed\u4e0a\u4e0b\u6587 DPO \u914d\u65b9\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.3.5 \u5de5\u5177\u4f7f\u7528<\/h4>\n<p>\u6559\u5bfc\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4f7f\u7528\u8bf8\u5982\u641c\u7d22\u5f15\u64ce\u6216\u4ee3\u7801\u89e3\u91ca\u5668\u4e4b\u7c7b\u7684\u5de5\u5177\uff0c\u53ef\u4ee5\u5927\u5927\u6269\u5c55\u5b83\u4eec\u80fd\u591f\u89e3\u51b3\u7684\u4efb\u52a1\u8303\u56f4\uff0c\u5c06\u5b83\u4eec\u4ece\u7eaf\u7cb9\u7684\u804a\u5929\u6a21\u578b\u8f6c\u53d8\u4e3a\u66f4\u901a\u7528\u7684\u52a9\u624b\uff08Nakano \u7b49\u4eba\uff0c2021 \u5e74\uff1bThoppilan \u7b49\u4eba\uff0c2022 \u5e74\uff1bParisi \u7b49\u4eba\uff0c2022 \u5e74\uff1bGao \u7b49\u4eba\uff0c2023 \u5e74\uff1bMialon \u7b49\u4eba\uff0c2023a\uff1bSchick \u7b49\u4eba\uff0c2024 \u5e74\uff09\u3002\u6211\u4eec\u8bad\u7ec3 Llama 3 \u4e0e\u4ee5\u4e0b\u5de5\u5177\u8fdb\u884c\u4ea4\u4e92\uff1a<\/p>\n<ul>\n<li>\u641c\u7d22\u5f15\u64ce\u3002Llama 3 \u88ab\u8bad\u7ec3\u4f7f\u7528 Brave Search7 \u6765\u56de\u7b54\u5173\u4e8e\u5176\u77e5\u8bc6\u622a\u6b62\u65e5\u671f\u4e4b\u540e\u7684\u8fd1\u671f\u4e8b\u4ef6\u7684\u95ee\u9898\uff0c\u6216\u8005\u9700\u8981\u4ece\u7f51\u7edc\u4e2d\u68c0\u7d22\u7279\u5b9a\u4fe1\u606f\u7684\u8bf7\u6c42\u3002<\/li>\n<li>Python \u89e3\u91ca\u5668\u3002Llama 3 \u53ef\u4ee5\u751f\u6210\u5e76\u6267\u884c\u4ee3\u7801\u4ee5\u6267\u884c\u590d\u6742\u7684\u8ba1\u7b97\uff0c\u8bfb\u53d6\u7528\u6237\u4e0a\u4f20\u7684\u6587\u4ef6\u5e76\u6839\u636e\u8fd9\u4e9b\u6587\u4ef6\u89e3\u51b3\u4efb\u52a1\uff0c\u4f8b\u5982\u95ee\u7b54\u3001\u6458\u8981\u3001\u6570\u636e\u5206\u6790\u6216\u53ef\u89c6\u5316\u3002<\/li>\n<li>\u6570\u5b66\u8ba1\u7b97\u5f15\u64ce\u3002Llama 3 \u53ef\u4ee5\u4f7f\u7528 Wolfram Alpha API8 \u66f4\u51c6\u786e\u5730\u89e3\u51b3\u6570\u5b66\u548c\u79d1\u5b66\u95ee\u9898\uff0c\u6216\u8005\u4ece Wolfram \u7684\u6570\u636e\u5e93\u4e2d\u68c0\u7d22\u51c6\u786e\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<p>\u751f\u6210\u7684\u6a21\u578b\u80fd\u591f\u5728\u804a\u5929\u8bbe\u7f6e\u4e2d\u4f7f\u7528\u8fd9\u4e9b\u5de5\u5177\u6765\u89e3\u51b3\u7528\u6237\u7684\u67e5\u8be2\uff0c\u5305\u62ec\u591a\u56de\u5408\u5bf9\u8bdd\u3002\u5982\u679c\u67e5\u8be2\u9700\u8981\u591a\u6b21\u8c03\u7528\u5de5\u5177\uff0c\u8be5\u6a21\u578b\u53ef\u4ee5\u7f16\u5199\u9010\u6b65\u8ba1\u5212\uff0c\u4f9d\u6b21\u8c03\u7528\u5de5\u5177\uff0c\u5e76\u5728\u6bcf\u6b21\u5de5\u5177\u8c03\u7528\u540e\u8fdb\u884c\u63a8\u7406\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u63d0\u9ad8\u4e86 Llama 3 \u7684\u96f6\u6837\u672c\u5de5\u5177\u4f7f\u7528\u80fd\u529b\u2014\u2014\u7ed9\u5b9a\u4e0a\u4e0b\u6587\u73af\u5883\u4e0b\u53ef\u80fd\u672a\u89c1\u8fc7\u7684\u5de5\u5177\u5b9a\u4e49\u548c\u7528\u6237\u67e5\u8be2\uff0c\u6211\u4eec\u8bad\u7ec3\u6a21\u578b\u751f\u6210\u6b63\u786e\u7684\u5de5\u5177\u8c03\u7528\u3002<\/p>\n<p><strong>\u5b9e\u73b0\u3002<\/strong>\u6211\u4eec\u5c06\u6838\u5fc3\u5de5\u5177\u5b9e\u65bd\u4e3a\u5177\u6709\u4e0d\u540c\u65b9\u6cd5\u7684 Python \u5bf9\u8c61\u3002\u96f6\u6837\u672c\u5de5\u5177\u53ef\u4ee5\u4f5c\u4e3a\u5e26\u6709\u63cf\u8ff0\u3001\u6587\u6863\uff08\u5373\u5982\u4f55\u4f7f\u7528\u5b83\u4eec\u7684\u793a\u4f8b\uff09\u7684 Python \u51fd\u6570\u6765\u5b9e\u73b0\uff0c\u6a21\u578b\u53ea\u9700\u8981\u51fd\u6570\u7b7e\u540d\u548c docstring \u4f5c\u4e3a\u4e0a\u4e0b\u6587\u5373\u53ef\u751f\u6210\u9002\u5f53\u7684\u8c03\u7528\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u5c06\u51fd\u6570\u5b9a\u4e49\u548c\u8c03\u7528\u8f6c\u6362\u4e3a JSON \u683c\u5f0f\uff0c\u4f8b\u5982\u7528\u4e8e Web API \u8c03\u7528\u3002\u6240\u6709\u5de5\u5177\u8c03\u7528\u90fd\u7531 Python \u89e3\u91ca\u5668\u6267\u884c\uff0cPython \u89e3\u91ca\u5668\u5fc5\u987b\u5728 Llama 3 \u7cfb\u7edf\u63d0\u793a\u4e2d\u542f\u7528\u3002\u53ef\u4ee5\u5728\u7cfb\u7edf\u63d0\u793a\u4e2d\u5355\u72ec\u542f\u7528\u6216\u7981\u7528\u6838\u5fc3\u5de5\u5177\u3002<\/p>\n<p><strong>\u6570\u636e\u6536\u96c6\u3002<\/strong>\u4e0e Schick \u7b49\u4eba (2024) \u4e0d\u540c\uff0c\u6211\u4eec\u4f9d\u9760\u4eba\u7c7b\u6807\u6ce8\u548c\u504f\u597d\u6765\u6559 Llama 3 \u4f7f\u7528\u5de5\u5177\u3002\u8fd9\u4e0e Llama 3 \u4e2d\u901a\u5e38\u4f7f\u7528\u7684\u540e\u8bad\u7ec3\u7ba1\u9053\u5b58\u5728\u4e24\u4e2a\u4e3b\u8981\u533a\u522b\uff1a<\/p>\n<ul>\n<li>\u5173\u4e8e\u5de5\u5177\uff0c\u5bf9\u8bdd\u5e38\u5e38\u5305\u542b\u4e0d\u6b62\u4e00\u6761\u52a9\u7406\u6d88\u606f\uff08\u4f8b\u5982\uff0c\u8c03\u7528\u5de5\u5177\u5e76\u5bf9\u5de5\u5177\u8f93\u51fa\u8fdb\u884c\u63a8\u7406\uff09\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u8fdb\u884c\u6d88\u606f\u7ea7\u522b\u7684\u6807\u6ce8\u4ee5\u6536\u96c6\u8be6\u7ec6\u7684\u53cd\u9988\uff1a\u6807\u6ce8\u8005\u5728\u76f8\u540c\u4e0a\u4e0b\u6587\u4e0b\u63d0\u4f9b\u4e24\u6761\u52a9\u7406\u6d88\u606f\u7684\u504f\u597d\uff0c\u6216\u8005\u5982\u679c\u4e24\u8005\u90fd\u5b58\u5728\u4e3b\u8981\u95ee\u9898\uff0c\u5219\u7f16\u8f91\u5176\u4e2d\u4e00\u6761\u6d88\u606f\u3002\u7136\u540e\u5c06\u88ab\u9009\u62e9\u6216\u4fee\u6539\u7684\u6d88\u606f\u6dfb\u52a0\u5230\u4e0a\u4e0b\u6587\u4e2d\uff0c\u5bf9\u8bdd\u7ee7\u7eed\u8fdb\u884c\u3002\u8fd9\u4e3a\u52a9\u7406\u8c03\u7528\u5de5\u5177\u548c\u63a8\u7406\u5de5\u5177\u8f93\u51fa\u7684\u80fd\u529b\u63d0\u4f9b\u4e86\u4eba\u7c7b\u53cd\u9988\u3002\u6807\u6ce8\u8005\u4e0d\u80fd\u5bf9\u5de5\u5177\u8f93\u51fa\u8fdb\u884c\u6392\u540d\u6216\u7f16\u8f91\u3002<\/li>\n<li>\u6211\u4eec\u6ca1\u6709\u6267\u884c\u62d2\u7edd\u91c7\u6837\uff0c\u56e0\u4e3a\u5728\u6211\u4eec\u7684\u5de5\u5177\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u6ca1\u6709\u89c2\u5bdf\u5230\u6536\u76ca\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u52a0\u901f\u6807\u6ce8\u8fc7\u7a0b\uff0c\u6211\u4eec\u9996\u5148\u901a\u8fc7\u5fae\u8c03\u6765\u81ea\u4e4b\u524d Llama 3 \u68c0\u67e5\u70b9\u7684\u5408\u6210\u6570\u636e\u6765\u5f15\u5bfc\u57fa\u672c\u7684\u5de5\u5177\u4f7f\u7528\u80fd\u529b\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u6807\u6ce8\u8005\u5c31\u9700\u8981\u8fdb\u884c\u8f83\u5c11\u7684\u7f16\u8f91\u64cd\u4f5c\u3002\u7c7b\u4f3c\u5730\uff0c\u968f\u7740 Llama 3 \u5728\u5f00\u53d1\u8fc7\u7a0b\u4e2d\u9010\u6e10\u6539\u8fdb\uff0c\u6211\u4eec\u4f1a\u9010\u6b65\u590d\u6742\u5316\u6211\u4eec\u7684\u4eba\u7c7b\u6807\u6ce8\u534f\u8bae\uff1a\u6211\u4eec\u4ece\u5355\u56de\u5408\u7684\u5de5\u5177\u4f7f\u7528\u6807\u6ce8\u5f00\u59cb\uff0c\u7136\u540e\u8f6c\u5411\u5bf9\u8bdd\u4e2d\u7684\u5de5\u5177\u4f7f\u7528\uff0c\u6700\u540e\u6ce8\u91ca\u591a\u6b65\u9aa4\u5de5\u5177\u4f7f\u7528\u548c\u6570\u636e\u5206\u6790\u3002<\/p>\n<p><strong>\u5de5\u5177\u6570\u636e\u96c6\u3002<\/strong>\u4e3a\u4e86\u521b\u5efa\u7528\u4e8e\u5de5\u5177\u4f7f\u7528\u5e94\u7528\u7a0b\u5e8f\u7684\u6570\u636e\uff0c\u6211\u4eec\u91c7\u7528\u4ee5\u4e0b\u6b65\u9aa4:<\/p>\n<ul>\n<li><strong>\u5355\u6b65\u5de5\u5177\u4f7f\u7528:<\/strong> \u6211\u4eec\u9996\u5148\u8fdb\u884c\u5c11\u91cf\u6837\u672c\u751f\u6210\u5408\u6210\u7528\u6237\u63d0\u793a\uff0c\u8fd9\u4e9b\u63d0\u793a\u901a\u8fc7\u6784\u9020\u9700\u8981\u8c03\u7528\u6211\u4eec\u7684\u6838\u5fc3\u5de5\u5177\u4e4b\u4e00\uff08\u4f8b\u5982\uff0c\u8d85\u8fc7\u6211\u4eec\u77e5\u8bc6\u622a\u6b62\u65e5\u671f\u7684\u95ee\u9898\uff09\u3002\u7136\u540e\uff0c\u4ecd\u7136\u4f9d\u9760\u5c11\u91cf\u6837\u672c\u751f\u6210\uff0c\u6211\u4eec\u4e3a\u8fd9\u4e9b\u63d0\u793a\u751f\u6210\u9002\u5f53\u7684\u5de5\u5177\u8c03\u7528\uff0c\u6267\u884c\u5b83\u4eec\uff0c\u5e76\u5c06\u8f93\u51fa\u6dfb\u52a0\u5230\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u4e2d\u3002\u6700\u540e\uff0c\u6211\u4eec\u518d\u6b21\u63d0\u793a\u6a21\u578b\u4ee5\u6839\u636e\u5de5\u5177\u8f93\u51fa\u751f\u6210\u5bf9\u7528\u6237\u67e5\u8be2\u7684\u6700\u7ec8\u7b54\u6848\u3002\u6211\u4eec\u6700\u7ec8\u5f97\u5230\u4ee5\u4e0b\u5f62\u5f0f\u7684\u8f68\u8ff9\uff1a\u7cfb\u7edf\u63d0\u793a\u3001\u7528\u6237\u63d0\u793a\u3001\u5de5\u5177\u8c03\u7528\u3001\u5de5\u5177\u8f93\u51fa\u3001\u6700\u7ec8\u7b54\u6848\u3002\u6211\u4eec\u8fd8\u8fc7\u6ee4\u6389\u4e86\u5927\u7ea6 30% \u7684\u6570\u636e\u96c6\uff0c\u4ee5\u5220\u9664\u65e0\u6cd5\u6267\u884c\u7684\u5de5\u5177\u8c03\u7528\u6216\u5176\u4ed6\u683c\u5f0f\u95ee\u9898\u3002<\/li>\n<li><strong>\u591a\u6b65\u9aa4\u5de5\u5177\u4f7f\u7528:<\/strong>\u00a0\u6211\u4eec\u9075\u5faa\u7c7b\u4f3c\u7684\u534f\u8bae\uff0c\u9996\u5148\u751f\u6210\u5408\u6210\u6570\u636e\u6765\u6559\u6388\u6a21\u578b\u57fa\u672c\u7684 \ub2e4\ub2e8\uacc4 \u5de5\u5177\u4f7f\u7528\u80fd\u529b\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u9996\u5148\u63d0\u793a Llama 3 \u751f\u6210\u81f3\u5c11\u9700\u8981\u4e24\u6b21\u5de5\u5177\u8c03\u7528\u7684\u7528\u6237\u63d0\u793a\uff08\u53ef\u4ee5\u662f\u6765\u81ea\u6211\u4eec\u7684\u6838\u5fc3\u96c6\u4e2d\u7684\u76f8\u540c\u5de5\u5177\u6216\u4e0d\u540c\u5de5\u5177\uff09\u3002\u7136\u540e\uff0c\u6839\u636e\u8fd9\u4e9b\u63d0\u793a\uff0c\u6211\u4eec\u8fdb\u884c\u5c11\u91cf\u6837\u672c\u63d0\u793a Llama 3 \u751f\u6210\u4e00\u4e2a\u89e3\u51b3\u65b9\u6848\uff0c\u8be5\u89e3\u51b3\u65b9\u6848\u7531\u4ea4\u7ec7\u7684\u63a8\u7406\u6b65\u9aa4\u548c\u5de5\u5177\u8c03\u7528\u7ec4\u6210\uff0c\u7c7b\u4f3c\u4e8e <a href=\"https:\/\/www.kdjingpai.com\/pt\/react\/\">ReAct<\/a> (Yao \u7b49\u4eba\uff0c2022)\u3002\u8bf7\u53c2\u89c1\u56fe 10\uff0c\u4ee5\u4e86\u89e3 Llama 3 \u6267\u884c\u6d89\u53ca\u591a\u6b65\u9aa4\u5de5\u5177\u4f7f\u7528\u4efb\u52a1\u7684\u793a\u4f8b\u3002<\/li>\n<li><strong>\u6587\u4ef6\u4e0a\u4f20:<\/strong>\u00a0\u6211\u4eec\u9488\u5bf9\u4ee5\u4e0b\u6587\u4ef6\u7c7b\u578b\u8fdb\u884c\u6807\u6ce8\uff1a.txt\uff0c .docx\uff0c .pdf\uff0c .pptx\uff0c .xlsx\uff0c .csv\uff0c .tsv\uff0c .py\uff0c .json\uff0c .jsonl\uff0c .html\uff0c .xml\u3002\u6211\u4eec\u7684\u63d0\u793a\u57fa\u4e8e\u63d0\u4f9b\u6587\u4ef6\uff0c\u5e76\u8981\u6c42\u603b\u7ed3\u6587\u4ef6\u5185\u5bb9\u3001\u67e5\u627e\u548c\u4fee\u590d\u9519\u8bef\u3001\u4f18\u5316\u4ee3\u7801\u6bb5\u3001\u6267\u884c\u6570\u636e\u5206\u6790\u6216\u53ef\u89c6\u5316\u3002\u56fe 11 \u663e\u793a\u4e86 Llama 3 \u6267\u884c\u6d89\u53ca\u6587\u4ef6\u4e0a\u4f20\u4efb\u52a1\u7684\u793a\u4f8b\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-3447\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b9d8be30ea0a9c3.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"1119\" height=\"750\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b9d8be30ea0a9c3.png 1119w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b9d8be30ea0a9c3-300x201.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b9d8be30ea0a9c3-1024x686.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b9d8be30ea0a9c3-768x515.png 768w\" sizes=\"auto, (max-width: 1119px) 100vw, 1119px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>\u5728\u5bf9\u6b64\u5408\u6210\u6570\u636e\u8fdb\u884c\u5fae\u8c03\u540e\uff0c\u6211\u4eec\u6536\u96c6\u4e86\u6765\u81ea\u5404\u79cd\u573a\u666f\u7684\u4eba\u7c7b\u6807\u6ce8\uff0c\u5305\u62ec\u591a\u8f6e\u4ea4\u4e92\u3001\u8d85\u8fc7\u4e09\u6b65\u7684\u5de5\u5177\u4f7f\u7528\u4ee5\u53ca\u5de5\u5177\u8c03\u7528\u65e0\u6cd5\u4ea7\u751f\u4ee4\u4eba\u6ee1\u610f\u7684\u7b54\u6848\u7684\u60c5\u51b5\u3002\u6211\u4eec\u7528\u4e0d\u540c\u7684\u7cfb\u7edf\u63d0\u793a\u589e\u5f3a\u5408\u6210\u6570\u636e\uff0c\u4ee5\u6559\u5bfc\u6a21\u578b\u4ec5\u5728\u88ab\u6fc0\u6d3b\u65f6\u624d\u4f7f\u7528\u5de5\u5177\u3002\u4e3a\u4e86\u8bad\u7ec3\u6a21\u578b\u907f\u514d\u5bf9\u7b80\u5355\u67e5\u8be2\u8fdb\u884c\u5de5\u5177\u8c03\u7528\uff0c\u6211\u4eec\u8fd8\u6dfb\u52a0\u4e86\u6765\u81ea\u6613\u4e8e\u8ba1\u7b97\u6216\u95ee\u7b54\u6570\u636e\u96c6\uff08Berant \u7b49\u4eba\uff0c2013\uff1bKoncel-Kedziorski \u7b49\u4eba\uff0c2016\uff1bJoshi \u7b49\u4eba\uff0c2017\uff1bAmini \u7b49\u4eba\uff0c2019\uff09\u7684\u67e5\u8be2\u53ca\u5176\u54cd\u5e94\uff0c\u8fd9\u4e9b\u54cd\u5e94\u4e0d\u4f7f\u7528\u5de5\u5177\uff0c\u4f46\u5728\u7cfb\u7edf\u63d0\u793a\u4e2d\u6fc0\u6d3b\u4e86\u5de5\u5177\u3002<\/p>\n<p><strong>\u96f6\u6837\u672c\u5de5\u5177\u4f7f\u7528\u6570\u636e:<\/strong>\u00a0\u6211\u4eec\u901a\u8fc7\u5bf9\u5927\u91cf\u4e14\u591a\u6837\u7684\u90e8\u5206\u5408\u6210\uff08\u51fd\u6570\u5b9a\u4e49\u3001\u7528\u6237\u67e5\u8be2\u3001\u76f8\u5e94\u8c03\u7528\uff09\u5143\u7ec4\u8fdb\u884c\u5fae\u8c03\uff0c\u4ece\u800c\u6539\u5584 Llama 3 \u7684\u96f6\u6837\u672c\u5de5\u5177\u4f7f\u7528\u80fd\u529b\uff08\u4e5f\u79f0\u4e3a\u529f\u80fd\u8c03\u7528\uff09\u3002\u6211\u4eec\u5728\u672a\u89c1\u8fc7\u5de5\u5177\u7684\u96c6\u5408\u4e0a\u8bc4\u4f30\u6211\u4eec\u7684\u6a21\u578b\u3002<\/p>\n<ul>\n<li><strong>\u5355\u4e00\u3001\u5d4c\u5957\u548c\u5e76\u884c\u51fd\u6570\u8c03\u7528\uff1a<\/strong>\u00a0\u8c03\u7528\u53ef\u4ee5\u662f\u7b80\u5355\u7684\u3001\u5d4c\u5957\u7684\uff08\u5373\u6211\u4eec\u5c06\u51fd\u6570\u8c03\u7528\u4f5c\u4e3a\u53e6\u4e00\u4e2a\u51fd\u6570\u7684\u53c2\u6570\u4f20\u9012\uff09\u6216\u5e73\u884c\u7684\uff08\u5373\u6a21\u578b\u8fd4\u56de\u4e00\u4e2a\u72ec\u7acb\u51fd\u6570\u8c03\u7528\u7684\u5217\u8868\uff09\u3002\u751f\u6210\u5404\u79cd\u51fd\u6570\u3001\u67e5\u8be2\u548c\u771f\u5b9e\u7ed3\u679c\u53ef\u80fd\u5177\u6709\u6311\u6218\u6027\uff08Mekala \u7b49\u4eba\uff0c2024\uff09\uff0c\u6211\u4eec\u4f9d\u9760\u6316\u6398 Stack\uff08Kocetkov \u7b49\u4eba\uff0c2022\uff09\u6765\u5c06\u6211\u4eec\u7684\u5408\u6210\u7528\u6237\u67e5\u8be2\u57fa\u4e8e\u771f\u5b9e\u7684\u51fd\u6570\u3002\u66f4\u51c6\u786e\u5730\u8bf4\uff0c\u6211\u4eec\u63d0\u53d6\u51fd\u6570\u8c03\u7528\u53ca\u5176\u5b9a\u4e49\uff0c\u6e05\u7406\u548c\u8fc7\u6ee4\u5b83\u4eec\uff08\u4f8b\u5982\uff0c\u7f3a\u5c11\u6587\u6863\u5b57\u7b26\u4e32\u6216\u4e0d\u53ef\u6267\u884c\u7684\u51fd\u6570\uff09\uff0c\u5e76\u4f7f\u7528 Llama 3 \u751f\u6210\u4e0e\u51fd\u6570\u8c03\u7528\u76f8\u5bf9\u5e94\u7684\u81ea\u7136\u8bed\u8a00\u67e5\u8be2\u3002<\/li>\n<li><strong>\u591a\u56de\u5408\u51fd\u6570\u8c03\u7528\uff1a<\/strong>\u00a0\u6211\u4eec\u8fd8\u4e3a\u5305\u542b\u51fd\u6570\u8c03\u7528\u7684\u591a\u56de\u5408\u5bf9\u8bdd\u751f\u6210\u5408\u6210\u6570\u636e\uff0c\u9075\u5faa\u7c7b\u4f3c\u4e8e Li \u7b49\u4eba\uff082023b\uff09\u4e2d\u63d0\u51fa\u7684\u534f\u8bae\u3002\u6211\u4eec\u4f7f\u7528\u591a\u4e2a\u4ee3\u7406\u6765\u751f\u6210\u57df\u3001API\u3001\u7528\u6237\u67e5\u8be2\u3001API \u8c03\u7528\u548c\u54cd\u5e94\uff0c\u540c\u65f6\u786e\u4fdd\u751f\u6210\u7684\u6578\u64da\u6db5\u76d6\u4e86\u4e00\u7cfb\u5217\u4e0d\u540c\u7684\u9886\u57df\u548c\u771f\u5b9e\u7684 API\u3002\u6240\u6709\u4ee3\u7406\u90fd\u662f Llama 3 \u7684\u53d8\u4f53\uff0c\u5176\u63d0\u793a\u65b9\u5f0f\u53d6\u51b3\u4e8e\u5b83\u4eec\u7684\u804c\u8d23\uff0c\u5e76\u4e14\u4ee5\u9010\u6b65\u7684\u65b9\u5f0f\u8fdb\u884c\u534f\u4f5c\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h4>4.3.6 \u4e8b\u5b9e\u6027<\/h4>\n<p>\u865a\u5e7b\u4ecd\u7136\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b\u9762\u4e34\u7684\u4e3b\u8981\u6311\u6218\u3002\u6a21\u578b\u5f80\u5f80\u8fc7\u4e8e\u81ea\u4fe1\uff0c\u5373\u4f7f\u5728\u5b83\u4eec\u7f3a\u4e4f\u77e5\u8bc6\u7684\u9886\u57df\u4e5f\u662f\u5982\u6b64\u3002\u5c3d\u7ba1\u5b58\u5728\u8fd9\u4e9b\u7f3a\u70b9\uff0c\u4f46\u5b83\u4eec\u7ecf\u5e38\u88ab\u7528\u4f5c\u77e5\u8bc6\u5e93\uff0c\u8fd9\u53ef\u80fd\u5bfc\u81f4\u5371\u9669\u7684\u7ed3\u679c\uff0c\u4f8b\u5982\u8bef\u4fe1\u606f\u7684\u4f20\u64ad\u3002\u867d\u7136\u6211\u4eec\u8ba4\u8bc6\u5230\u771f\u5b9e\u6027\u8d85\u8d8a\u4e86\u865a\u5e7b\uff0c\u4f46\u6211\u4eec\u5728\u8fd9\u91cc\u91c7\u53d6\u4e86\u4e00\u79cd\u4ee5\u865a\u5e7b\u4e3a\u5148\u7684\u65b9\u6cd5\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-3439\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217\" width=\"1039\" height=\"624\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621.png 1039w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621-300x180.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621-1024x615.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/11425ea2e7fd621-768x461.png 768w\" sizes=\"auto, (max-width: 1039px) 100vw, 1039px\" \/><\/p>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3450\" style=\"width: 2505px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3450\" class=\"wp-image-3450 size-full\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aa58bbc8041c830.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217\" width=\"2495\" height=\"1882\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aa58bbc8041c830.png 2495w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aa58bbc8041c830-300x226.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aa58bbc8041c830-1024x772.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aa58bbc8041c830-768x579.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aa58bbc8041c830-1536x1159.png 1536w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aa58bbc8041c830-2048x1545.png 2048w\" sizes=\"auto, (max-width: 2495px) 100vw, 2495px\" \/><p id=\"caption-attachment-3450\" class=\"wp-caption-text\">\u56fe 11 \u6587\u4ef6\u4e0a\u4f20\u5904\u7406\u3002\u793a\u4f8b\u5c55\u793a\u4e86 Llama 3 \u5bf9\u4e0a\u4f20\u6587\u4ef6\u8fdb\u884c\u5206\u6790\u548c\u53ef\u89c6\u5316\u7684\u8fc7\u7a0b\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u9075\u5faa\u8fd9\u6837\u7684\u539f\u5219\uff1a\u540e\u8bad\u7ec3\u5e94\u8be5\u4f7f\u6a21\u578b\u4e0e\u201c\u77e5\u9053\u5b83\u77e5\u9053\u4ec0\u4e48\u201d\u4fdd\u6301\u4e00\u81f4\uff0c\u800c\u4e0d\u662f\u6dfb\u52a0\u77e5\u8bc6\uff08Gekhman \u7b49\u4eba\uff0c2024\uff1bMielke \u7b49\u4eba\uff0c2020\uff09\u3002\u6211\u4eec\u7684\u4e3b\u8981\u65b9\u6cd5\u6d89\u53ca\u751f\u6210\u5c06\u6a21\u578b\u751f\u6210\u4e0e\u9884\u8bad\u7ec3\u6570\u636e\u4e2d\u5b58\u5728\u7684\u771f\u5b9e\u6570\u636e\u5b50\u96c6\u5bf9\u9f50\u7684\u6570\u636e\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u79cd\u5229\u7528 Llama 3 \u4e0a\u4e0b\u6587\u80fd\u529b\u7684\u77e5\u8bc6\u63a2\u6d4b\u6280\u672f\u3002\u8fd9\u4e2a\u6570\u636e\u751f\u6210\u8fc7\u7a0b\u5305\u62ec\u4ee5\u4e0b\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li>\u4ece\u9884\u8bad\u7ec3\u6570\u636e\u4e2d\u63d0\u53d6\u4e00\u4e2a\u6570\u636e\u7247\u6bb5\u3002<\/li>\n<li>\u901a\u8fc7\u63d0\u793a Llama 3 \u751f\u6210\u5173\u4e8e\u8fd9\u4e9b\u7247\u6bb5\uff08\u4e0a\u4e0b\u6587\uff09\u7684\u4e8b\u5b9e\u6027\u95ee\u9898\u3002<\/li>\n<li>\u4ece Llama 3 \u4e2d\u91c7\u6837\u5bf9\u8be5\u95ee\u9898\u7684\u56de\u7b54\u3002<\/li>\n<li>\u4f7f\u7528\u539f\u59cb\u4e0a\u4e0b\u6587\u4f5c\u4e3a\u53c2\u8003\uff0c\u5e76\u4f7f\u7528 Llama 3 \u4f5c\u4e3a\u8bc4\u5224\u8005\u6765\u8bc4\u5206\u751f\u6210\u7684\u6b63\u786e\u6027\u3002<\/li>\n<li>\u4f7f\u7528 Llama 3 \u4f5c\u4e3a\u8bc4\u5224\u8005\u6765\u8bc4\u5206\u751f\u6210\u7684\u4e30\u5bcc\u5ea6\u3002<\/li>\n<li>\u4e3a\u90a3\u4e9b\u5728\u591a\u6b21\u751f\u6210\u4e2d\u59cb\u7ec8\u4fe1\u606f\u4e30\u5bcc\u4e14\u4e0d\u6b63\u786e\u7684\u56de\u7b54\u751f\u6210\u62d2\u7edd\u7406\u7531\uff0c\u5e76\u4f7f\u7528 Llama 3<\/li>\n<\/ol>\n<p>\u6211\u4eec\u4f7f\u7528\u4ece\u77e5\u8bc6\u63a2\u6d4b\u4e2d\u751f\u6210\u7684\u6570\u636e\u6765\u9f13\u52b1\u6a21\u578b\u53ea\u56de\u7b54\u5b83\u77e5\u9053\u7684\u95ee\u9898\uff0c\u5e76\u62d2\u7edd\u56de\u7b54\u5b83\u4e0d\u786e\u5b9a\u7684\u95ee\u9898\u3002\u6b64\u5916\uff0c\u9884\u8bad\u7ec3\u6570\u636e\u5e76\u4e0d\u603b\u662f\u4e8b\u5b9e\u4e00\u81f4\u6216\u6b63\u786e\u7684\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u8fd8\u6536\u96c6\u4e86\u4e00\u7ec4\u6709\u9650\u7684\u6807\u8bb0\u771f\u5b9e\u6027\u6570\u636e\uff0c\u8fd9\u4e9b\u6570\u636e\u5904\u7406\u4e86\u654f\u611f\u8bdd\u9898\uff0c\u5176\u4e2d\u5b58\u5728\u8bb8\u591a\u4e0e\u4e8b\u5b9e\u76f8\u77db\u76fe\u6216\u4e0d\u6b63\u786e\u7684\u9648\u8ff0\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>4.3.7 \u53ef\u63a7\u6027<\/h4>\n<p>\u53ef\u63a7\u6027\u662f\u6307\u80fd\u591f\u5c06\u6a21\u578b\u7684\u884c\u4e3a\u548c\u7ed3\u679c\u5f15\u5bfc\u81f3\u6ee1\u8db3\u5f00\u53d1\u8005\u548c\u7528\u6237\u9700\u6c42\u7684\u80fd\u529b\u3002\u7531\u4e8eLlama 3\u662f\u4e00\u4e2a\u901a\u7528\u7684\u57fa\u7840\u6a21\u578b\uff0c\u56e0\u6b64\u5e94\u8be5\u80fd\u591f\u8f7b\u677e\u5730\u5c06\u5176\u5f15\u5bfc\u81f3\u4e0d\u540c\u7684\u4e0b\u6e38\u7528\u4f8b\u3002\u4e3a\u4e86\u63d0\u9ad8Llama 3\u7684\u53ef\u63a7\u6027\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u901a\u8fc7\u7cfb\u7edf\u63d0\u793a\uff08\u4f7f\u7528\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\uff09\u6765\u589e\u5f3a\u5176\u53ef\u63a7\u6027\uff0c\u7279\u522b\u662f\u5173\u4e8e\u54cd\u5e94\u957f\u5ea6\u3001\u683c\u5f0f\u3001\u8bed\u6c14\u548c\u89d2\u8272\/\u4eba\u7269\u8bbe\u5b9a\u65b9\u9762\u3002<\/p>\n<p><strong>\u6570\u636e\u6536\u96c6:<\/strong>\u00a0\u6211\u4eec\u901a\u8fc7\u8981\u6c42\u6807\u6ce8\u8005\u4e3aLlama 3\u8bbe\u8ba1\u4e0d\u540c\u7684\u7cfb\u7edf\u63d0\u793a\uff0c\u5728\u901a\u7528\u82f1\u8bed\u7c7b\u522b\u4e2d\u6536\u96c6\u53ef\u63a7\u6027\u504f\u597d\u6837\u672c\u3002\u7136\u540e\uff0c\u6807\u6ce8\u8005\u4e0e\u6a21\u578b\u8fdb\u884c\u5bf9\u8bdd\uff0c\u4ee5\u8bc4\u4f30\u6a21\u578b\u5728\u6574\u4e2a\u5bf9\u8bdd\u8fc7\u7a0b\u4e2d\u662f\u5426\u80fd\u591f\u59cb\u7ec8\u5982\u4e00\u5730\u9075\u5faa\u7cfb\u7edf\u63d0\u793a\u4e2d\u5b9a\u4e49\u7684\u6307\u4ee4\u3002\u4ee5\u4e0b\u662f\u7528\u4e8e\u589e\u5f3a\u53ef\u63a7\u6027\u7684\u5b9a\u5236\u7cfb\u7edf\u63d0\u793a\u793a\u4f8b\uff1a<\/p>\n<blockquote><p>\u201c\u4f60\u662f\u4e00\u4e2a\u4e50\u4e8e\u52a9\u4eba\u4e14\u5145\u6ee1\u6d3b\u529b\u7684 AI \u804a\u5929\u673a\u5668\u4eba\uff0c\u62c5\u4efb\u5fd9\u788c\u5bb6\u5ead\u7684\u81b3\u98df\u8ba1\u5212\u52a9\u624b\u3002\u5de5\u4f5c\u65e5\u9910\u98df\u5e94\u8be5\u5feb\u901f\u4fbf\u6377\u3002\u65e9\u9910\u548c\u5348\u9910\u5e94\u4f18\u5148\u9009\u62e9\u65b9\u4fbf\u98df\u54c1\uff0c\u4f8b\u5982\u8c37\u7269\u3001\u82f1\u5f0f\u677e\u997c\u914d\u9884\u5148\u716e\u719f\u7684\u57f9\u6839\u548c\u5176\u4ed6\u5feb\u901f\u6613\u505a\u7684\u98df\u7269\u3002\u8fd9\u4e2a\u5bb6\u5ead\u5f88\u5fd9\u3002\u8bf7\u52a1\u5fc5\u8be2\u95ee\u4ed6\u4eec\u624b\u8fb9\u662f\u5426\u6709\u5fc5\u9700\u54c1\u548c\u559c\u6b22\u7684\u996e\u54c1\uff0c\u4f8b\u5982\u5496\u5561\u6216\u80fd\u91cf\u996e\u6599\uff0c\u4ee5\u514d\u4ed6\u4eec\u5fd8\u8bb0\u8d2d\u4e70\u3002\u9664\u975e\u662f\u7279\u6b8a\u573a\u5408\uff0c\u5426\u5219\u8bf7\u8bb0\u4f4f\u8282\u7ea6\u9884\u7b97\u3002\u201d<\/p><\/blockquote>\n<p><strong>\u5efa\u6a21:<\/strong>\u00a0\u6536\u96c6\u504f\u597d\u6570\u636e\u540e\uff0c\u6211\u4eec\u5c06\u8fd9\u4e9b\u6570\u636e\u7528\u4e8e\u5956\u52b1\u5efa\u6a21\u3001\u62d2\u7edd\u91c7\u6837\u3001SFT\uff08\u6301\u7eed\u7cbe\u7ec6\u8c03\u6574\uff09\u548cDPO\uff08\u6570\u636e\u9a71\u52a8\u7684\u53c2\u6570\u4f18\u5316\uff09\uff0c\u4ee5\u589e\u5f3aLlama 3\u7684\u53ef\u63a7\u6027\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>5 \u7ed3\u679c<\/h2>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u5bf9 Llama 3 \u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u5e7f\u6cdb\u7684\u8bc4\u4f30\uff0c\u8c03\u67e5\u4e86\u4ee5\u4e0b\u65b9\u9762\u7684\u6027\u80fd\uff1a(1) \u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c(2) \u540e\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c\u4ee5\u53ca (3) Llama 3 \u7684\u5b89\u5168\u7279\u6027\u3002\u6211\u4eec\u5728\u4e0b\u9762\u7684\u72ec\u7acb\u5b50\u8282\u4e2d\u5c55\u793a\u8fd9\u4e9b\u8bc4\u4f30\u7684\u7ed3\u679c\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>5.1 \u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b<\/h3>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u62a5\u544a\u9884\u8bad\u7ec3 Llama 3 (\u7b2c\u4e09\u90e8\u5206) \u7684\u8bc4\u4f30\u7ed3\u679c\uff0c\u5e76\u5c06\u5176\u4e0e\u5176\u4ed6\u53ef\u6bd4\u89c4\u6a21\u7684\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\u3002\u6211\u4eec\u4f1a\u5c3d\u53ef\u80fd\u91cd\u73b0\u7ade\u54c1\u6a21\u578b\u7684\u7ed3\u679c\u3002\u5bf9\u4e8e\u975e Llama \u6a21\u578b\uff0c\u6211\u4eec\u5c06\u62a5\u544a\u5728\u516c\u5f00\u62a5\u9053\u7684\u7ed3\u679c\u4e2d\u6216\uff08\u5728\u53ef\u80fd\u7684\u60c5\u51b5\u4e0b\uff09\u6211\u4eec\u81ea\u5df1\u91cd\u73b0\u7684\u7ed3\u679c\u4e2d\u7684\u6700\u4f73\u5206\u6570\u3002\u8fd9\u4e9b\u8bc4\u4f30\u7684\u5177\u4f53\u7ec6\u8282\uff0c\u5305\u62ec\u5c04\u6570\u3001\u6307\u6807\u548c\u5176\u4ed6\u76f8\u5173\u7684\u8d85\u53c2\u6570\u548c\u8bbe\u7f6e\u7b49\u914d\u7f6e\uff0c\u53ef\u4ee5\u5728\u6211\u4eec\u7684 Github \u4ed3\u5e93\u4e2d\u83b7\u5f97\uff1a [\u5728\u6b64\u63d2\u5165\u94fe\u63a5]\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u4f1a\u53d1\u5e03\u4f5c\u4e3a\u516c\u5f00\u57fa\u51c6\u8bc4\u4f30\u7684\u4e00\u90e8\u5206\u751f\u6210\u7684\u6570\u636e\uff0c\u8fd9\u4e9b\u6570\u636e\u53ef\u4ee5\u5728\u8fd9\u91cc\u627e\u5230\uff1a [\u5728\u6b64\u63d2\u5165\u94fe\u63a5]\u3002<\/p>\n<p>\u6211\u4eec\u5c06\u6839\u636e\u6807\u51c6\u57fa\u51c6 (\u7b2c\u4e94\u90e8\u5206 5.1.1) \u5bf9\u6a21\u578b\u8d28\u91cf\u8fdb\u884c\u8bc4\u4f30\uff0c\u6d4b\u8bd5\u591a\u9879\u9009\u62e9\u9898\u8bbe\u7f6e\u53d8\u5316\u7684\u9c81\u68d2\u6027 (\u7b2c\u4e94\u90e8\u5206 5.1.2)\uff0c\u4ee5\u53ca\u8fdb\u884c\u5bf9\u6297\u6027\u8bc4\u4f30 (\u7b2c\u4e94\u90e8\u5206 5.1.3)\u3002\u6211\u4eec\u8fd8\u5c06\u8fdb\u884c\u6c61\u67d3\u5206\u6790\uff0c\u4ee5\u4f30\u8ba1\u8bad\u7ec3\u6570\u636e\u6c61\u67d3\u5bf9\u6211\u4eec\u8bc4\u4f30\u7684\u5f71\u54cd\u7a0b\u5ea6 (\u7b2c\u4e94\u90e8\u5206 5.1.4)\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>5.1.1 \u6807\u51c6\u57fa\u51c6<\/h4>\n<p>\u4e3a\u4e86\u5c06\u6211\u4eec\u7684\u6a21\u578b\u4e0e\u5f53\u524d\u6700\u5148\u8fdb\u7684\u6c34\u5e73\u8fdb\u884c\u6bd4\u8f83\uff0c\u6211\u4eec\u5bf9 Llama 3 \u5728\u5927\u91cf\u6807\u51c6\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u8fd9\u4e9b\u6d4b\u8bd5\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n<p>(1) \u5e38\u8bc6\u63a8\u7406\uff1b(2) \u77e5\u8bc6\uff1b(3) \u9605\u8bfb\u7406\u89e3\uff1b(4) \u6570\u5b66\u3001\u63a8\u7406\u548c\u95ee\u9898\u89e3\u51b3\uff1b(5) \u957f\u4e0a\u4e0b\u6587\uff1b(6) \u4ee3\u7801\uff1b(7) \u5bf9\u6297\u6027\u8bc4\u4f30\uff1b\u4ee5\u53ca (8) \u603b\u4f53\u8bc4\u4f30\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3451\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7395d18ec55ad44.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"796\" height=\"456\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7395d18ec55ad44.png 1046w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7395d18ec55ad44-300x172.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7395d18ec55ad44-1024x586.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7395d18ec55ad44-768x440.png 768w\" sizes=\"auto, (max-width: 796px) 100vw, 796px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u5b9e\u9a8c\u8bbe\u7f6e\u3002<\/strong>\u5bf9\u4e8e\u6bcf\u4e2a\u57fa\u51c6\u6d4b\u8bd5\uff0c\u6211\u4eec\u8ba1\u7b97 Llama 3 \u7684\u5206\u6570\u4ee5\u53ca\u5176\u4ed6\u5177\u6709\u53ef\u6bd4\u5c3a\u5bf8\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u5206\u6570\u3002\u5728\u53ef\u80fd\u7684\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u81ea\u5df1\u7684\u7ba1\u9053\u91cd\u65b0\u8ba1\u7b97\u5176\u4ed6\u6a21\u578b\u7684\u6570\u636e\u3002\u4e3a\u4e86\u786e\u4fdd\u516c\u5e73\u6bd4\u8f83\uff0c\u6211\u4eec\u7136\u540e\u9009\u62e9\u6211\u4eec\u5728\u8ba1\u7b97\u7684\u6570\u636e\u548c\u8be5\u6a21\u578b\u62a5\u544a\u7684\u6570\u5b57\uff08\u4f7f\u7528\u76f8\u540c\u6216\u66f4\u4fdd\u5b88\u7684\u8bbe\u7f6e\uff09\u4e4b\u95f4\u7684\u6700\u4f73\u5206\u6570\u3002\u60a8\u53ef\u4ee5\u5728\u6b64\u5904\u627e\u5230\u6709\u5173\u6211\u4eec\u7684\u8bc4\u4f30\u8bbe\u7f6e\u7684\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\u3002\u5bf9\u4e8e\u67d0\u4e9b\u6a21\u578b\uff0c\u4f8b\u5982\u7531\u4e8e\u672a\u53d1\u5e03\u9884\u8bad\u7ec3\u6a21\u578b\u6216API\u4e0d\u63d0\u4f9b\u5bf9\u6570\u6982\u7387\u8bbf\u95ee\u6743\u9650\uff0c\u65e0\u6cd5\u91cd\u65b0\u8ba1\u7b97\u57fa\u51c6\u6d4b\u8bd5\u503c\u3002\u8fd9\u5c24\u5176\u9002\u7528\u4e8e\u4e0e Llama 3 405B \u53ef\u6bd4\u7684\u6240\u6709\u6a21\u578b\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u6ca1\u6709\u62a5\u544a Llama 3 405B \u7684\u7c7b\u522b\u5e73\u5747\u503c\uff0c\u56e0\u4e3a\u9700\u8981\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u7684\u6570\u5b57\u90fd\u53ef\u7528\u3002<\/p>\n<p><strong>\u663e\u8457\u6027\u503c\u3002<\/strong>\u5728\u8ba1\u7b97\u57fa\u51c6\u6d4b\u8bd5\u5206\u6570\u65f6\uff0c\u7531\u4e8e\u6709\u51e0\u79cd\u65b9\u5dee\u6765\u6e90\u4f1a\u5bfc\u81f4\u5bf9\u57fa\u51c6\u6d4b\u8bd5\u65e8\u5728\u6d4b\u91cf\u7684\u6a21\u578b\u6027\u80fd\u4ea7\u751f\u4e0d\u7cbe\u786e\u7684\u4f30\u8ba1\uff0c\u4f8b\u5982\u5c11\u91cf\u6f14\u793a\u3001\u968f\u673a\u79cd\u5b50\u548c\u6279\u91cf\u5927\u5c0f\u3002\u8fd9\u4f7f\u5f97\u7406\u89e3\u4e00\u4e2a\u6a21\u578b\u662f\u5426\u5728\u7edf\u8ba1\u4e0a\u663e\u8457\u4f18\u4e8e\u53e6\u4e00\u4e2a\u6a21\u578b\u53d8\u5f97\u5177\u6709\u6311\u6218\u6027\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5c06\u5206\u6570\u4e0e 95% \u7f6e\u4fe1\u533a\u95f4 (CI) \u4e00\u8d77\u62a5\u544a\uff0c\u4ee5\u53cd\u6620\u57fa\u51c6\u6570\u636e\u9009\u62e9\u5e26\u6765\u7684\u65b9\u5dee\u3002\u6211\u4eec\u4f7f\u7528\u516c\u5f0f\uff08Madaan \u7b49\u4eba\uff0c2024b\uff09\u5206\u6790\u5730\u8ba1\u7b97 95% CI\uff1a<\/p>\n<div id=\"attachment_3452\" style=\"width: 380px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3452\" class=\"wp-image-3452\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/c2592f0e3da6f61.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"370\" height=\"78\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/c2592f0e3da6f61.png 455w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/c2592f0e3da6f61-300x63.png 300w\" sizes=\"auto, (max-width: 370px) 100vw, 370px\" \/><p id=\"caption-attachment-3452\" class=\"wp-caption-text\">CI_analytic(S) = 1.96 * sqrt(S * (1 &#8211; S) \/ N)<\/p><\/div>\n<p>\u5176\u4e2d S \u662f\u9996\u9009\u7684\u57fa\u51c6\u5206\u6570\uff0cN \u662f\u57fa\u51c6\u7684\u6837\u672c\u91cf\u3002\u6211\u4eec\u6ce8\u610f\u5230\uff0c\u7531\u4e8e\u57fa\u51c6\u6570\u636e\u4e2d\u7684\u65b9\u5dee\u4e0d\u662f\u552f\u4e00\u7684\u65b9\u5dee\u6765\u6e90\uff0c\u56e0\u6b64\u8fd9\u4e9b 95% CI \u662f\u5b9e\u9645\u80fd\u529b\u4f30\u8ba1\u65b9\u5dee\u7684\u4e0b\u9650\u3002\u5bf9\u4e8e\u4e0d\u662f\u7b80\u5355\u5e73\u5747\u503c\u7684\u6307\u6807\uff0c\u5c06\u7701\u7565 CI\u3002<\/p>\n<p><strong>Llama 3 8B \u548c 70B \u6a21\u578b\u7684\u7ed3\u679c\u3002<\/strong>\u56fe 12 \u663e\u793a\u4e86 Llama 3 8B \u548c 70B \u5728\u5e38\u8bc6\u63a8\u7406\u3001\u77e5\u8bc6\u3001\u9605\u8bfb\u7406\u89e3\u3001\u6570\u5b66\u548c\u63a8\u7406\u4ee5\u53ca\u4ee3\u7801\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u5e73\u5747\u6027\u80fd\u3002\u7ed3\u679c\u8868\u660e\uff0cLlama 3 8B \u5728\u51e0\u4e4e\u6240\u6709\u7c7b\u522b\u4e2d\u90fd\u4f18\u4e8e\u7ade\u4e89\u6a21\u578b\uff0c\u65e0\u8bba\u662f\u6309\u7c7b\u522b\u80dc\u7387\u8fd8\u662f\u6309\u7c7b\u522b\u5e73\u5747\u6027\u80fd\u800c\u8a00\u3002\u6211\u4eec\u8fd8\u53d1\u73b0\uff0c Llama 3 70B \u5728\u5927\u591a\u6570\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u90fd\u6bd4\u5176\u524d\u8eab Llama 2 70B \u5927\u5e45\u63d0\u9ad8\u4e86\u6027\u80fd\uff0c\u9664\u4e86\u53ef\u80fd\u5df2\u9971\u548c\u7684\u5e38\u8bc6\u57fa\u51c6\u6d4b\u8bd5\u4e4b\u5916\u3002Llama 3 70B \u4e5f\u4f18\u4e8e Mixtral 8x22B\u3002<\/p>\n<p><strong>8B\u548c70B\u6a21\u578b\u7684\u7ed3\u679c\u3002<\/strong>\u56fe12\u663e\u793a\u4e86Llama 3 8B\u548c70B\u5728\u5e38\u8bc6\u63a8\u7406\u3001\u77e5\u8bc6\u3001\u9605\u8bfb\u7406\u89e3\u3001\u6570\u5b66\u4e0e\u63a8\u7406\u4ee5\u53ca\u4ee3\u7801\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u5e73\u5747\u8868\u73b0\u3002\u7ed3\u679c\u8868\u660e\uff0cLlama 3 8B\u5728\u51e0\u4e4e\u6bcf\u4e2a\u7c7b\u522b\u4e2d\u90fd\u8d85\u8d8a\u4e86\u7ade\u4e89\u6a21\u578b\uff0c\u65e0\u8bba\u662f\u6309\u7c7b\u522b\u7684\u80dc\u7387\u8fd8\u662f\u5e73\u5747\u6bcf\u7c7b\u522b\u7684\u8868\u73b0\u3002\u6211\u4eec\u8fd8\u53d1\u73b0\uff0cLlama 3 70B\u5728\u5927\u591a\u6570\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u6bd4\u5176\u524d\u4ee3\u4ea7\u54c1Llama 2 70B\u6709\u4e86\u5f88\u5927\u7684\u63d0\u5347\uff0c\u9664\u4e86\u53ef\u80fd\u5df2\u8fbe\u5230\u9971\u548c\u7684\u5e38\u8bc6\u57fa\u51c6\u6d4b\u8bd5\u4e4b\u5916\u3002Llama 3 70B\u8fd8\u8d85\u8d8a\u4e86Mixtral 8x22B\u3002<\/p>\n<p><strong>\u6240\u6709\u6a21\u578b\u7684\u8be6\u7ec6\u7ed3\u679c\u3002<\/strong>\u88689\u300110\u300111\u300112\u300113\u548c14\u5c55\u793a\u4e86\u9884\u8bad\u7ec3\u7684Llama 3 8B\u300170B\u548c405B\u6a21\u578b\u5728\u9605\u8bfb\u7406\u89e3\u4efb\u52a1\u3001\u7f16\u7801\u4efb\u52a1\u3001\u5e38\u8bc6\u7406\u89e3\u4efb\u52a1\u3001\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u548c\u5e38\u89c4\u4efb\u52a1\u4e2d\u7684\u57fa\u51c6\u6d4b\u8bd5\u8868\u73b0\u3002\u8fd9\u4e9b\u8868\u683c\u5c06Llama 3\u7684\u8868\u73b0\u4e0e\u540c\u7c7b\u5927\u5c0f\u7684\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u7ed3\u679c\u663e\u793a\uff0cLlama 3 405B\u5728\u5176\u7c7b\u522b\u4e2d\u5177\u6709\u7ade\u4e89\u529b\uff0c\u5c24\u5176\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u8d85\u8d8a\u4e86\u5148\u524d\u7684\u5f00\u6e90\u6a21\u578b\u3002\u5173\u4e8e\u957f\u4e0a\u4e0b\u6587\u7684\u6d4b\u8bd5\uff0c\u6211\u4eec\u5728\u7b2c5.2\u8282\u4e2d\u63d0\u4f9b\u4e86\u66f4\u5168\u9762\u7684\u7ed3\u679c\uff08\u5305\u62ec\u9488\u5728\u5927\u6d77\u635e\u9488\u7b49\u63a2\u6d4b\u4efb\u52a1\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3453\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/6701885198f681f.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"889\" height=\"440\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/6701885198f681f.png 1262w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/6701885198f681f-300x149.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/6701885198f681f-1024x507.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/6701885198f681f-768x380.png 768w\" sizes=\"auto, (max-width: 889px) 100vw, 889px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3454\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/adbe121faa1be04.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"862\" height=\"363\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/adbe121faa1be04.png 1260w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/adbe121faa1be04-300x126.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/adbe121faa1be04-1024x432.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/adbe121faa1be04-768x324.png 768w\" sizes=\"auto, (max-width: 862px) 100vw, 862px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h4>5.1.2 \u6a21\u578b\u7a33\u5065\u6027\uff08Model Robustness\uff09<\/h4>\n<p>\u9664\u4e86\u57fa\u51c6\u6d4b\u8bd5\u6027\u80fd\u4e4b\u5916\uff0c\u7a33\u5065\u6027\u4e5f\u662f\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u8d28\u91cf\u7684\u91cd\u8981\u56e0\u7d20\u3002\u6211\u4eec\u7814\u7a76\u4e86\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5728\u591a\u9879\u9009\u62e9\u9898 (MCQ) \u8bbe\u7f6e\u4e2d\u7684\u8bbe\u8ba1\u9009\u62e9\u7a33\u5065\u6027\u3002\u5148\u524d\u7684\u7814\u7a76\u8868\u660e\uff0c\u6a21\u578b\u6027\u80fd\u53ef\u80fd\u5bf9\u8fd9\u4e9b\u8bbe\u7f6e\u4e2d\u770b\u4f3c\u4efb\u610f\u7684\u8bbe\u8ba1\u9009\u62e9\u5f88\u654f\u611f\uff0c\u4f8b\u5982\uff0c\u6a21\u578b\u5f97\u5206\u751a\u81f3\u6392\u540d\u53ef\u80fd\u4f1a\u968f\u7740\u4e0a\u4e0b\u6587\u793a\u4f8b\u7684\u987a\u5e8f\u548c\u6807\u7b7e\u800c\u6539\u53d8\uff08Lu \u7b49\u4eba\uff0c2022\uff1bZhao \u7b49\u4eba\uff0c2021\uff1bRobinson \u548c Wingate\uff0c2023\uff1bLiang \u7b49\u4eba\uff0c2022\uff1bGupta \u7b49\u4eba\uff0c2024\uff09\uff0c\u63d0\u793a\u7684\u786e\u5207\u683c\u5f0f\uff08Weber \u7b49\u4eba\uff0c2023b\uff1bMishra \u7b49\u4eba\uff0c2022\uff09 \u6216\u7b54\u6848\u9009\u9879\u7684\u683c\u5f0f\u548c\u987a\u5e8f\uff08Alzahrani \u7b49\u4eba\uff0c2024\uff1bWang \u7b49\u4eba\uff0c2024a\uff1bZheng \u7b49\u4eba\uff0c2023\uff09\u3002\u53d7\u6b64\u5de5\u4f5c\u542f\u53d1\uff0c\u6211\u4eec\u4f7f\u7528 MMLU \u57fa\u51c6\u6765\u8bc4\u4f30\u9884\u8bad\u7ec3\u6a21\u578b\u5bf9\u4ee5\u4e0b\u65b9\u9762\u7684\u7a33\u5065\u6027\uff1a\uff081\uff09\u5c11\u6570\u955c\u5934\u6807\u7b7e\u504f\u5dee\uff0c\uff082\uff09\u6807\u7b7e\u53d8\u4f53\uff0c\uff083\uff09\u7b54\u6848\u987a\u5e8f\uff0c\u4ee5\u53ca\uff084\uff09\u63d0\u793a\u683c\u5f0f\uff1a<\/p>\n<ul>\n<li><strong>\u5c11\u6570\u955c\u5934\u6807\u7b7e\u504f\u5dee\u3002<\/strong>\u00a0\u9075\u5faa Zheng \u7b49\u4eba\uff082023\uff09\uff0c&#8230;\uff08\u6b64\u5904\u7701\u7565\u5b9e\u9a8c\u7ec6\u8282\u548c\u7ed3\u679c\u63cf\u8ff0\uff09\u3002<\/li>\n<li><strong>\u6807\u7b7e\u53d8\u4f53\u3002<\/strong>\u00a0\u6211\u4eec\u8fd8\u7814\u7a76\u4e86\u6a21\u578b\u5bf9\u4e0d\u540c\u9009\u62e9\u6807\u8bb0\u96c6\u7684\u54cd\u5e94\u3002\u6211\u4eec\u8003\u8651\u4e86 Alzahrani \u7b49\u4eba\uff082024\uff09\u63d0\u51fa\u7684\u4e24\u4e2a\u6807\u8bb0\u96c6\uff1a\u5373\u4e00\u7ec4\u5e38\u89c1\u8bed\u8a00\u65e0\u5173\u6807\u8bb0\uff08$ &amp; # @\uff09\u548c\u4e00\u7ec4\u4e0d\u5177\u6709\u4efb\u4f55\u9690\u542b\u76f8\u5bf9\u987a\u5e8f\u7684\u7a00\u6709\u6807\u8bb0\uff08oe \u00a7 \u0437 \u00fc\uff09\u3002\u6211\u4eec\u8fd8\u8003\u8651\u4e86\u89c4\u8303\u6807\u7b7e\u7684\u4e24\u79cd\u7248\u672c\uff08A. B. C. D. \u548c A) B) C) D\uff09\uff09 \u548c\u4e00\u4e2a\u6570\u5b57\u5217\u8868\uff081. 2. 3. 4.\uff09\u3002<\/li>\n<li><strong>\u7b54\u6848\u987a\u5e8f\u3002<\/strong>\u00a0\u9075\u5faa Wang \u7b49\u4eba\uff082024a\uff09\uff0c\u6211\u4eec\u8ba1\u7b97\u7ed3\u679c\u5728\u4e0d\u540c\u7b54\u6848\u987a\u5e8f\u4e0b\u7684\u7a33\u5b9a\u6027\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u5c06\u6570\u636e\u96c6\u4e2d\u7684\u6240\u6709\u7b54\u6848\u6839\u636e\u56fa\u5b9a\u7684\u6392\u5217\u91cd\u65b0\u6620\u5c04\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u6392\u5217 A B C D\uff0c\u6240\u6709\u6807\u8bb0\u4e3a A \u548c B \u7684\u7b54\u6848\u9009\u9879\u4fdd\u6301\u5176\u6807\u7b7e\uff0c\u800c\u6807\u8bb0\u4e3a C \u7684\u6240\u6709\u7b54\u6848\u9009\u9879\u83b7\u53d6\u6807\u7b7e D\uff0c\u53cd\u4e4b\u4ea6\u7136\u3002<\/li>\n<li><strong>\u63d0\u793a\u683c\u5f0f\u3002<\/strong>\u00a0\u6211\u4eec\u8bc4\u4f30\u4e86\u4e94\u79cd\u4efb\u52a1\u63d0\u793a\u7684\u6027\u80fd\u5dee\u5f02\uff0c\u8fd9\u4e9b\u63d0\u793a\u7684\u4fe1\u606f\u91cf\u4e0d\u540c\uff1a\u4e00\u79cd\u63d0\u793a\u7b80\u5355\u5730\u8981\u6c42\u6a21\u578b\u56de\u7b54\u95ee\u9898\uff0c\u800c\u5176\u4ed6\u63d0\u793a\u5219\u65ad\u8a00\u6a21\u578b\u7684\u4e13\u4e1a\u77e5\u8bc6\u6216\u5e94\u8be5\u9009\u62e9\u6700\u4f73\u7b54\u6848\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3455\" style=\"width: 824px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3455\" class=\"wp-image-3455\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/809e9891cb99842.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"814\" height=\"297\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/809e9891cb99842.png 1260w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/809e9891cb99842-300x110.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/809e9891cb99842-1024x374.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/809e9891cb99842-768x280.png 768w\" sizes=\"auto, (max-width: 814px) 100vw, 814px\" \/><p id=\"caption-attachment-3455\" class=\"wp-caption-text\">\u8868 11 \u9884\u8bad\u7ec3\u6a21\u578b\u5728\u5e38\u8bc6\u7406\u89e3\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u3002\u7ed3\u679c\u5305\u62ec 95% \u7f6e\u4fe1\u533a\u95f4\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3456\" style=\"width: 767px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3456\" class=\"wp-image-3456\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/374f644ec20a2dd.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"757\" height=\"317\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/374f644ec20a2dd.png 1264w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/374f644ec20a2dd-300x126.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/374f644ec20a2dd-1024x429.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/374f644ec20a2dd-768x322.png 768w\" sizes=\"auto, (max-width: 757px) 100vw, 757px\" \/><p id=\"caption-attachment-3456\" class=\"wp-caption-text\">\u8868 12 \u9884\u8bad\u7ec3\u6a21\u578b\u5728\u6570\u5b66\u548c\u63a8\u7406\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u3002\u7ed3\u679c\u5305\u62ec 95% \u7f6e\u4fe1\u533a\u95f4\u300211 \u6b21\u5c04\u51fb\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3457\" style=\"width: 778px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3457\" class=\"wp-image-3457\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc2957d2a5e5dd9.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"768\" height=\"338\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc2957d2a5e5dd9.png 1134w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc2957d2a5e5dd9-300x132.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc2957d2a5e5dd9-1024x451.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc2957d2a5e5dd9-768x339.png 768w\" sizes=\"auto, (max-width: 768px) 100vw, 768px\" \/><p id=\"caption-attachment-3457\" class=\"wp-caption-text\">\u8868 13 \u9884\u8bad\u7ec3\u6a21\u578b\u5728\u901a\u7528\u8bed\u8a00\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u3002\u7ed3\u679c\u5305\u62ec 95% \u7f6e\u4fe1\u533a\u95f4\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3458\" style=\"width: 827px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3458\" class=\"wp-image-3458\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d2b10349107359d.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"817\" height=\"314\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d2b10349107359d.png 1259w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d2b10349107359d-300x115.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d2b10349107359d-1024x394.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/d2b10349107359d-768x295.png 768w\" sizes=\"auto, (max-width: 817px) 100vw, 817px\" \/><p id=\"caption-attachment-3458\" class=\"wp-caption-text\">\u56fe 13 \u6211\u4eec\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5728 MMLU \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5bf9\u4e0d\u540c\u8bbe\u8ba1\u9009\u62e9\u7684\u9c81\u68d2\u6027\u3002\u5de6\u4fa7\uff1a\u4e0d\u540c\u6807\u7b7e\u53d8\u4f53\u7684\u6027\u80fd\u3002\u53f3\u4fa7\uff1a\u5c11\u6837\u672c\u793a\u4f8b\u4e2d\u5b58\u5728\u4e0d\u540c\u6807\u7b7e\u7684\u6027\u80fd\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3459\" style=\"width: 790px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3459\" class=\"wp-image-3459\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7dbea9fca2abaa2.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"780\" height=\"390\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7dbea9fca2abaa2.png 1254w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7dbea9fca2abaa2-300x150.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7dbea9fca2abaa2-1024x512.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7dbea9fca2abaa2-768x384.png 768w\" sizes=\"auto, (max-width: 780px) 100vw, 780px\" \/><p id=\"caption-attachment-3459\" class=\"wp-caption-text\">\u56fe 14 \u6211\u4eec\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5728 MMLU \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5bf9\u4e0d\u540c\u8bbe\u8ba1\u9009\u62e9\u7684\u9c81\u68d2\u6027\u3002\u5de6\u4fa7\uff1a\u4e0d\u540c\u7b54\u6848\u987a\u5e8f\u7684\u6027\u80fd\u3002\u53f3\u4fa7\uff1a\u4e0d\u540c\u63d0\u793a\u683c\u5f0f\u7684\u6027\u80fd\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<p>\u56fe 13 \u5c55\u793a\u4e86\u6211\u4eec\u7814\u7a76\u6807\u7b7e\u53d8\u4f53\uff08\u5de6\uff09\u548c\u5c11\u6570\u955c\u5934\u6807\u7b7e\u504f\u5dee\uff08\u53f3\uff09\u6a21\u578b\u6027\u80fd\u7a33\u5065\u6027\u7684\u5b9e\u9a8c\u7ed3\u679c\u3002\u7ed3\u679c\u8868\u660e\uff0c\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5bf9 MCQ \u6807\u7b7e\u7684\u53d8\u5316\u4ee5\u53ca\u5c11\u6570\u955c\u5934\u63d0\u793a\u6807\u7b7e\u7684\u7ed3\u6784\u90fd\u975e\u5e38\u7a33\u5065\u3002\u8fd9\u79cd\u7a33\u5065\u6027\u5bf9\u4e8e 405B \u53c2\u6570\u6a21\u578b\u5c24\u5176\u660e\u663e\u3002<\/p>\n<p>\u56fe 14 \u5c55\u793a\u4e86\u6211\u4eec\u5bf9\u7b54\u6848\u987a\u5e8f\u548c\u63d0\u793a\u683c\u5f0f\u7a33\u5065\u6027\u7814\u7a76\u7684\u7ed3\u679c\u3002\u8fd9\u4e9b\u7ed3\u679c\u8fdb\u4e00\u6b65\u5f3a\u8c03\u4e86\u6211\u4eec\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u6027\u80fd\u7684\u7a33\u5065\u6027\uff0c\u7279\u522b\u662f Llama 3 405B \u7684\u7a33\u5065\u6027\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3460\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/74236358dbcf641.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"802\" height=\"393\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/74236358dbcf641.png 1259w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/74236358dbcf641-300x147.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/74236358dbcf641-1024x501.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/74236358dbcf641-768x376.png 768w\" sizes=\"auto, (max-width: 802px) 100vw, 802px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h4>5.1.3 \u5bf9\u6297\u6027\u57fa\u51c6\u6d4b\u8bd5<\/h4>\n<p>\u9664\u4e86\u4e0a\u9762\u63d0\u5230\u7684\u57fa\u51c6\u6d4b\u8bd5\u5916\uff0c\u6211\u4eec\u8fd8\u5bf9\u4e09\u4e2a\u9886\u57df\u7684\u51e0\u4e2a\u5bf9\u6297\u6027\u57fa\u51c6\u8fdb\u884c\u4e86\u8bc4\u4f30\uff1a\u95ee\u7b54\u3001\u6570\u5b66\u63a8\u7406\u548c\u53e5\u5f0f\u6539\u5199\u68c0\u6d4b\u3002\u8fd9\u4e9b\u6d4b\u8bd5\u65e8\u5728\u63a2\u6d4b\u6a21\u578b\u5728\u4e13\u95e8\u8bbe\u8ba1\u6210\u5177\u6709\u6311\u6218\u6027\u7684\u4efb\u52a1\u4e0a\u7684\u80fd\u529b\uff0c\u5e76\u53ef\u80fd\u6307\u51fa\u6a21\u578b\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u8fc7\u5ea6\u62df\u5408\u95ee\u9898\u3002<\/p>\n<ul>\n<li><strong>\u95ee\u7b54\u65b9\u9762<\/strong>\uff0c\u6211\u4eec\u4f7f\u7528\u4e86Adversarial SQuAD (Jia \u548c Liang, 2017) \u548c Dynabench SQuAD (Kiela \u7b49\u4eba, 2021)\u3002<\/li>\n<li><strong>\u6570\u5b66\u63a8\u7406\u65b9\u9762<\/strong>\uff0c\u6211\u4eec\u4f7f\u7528\u4e86 GSM-Plus (Li \u7b49\u4eba, 2024c)\u3002<\/li>\n<li><strong>\u53e5\u5f0f\u6539\u5199\u68c0\u6d4b\u65b9\u9762<\/strong>\uff0c\u6211\u4eec\u4f7f\u7528\u4e86 PAWS (Zhang \u7b49\u4eba, 2019)\u3002<\/li>\n<\/ul>\n<p>\u56fe 15 \u5c55\u793a\u4e86 Llama 3 8B\u300170B \u548c 405B \u5728\u5bf9\u6297\u6027\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u5f97\u5206\uff0c\u4f5c\u4e3a\u5176\u5728\u975e\u5bf9\u6297\u6027\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u6027\u80fd\u7684\u51fd\u6570\u3002\u6211\u4eec\u4f7f\u7528\u7684\u975e\u5bf9\u6297\u6027\u57fa\u51c6\u6d4b\u8bd5\u662f\uff1a\u7528\u4e8e\u95ee\u7b54\u7684 SQuAD (Rajpurkar \u7b49\u4eba, 2016)\u3001\u7528\u4e8e\u6570\u5b66\u63a8\u7406\u7684 GSM8K \u548c\u7528\u4e8e\u53e5\u5f0f\u6539\u5199\u68c0\u6d4b\u7684 QQP (Wang \u7b49\u4eba, 2017)\u3002\u6bcf\u4e2a\u6570\u636e\u70b9\u4ee3\u8868\u4e00\u4e2a\u5bf9\u6297\u6027\u6570\u636e\u96c6\u548c\u975e\u5bf9\u6297\u6027\u6570\u636e\u96c6\u5bf9\uff08\u4f8b\u5982 QQP \u4e0e PAWS \u914d\u5bf9\uff09\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u7c7b\u522b\u5185\u7684\u6240\u6709\u53ef\u80fd\u914d\u5bf9\u3002\u5bf9\u89d2\u7ebf\u4e0a\u7684\u9ed1\u8272\u7ebf\u8868\u793a\u5bf9\u6297\u6027\u548c\u975e\u5bf9\u6297\u6027\u6570\u636e\u96c6\u4e4b\u95f4\u7684\u5e73\u4ef7\u2014\u2014\u5728\u8be5\u7ebf\u4e0a\u8868\u793a\u6a21\u578b\u65e0\u8bba\u5bf9\u6297\u6027\u4e0e\u5426\u90fd\u5177\u6709\u7c7b\u4f3c\u7684\u6027\u80fd\u3002<\/p>\n<p>\u5728\u53e5\u5f0f\u6539\u5199\u68c0\u6d4b\u65b9\u9762\uff0c\u65e0\u8bba\u662f\u9884\u8bad\u7ec3\u6a21\u578b\u8fd8\u662f\u540e\u8bad\u7ec3\u6a21\u578b\uff0c\u4f3c\u4e4e\u90fd\u6ca1\u6709\u53d7\u5230 PAWS \u6240\u6784\u5efa\u7684\u5bf9\u7acb\u6027\u7684\u5f71\u54cd\uff0c\u8fd9\u4ee3\u8868\u4e86\u76f8\u5bf9\u4e8e\u4e0a\u4e00\u4ee3\u6a21\u578b\u7684\u5de8\u5927\u8fdb\u6b65\u3002\u8fd9\u4e00\u7ed3\u679c\u8bc1\u5b9e\u4e86 Weber \u7b49\u4eba (2023a) \u7684\u53d1\u73b0\uff0c\u4ed6\u4eec\u4e5f\u53d1\u73b0\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5bf9\u51e0\u4e2a\u5bf9\u6297\u6027\u6570\u636e\u96c6\u4e2d\u7684\u865a\u5047\u76f8\u5173\u6027\u66f4\u5c11\u654f\u611f\u3002\u7136\u800c\uff0c\u5bf9\u4e8e\u6570\u5b66\u63a8\u7406\u548c\u95ee\u7b54\uff0c\u5bf9\u6297\u6027\u8868\u73b0\u660e\u663e\u4f4e\u4e8e\u975e\u5bf9\u6297\u6027\u8868\u73b0\u3002\u8fd9\u79cd\u6a21\u5f0f\u9002\u7528\u4e8e\u9884\u8bad\u7ec3\u6a21\u578b\u548c\u540e\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>5.1.4 \u6c61\u67d3\u5206\u6790<\/h4>\n<p>\u6211\u4eec\u8fdb\u884c\u4e86\u4e00\u9879\u6c61\u67d3\u5206\u6790\uff0c\u4ee5\u4f30\u8ba1\u57fa\u51c6\u5206\u6570\u53ef\u80fd\u53d7\u5230\u9884\u8bad\u7ec3\u8bed\u6599\u5e93\u4e2d\u8bc4\u4f30\u6570\u636e\u6c61\u67d3\u5f71\u54cd\u7684\u7a0b\u5ea6\u3002\u5148\u524d\u7684\u4e00\u4e9b\u5de5\u4f5c\u4f7f\u7528\u4e86\u591a\u79cd\u4e0d\u540c\u7684\u6c61\u67d3\u65b9\u6cd5\u548c\u8d85\u53c2\u6570\u2014\u2014\u6211\u4eec\u53c2\u8003\u4e86 Singh \u7b49\u4eba\u7684\u7814\u7a76\uff082024\uff09\u3002\u7ed3\u679c\u8868\u660e\uff0c\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5bf9\u591a\u9879\u9009\u62e9\u9898\u6807\u7b7e\u7684\u53d8\u5316\u4ee5\u53ca\u5c11\u6837\u672c\u63d0\u793a\u6807\u7b7e\u7ed3\u6784\u53d8\u5316\u975e\u5e38\u7a33\u5065\uff082024 \u5e74\u6982\u8ff0\uff09\u3002\u4efb\u4f55\u8fd9\u4e9b\u65b9\u6cd5\u90fd\u53ef\u80fd\u51fa\u73b0\u5047\u9633\u6027\u548c\u5047\u9634\u6027\uff0c\u5982\u4f55\u6700\u4f73\u5730\u8fdb\u884c\u6c61\u67d3\u5206\u6790\u76ee\u524d\u4ecd\u7136\u662f\u4e00\u4e2a\u5f00\u653e\u7684\u7814\u7a76\u9886\u57df\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u4e3b\u8981\u9075\u5faa Singh \u7b49\u4eba\u7684\u5efa\u8bae\uff082024\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3461\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc6cf0e41112bc7.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"448\" height=\"187\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc6cf0e41112bc7.png 649w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cc6cf0e41112bc7-300x125.png 300w\" sizes=\"auto, (max-width: 448px) 100vw, 448px\" \/><\/strong><\/p>\n<p>&nbsp;<\/p>\n<p><strong><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3462\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9a35661657528cd.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"409\" height=\"551\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9a35661657528cd.png 457w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9a35661657528cd-223x300.png 223w\" sizes=\"auto, (max-width: 409px) 100vw, 409px\" \/><\/strong><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u65b9\u6cd5\uff1a<\/strong>\u5177\u4f53\u6765\u8bf4\uff0cSingh \u7b49\u4eba (2024) \u5efa\u8bae\u6839\u636e\u54ea\u79cd\u65b9\u6cd5\u5bfc\u81f4\u201c\u5e72\u51c0\u201d\u6570\u636e\u96c6\u548c\u6574\u4e2a\u6570\u636e\u96c6\u4e4b\u95f4\u7684\u6700\u5927\u5dee\u5f02\u6765\u7ecf\u9a8c\u9009\u62e9\u6c61\u67d3\u68c0\u6d4b\u65b9\u6cd5\uff0c\u4ed6\u4eec\u5c06\u5176\u79f0\u4e3a\u4f30\u8ba1\u7684\u6027\u80fd\u589e\u76ca\u3002\u5bf9\u4e8e\u6240\u6709\u8bc4\u4f30\u6570\u636e\u96c6\uff0c\u6211\u4eec\u57fa\u4e8e 8-gram \u91cd\u53e0\u8fdb\u884c\u8bc4\u5206\uff0cSingh \u7b49\u4eba (2024) \u53d1\u73b0\u8fd9\u79cd\u65b9\u6cd5\u5bf9\u8bb8\u591a\u6570\u636e\u96c6\u90fd\u662f\u51c6\u786e\u7684\u3002\u6211\u4eec\u5c06\u6570\u636e\u96c6 D \u7684\u4e00\u4e2a\u793a\u4f8b\u89c6\u4e3a\u53d7\u6c61\u67d3\u7684\uff0c\u5982\u679c\u5176\u6807\u8bb0 T<sub>D<\/sub>\u00a0\u4e2d\u7684\u4e00\u4e2a\u6bd4\u4f8b\u81f3\u5c11\u5728\u9884\u8bad\u7ec3\u8bed\u6599\u5e93\u4e2d\u51fa\u73b0\u4e00\u6b21\u3002\u6211\u4eec\u9488\u5bf9\u6bcf\u4e2a\u6570\u636e\u96c6\u5355\u72ec\u9009\u62e9 T<sub>D<\/sub>\uff0c\u6839\u636e\u54ea\u4e2a\u503c\u663e\u793a\u6700\u5927\u663e\u8457\u4f30\u8ba1\u6027\u80fd\u589e\u76ca\uff08\u8de8\u4e09\u4e2a\u6a21\u578b\u5927\u5c0f\uff09\u3002<\/p>\n<p><strong>\u7ed3\u679c\uff1a<\/strong>\u8868 15 \u663e\u793a\u4e86\u5982\u4e0a\u6240\u8ff0\uff0c\u6240\u6709\u4e3b\u8981\u57fa\u51c6\u7684\u8bc4\u4f30\u6570\u636e\u4e2d\u88ab\u8ba4\u4e3a\u53d7\u6c61\u67d3\u7684\u767e\u5206\u6bd4\uff0c\u4ee5\u83b7\u5f97\u6700\u5927\u7684\u4f30\u8ba1\u6027\u80fd\u589e\u76ca\u3002\u4ece\u8be5\u8868\u4e2d\uff0c\u6211\u4eec\u6392\u9664\u4e86\u7ed3\u679c\u4e0d\u663e\u7740\u7684\u57fa\u51c6\u6570\u5b57\uff0c\u4f8b\u5982\u7531\u4e8e\u5e72\u51c0\u6216\u53d7\u6c61\u67d3\u7684\u96c6\u5408\u6837\u672c\u592a\u5c11\uff0c\u6216\u8005\u89c2\u5bdf\u5230\u7684\u6027\u80fd\u589e\u76ca\u4f30\u8ba1\u663e\u793a\u51fa\u6781\u5176\u4e0d\u7a33\u5b9a\u884c\u4e3a\u3002<\/p>\n<p>\u5728\u8868 15 \u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u5bf9\u4e8e\u4e00\u4e9b\u6570\u636e\u96c6\uff0c\u6c61\u67d3\u5177\u6709\u5f88\u5927\u5f71\u54cd\uff0c\u800c\u5bf9\u5176\u4ed6\u6570\u636e\u96c6\u5219\u6ca1\u6709\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e PiQA \u548c HellaSwag\uff0c\u6c61\u67d3\u4f30\u8ba1\u548c\u6027\u80fd\u589e\u76ca\u4f30\u8ba1\u90fd\u8f83\u9ad8\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5bf9\u4e8e Natural Questions\uff0c\u4f30\u8ba1\u7684 52% \u6c61\u67d3\u4f3c\u4e4e\u5bf9\u6027\u80fd\u51e0\u4e4e\u6ca1\u6709\u5f71\u54cd\u3002\u5bf9\u4e8e SQuAD \u548c MATH\uff0c\u4f4e\u9608\u503c\u4f1a\u5bfc\u81f4\u9ad8\u6c34\u5e73\u7684\u6c61\u67d3\uff0c\u4f46\u4e0d\u4f1a\u5e26\u6765\u6027\u80fd\u63d0\u5347\u3002\u8fd9\u8868\u660e\u6c61\u67d3\u53ef\u80fd\u5bf9\u8fd9\u4e9b\u6570\u636e\u96c6\u6ca1\u6709\u5e2e\u52a9\uff0c\u6216\u8005\u9700\u8981\u66f4\u5927\u7684 n \u6765\u83b7\u5f97\u66f4\u597d\u7684\u4f30\u8ba1\u3002\u6700\u540e\uff0c\u5bf9\u4e8e MBPP\u3001HumanEval\u3001MMLU \u548c MMLU-Pro\uff0c\u53ef\u80fd\u9700\u8981\u5176\u4ed6\u6c61\u67d3\u68c0\u6d4b\u65b9\u6cd5\uff1a\u5373\u4f7f\u4f7f\u7528\u66f4\u9ad8\u7684\u9608\u503c\uff0c8-gram \u91cd\u53e0\u4e5f\u4f1a\u7ed9\u51fa\u5982\u6b64\u9ad8\u7684\u6c61\u67d3\u5206\u6570\uff0c\u4ee5\u81f3\u4e8e\u65e0\u6cd5\u83b7\u5f97\u826f\u597d\u7684\u6027\u80fd\u589e\u76ca\u4f30\u8ba1\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>5.2 \u5fae\u8c03\u8bed\u8a00\u6a21\u578b<\/h3>\n<p>\u6211\u4eec\u5c55\u793a\u4e86\u5728\u4e0d\u540c\u80fd\u529b\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u8bad\u7ec3\u540e\u7684Llama 3\u6a21\u578b\u7684\u7ed3\u679c\u3002\u4e0e\u9884\u8bad\u7ec3\u7c7b\u4f3c\uff0c\u6211\u4eec\u5c06\u4f5c\u4e3a\u8bc4\u4f30\u7684\u4e00\u90e8\u5206\u751f\u6210\u7684\u6570\u636e\u53d1\u5e03\u5230\u516c\u5171\u53ef\u7528\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u8fd9\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u53ef\u4ee5\u5728Huggingface\u4e0a\u627e\u5230\uff08\u6b64\u5904\u63d2\u5165\u94fe\u63a5\uff09\u3002\u6709\u5173\u6211\u4eec\u8bc4\u4f30\u8bbe\u7f6e\u7684\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u89c1\u6b64\u5904\uff08\u6b64\u5904\u63d2\u5165\u94fe\u63a5\uff09\u3002<\/p>\n<p><strong>\u57fa\u51c6\u6d4b\u8bd5\u548c\u6307\u6807\u3002<\/strong>\u886816\u6982\u8ff0\u4e86\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\uff0c\u6309\u80fd\u529b\u5206\u7c7b\u3002\u6211\u4eec\u5c06\u901a\u8fc7\u4e0e\u6bcf\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u63d0\u793a\u8fdb\u884c\u7cbe\u786e\u5339\u914d\u6765\u5bf9\u540e\u8bad\u7ec3\u6570\u636e\u8fdb\u884c\u53bb\u6c61\u67d3\u5904\u7406\u3002\u9664\u4e86\u6807\u51c6\u7684\u5b66\u672f\u57fa\u51c6\u6d4b\u8bd5\u5916\uff0c\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u4eba\u5de5\u8bc4\u4f30\u4e0d\u540c\u80fd\u529b\u7684\u6d4b\u8bd5\u3002\u8be6\u7ec6\u4fe1\u606f\u8bf7\u53c2\u89c1\u7b2c5.3\u8282\u3002<\/p>\n<p><strong>\u5b9e\u9a8c\u8bbe\u7f6e\u3002<\/strong>\u6211\u4eec\u91c7\u7528\u4e0e\u9884\u8bad\u7ec3\u9636\u6bb5\u7c7b\u4f3c\u7684\u5b9e\u9a8c\u8bbe\u7f6e\uff0c\u5e76\u5bf9Llama 3\u4e0e\u5176\u4ed6\u5177\u6709\u53ef\u6bd4\u5c3a\u5bf8\u548c\u80fd\u529b\u7684\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\u5206\u6790\u3002\u5728\u53ef\u80fd\u7684\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u5c06\u81ea\u5df1\u8bc4\u4f30\u5176\u4ed6\u6a21\u578b\u7684\u6027\u80fd\u5e76\u5c06\u5176\u7ed3\u679c\u4e0e\u62a5\u544a\u6570\u5b57\u8fdb\u884c\u6bd4\u8f83\uff0c\u9009\u62e9\u6700\u4f73\u5206\u6570\u3002\u6709\u5173\u6211\u4eec\u8bc4\u4f30\u8bbe\u7f6e\u7684\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u89c1\u6b64\u5904\uff08\u6b64\u5904\u63d2\u5165\u94fe\u63a5\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3463\" style=\"width: 806px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3463\" class=\"wp-image-3463\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4640c7066610f60.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"796\" height=\"363\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4640c7066610f60.png 1105w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4640c7066610f60-300x137.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4640c7066610f60-1024x467.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4640c7066610f60-768x350.png 768w\" sizes=\"auto, (max-width: 796px) 100vw, 796px\" \/><p id=\"caption-attachment-3463\" class=\"wp-caption-text\">\u886816 \u6309\u7c7b\u522b\u5217\u51fa\u7684\u540e\u8bad\u7ec3\u57fa\u51c6\u6d4b\u8bd5\u3002\u6211\u4eec\u7528\u6765\u8bc4\u4f30\u8bad\u7ec3\u540e\u7684Llama 3\u6a21\u578b\u7684\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u7684\u6982\u8ff0\uff0c\u6309\u80fd\u529b\u6392\u5e8f\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<h4>5.2.1 \u901a\u7528\u77e5\u8bc6\u548c\u6307\u4ee4\u9075\u5faa\u57fa\u51c6\u6d4b\u8bd5<\/h4>\n<p>\u6211\u4eec\u4f7f\u7528\u8868 2 \u4e2d\u5217\u51fa\u7684\u57fa\u51c6\u6d4b\u8bd5\u6765\u8bc4\u4f30 Llama 3 \u5728\u901a\u7528\u77e5\u8bc6\u548c\u6307\u4ee4\u9075\u5faa\u65b9\u9762\u7684\u80fd\u529b\u3002<\/p>\n<p><strong>\u901a\u7528\u77e5\u8bc6\uff1a<\/strong>\u00a0\u6211\u4eec\u5229\u7528 MMLU (Hendrycks et al., 2021a) \u548c MMLU-Pro (Wang et al., 2024b) \u6765\u8bc4\u4f30 Llama 3 \u5728\u57fa\u4e8e\u77e5\u8bc6\u7684\u95ee\u7b54\u80fd\u529b\u4e0a\u7684\u8868\u73b0\u3002\u5bf9\u4e8e MMLU\uff0c\u6211\u4eec\u5728\u6ca1\u6709 CoT \u7684 5 \u6b21\u793a\u4f8b\u6807\u51c6\u8bbe\u7f6e\u4e0b\u62a5\u544a\u5b50\u4efb\u52a1\u51c6\u786e\u7387\u7684\u5b8f\u89c2\u5e73\u5747\u503c\u3002MMLU-Pro \u662f MMLU \u7684\u6269\u5c55\u7248\u672c\uff0c\u5305\u542b\u66f4\u5177\u6311\u6218\u6027\u7684\u3001\u4ee5\u63a8\u7406\u4e3a\u91cd\u70b9\u7684\u95ee\u9898\uff0c\u6d88\u9664\u4e86\u566a\u58f0\u95ee\u9898\uff0c\u5e76\u5c06\u9009\u62e9\u8303\u56f4\u4ece\u56db\u4e2a\u9009\u9879\u6269\u5c55\u5230\u5341\u4e2a\u9009\u9879\u3002\u8003\u8651\u5230\u5b83\u4fa7\u91cd\u4e8e\u590d\u6742\u63a8\u7406\uff0c\u6211\u4eec\u62a5\u544a\u4e86 MMLU-Pro \u7684 5 \u6b21\u793a\u4f8b CoT\u3002\u6240\u6709\u4efb\u52a1\u90fd\u683c\u5f0f\u5316\u4e3a\u751f\u6210\u4efb\u52a1\uff0c\u7c7b\u4f3c\u4e8e simple-evals (OpenAI, 2024)\u3002<\/p>\n<p>\u5982\u8868 2 \u6240\u793a\uff0c\u6211\u4eec\u7684 8B \u548c 70B Llama 3 \u53d8\u4f53\u5728\u4e24\u4e2a\u901a\u7528\u77e5\u8bc6\u4efb\u52a1\u4e0a\u90fd\u4f18\u4e8e\u5176\u4ed6\u76f8\u4f3c\u89c4\u6a21\u7684\u6a21\u578b\u3002\u6211\u4eec\u7684 405B \u6a21\u578b\u4f18\u4e8e GPT-4 \u548c Nemotron 4 340B\uff0cClaude 3.5 Sonnet \u5728\u66f4\u5927\u6a21\u578b\u4e2d\u9886\u5148\u3002<\/p>\n<p><strong>\u6307\u4ee4\u9075\u5faa\uff1a<\/strong>\u00a0\u6211\u4eec\u4f7f\u7528 IFEval (Zhou et al., 2023) \u6765\u8bc4\u4f30 Llama 3 \u548c\u5176\u4ed6\u6a21\u578b\u9075\u5faa\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\u7684\u80fd\u529b\u3002IFEval \u5305\u62ec\u5927\u7ea6 500 \u4e2a\u201c\u53ef\u9a8c\u8bc1\u6307\u4ee4\u201d\uff0c\u4f8b\u5982\u201c\u7528\u8d85\u8fc7 400 \u4e2a\u5b57\u5199\u201d \uff0c\u8fd9\u4e9b\u6307\u4ee4\u53ef\u4ee5\u4f7f\u7528\u542f\u53d1\u5f0f\u65b9\u6cd5\u8fdb\u884c\u9a8c\u8bc1\u3002\u6211\u4eec\u5728\u8868 2 \u4e2d\u62a5\u544a\u4e86\u4e25\u683c\u548c\u5bbd\u677e\u7ea6\u675f\u4e0b\u7684\u63d0\u793a\u7ea7\u548c\u6307\u4ee4\u7ea7\u51c6\u786e\u7387\u7684\u5e73\u5747\u503c\u3002\u8bf7\u6ce8\u610f\uff0c\u6240\u6709 Llama 3 \u53d8\u4f53\u5728 IFEval \u4e0a\u90fd\u4f18\u4e8e\u53ef\u6bd4\u8f83\u7684\u6a21\u578b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>5.2.2 \u80fd\u529b\u8003\u8bd5<\/h4>\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u5728\u6700\u521d\u8bbe\u8ba1\u7528\u4e8e\u6d4b\u8bd5\u4eba\u7c7b\u7684\u4e00\u7cfb\u5217\u80fd\u529b\u6d4b\u8bd5\u4e0a\u8bc4\u4f30\u6211\u4eec\u7684\u6a21\u578b\u3002\u6211\u4eec\u4ece\u516c\u5f00\u53ef\u7528\u7684\u5b98\u65b9\u6765\u6e90\u83b7\u53d6\u8fd9\u4e9b\u8003\u8bd5\uff1b\u5bf9\u4e8e\u67d0\u4e9b\u8003\u8bd5\uff0c\u6211\u4eec\u5c06\u4e0d\u540c\u8003\u8bd5\u96c6\u7684\u5e73\u5747\u5206\u6570\u62a5\u544a\u4e3a\u6bcf\u4e2a\u80fd\u529b\u6d4b\u8bd5\u7684\u7ed3\u679c\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5e73\u5747\uff1a<\/p>\n<ul>\n<li>GRE\uff1a\u6559\u80b2\u6d4b\u8bd5\u670d\u52a1\u63d0\u4f9b\u7684\u5b98\u65b9 GRE \u5b9e\u8df5\u8003\u8bd5 1 \u548c 2\uff1b<\/li>\n<li>LSAT\uff1a\u5b98\u65b9\u9884\u6d4b\u8bd5 71\u300173\u300180 \u548c 93\uff1b<\/li>\n<li>SAT\uff1a\u6765\u81ea\u300a\u5b98\u65b9 SAT \u5b66\u4e60\u6307\u5357\u300b2018 \u7248\u7684 8 \u4e2a\u8003\u8bd5\uff1b<\/li>\n<li>AP\uff1a\u6bcf\u4e2a\u5b66\u79d1\u4e00\u4e2a\u5b98\u65b9\u7ec3\u4e60\u8003\u8bd5\uff1b<\/li>\n<li>GMAT\uff1a\u5b98\u65b9 GMAT \u5728\u7ebf\u8003\u8bd5\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u8003\u8bd5\u4e2d\u7684\u95ee\u9898\u5305\u542b\u4e86\u591a\u9879\u9009\u62e9\u9898\u548c\u751f\u6210\u5f0f\u95ee\u9898\u3002\u6211\u4eec\u5c06\u6392\u9664\u4efb\u4f55\u9644\u5e26\u56fe\u50cf\u7684\u95ee\u9898\u3002\u5bf9\u4e8e\u5305\u542b\u591a\u4e2a\u6b63\u786e\u9009\u9879\u7684 GRE \u9898\u76ee\uff0c\u53ea\u6709\u5f53\u6a21\u578b\u9009\u62e9\u4e86\u6240\u6709\u6b63\u786e\u9009\u9879\u65f6\uff0c\u6211\u4eec\u624d\u5c06\u8f93\u51fa\u9650\u5b9a\u4e3a\u6b63\u786e\u3002\u5728\u6709\u8d85\u8fc7\u4e00\u4e2a\u8003\u8bd5\u96c6\u7684\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u4f1a\u4f7f\u7528\u5c11\u91cf\u63d0\u793a\u8fdb\u884c\u8bc4\u4f30\u3002\u6211\u4eec\u5c06\u5206\u6570\u8c03\u6574\u5230 130-170 \u7684\u8303\u56f4\u5185\uff08\u7528\u4e8e GRE\uff09\uff0c\u5e76\u62a5\u544a\u5176\u4ed6\u6240\u6709\u8003\u8bd5\u7684\u51c6\u786e\u7387\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3464\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03eb7747bd28c02.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"735\" height=\"659\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03eb7747bd28c02.png 841w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03eb7747bd28c02-300x269.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03eb7747bd28c02-768x689.png 768w\" sizes=\"auto, (max-width: 735px) 100vw, 735px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u7684\u7ed3\u679c\u89c1\u8868 17\u3002\u6211\u4eec\u53d1\u73b0\u6211\u4eec\u7684 Llama 3 405B \u6a21\u578b\u7684\u8868\u73b0\u4e0e <a href=\"https:\/\/www.kdjingpai.com\/pt\/claudeanquanfubai\/\">Claude<\/a> 3.5 Sonnet \u548c GPT-4 4o \u975e\u5e38\u76f8\u4f3c\u3002\u800c\u6211\u4eec\u7684 70B \u6a21\u578b\u5219\u5c55\u73b0\u51fa\u66f4\u4e3a\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u8868\u73b0\u3002\u5b83\u6bd4 GPT-3.5 Turbo \u663e\u8457\u66f4\u597d\uff0c\u5e76\u5728\u8bb8\u591a\u6d4b\u8bd5\u4e2d\u8d85\u8fc7\u4e86 Nemotron 4 340B\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>5.2.3 \u7f16\u7801\u57fa\u51c6<\/h4>\n<p>\u6211\u4eec\u5728\u591a\u4e2a\u6d41\u884c\u7684 Python \u548c\u591a\u8bed\u8a00\u7f16\u7a0b\u57fa\u51c6\u4e0a\u8bc4\u4f30 Llama 3 \u7684\u4ee3\u7801\u751f\u6210\u80fd\u529b\u3002\u4e3a\u4e86\u8861\u91cf\u6a21\u578b\u751f\u6210\u529f\u80fd\u6b63\u786e\u4ee3\u7801\u7684\u6709\u6548\u6027\uff0c\u6211\u4eec\u4f7f\u7528 pass@N \u6307\u6807\uff0c\u8be5\u6307\u6807\u8bc4\u4f30 N \u4e2a\u751f\u6210\u4e2d\u7684\u4e00\u7ec4\u5355\u5143\u6d4b\u8bd5\u901a\u8fc7\u7387\u3002\u6211\u4eec\u62a5\u544a pass@1 \u7684\u7ed3\u679c\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3465\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/29284b188591aa8.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"646\" height=\"303\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/29284b188591aa8.png 1035w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/29284b188591aa8-300x141.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/29284b188591aa8-1024x481.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/29284b188591aa8-768x361.png 768w\" sizes=\"auto, (max-width: 646px) 100vw, 646px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3466\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a90ba121ee90273.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"760\" height=\"215\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a90ba121ee90273.png 1029w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a90ba121ee90273-300x85.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a90ba121ee90273-1024x290.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/a90ba121ee90273-768x217.png 768w\" sizes=\"auto, (max-width: 760px) 100vw, 760px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><strong>Python \u4ee3\u7801\u751f\u6210\u3002<\/strong> HumanEval\uff08Chen \u7b49\u4eba\uff0c2021\uff09\u548c MBPP\uff08Austin \u7b49\u4eba\uff0c2021\uff09\u662f\u6d41\u884c\u7684 Python \u4ee3\u7801\u751f\u6210\u57fa\u51c6\u6d4b\u8bd5\uff0c\u5b83\u4eec\u4fa7\u91cd\u4e8e\u76f8\u5bf9\u7b80\u5355\u3001\u81ea\u5305\u542b\u7684\u529f\u80fd\u3002HumanEval+\uff08Liu \u7b49\u4eba\uff0c2024a\uff09\u662f HumanEval \u7684\u589e\u5f3a\u7248\u672c\uff0c\u5176\u4e2d\u751f\u6210\u4e86\u66f4\u591a\u6d4b\u8bd5\u7528\u4f8b\u4ee5\u907f\u514d\u5047\u9633\u6027\u3002MBPP EvalPlus \u57fa\u51c6\u7248\u672c (v0.2.0) \u662f\u4ece\u539f\u59cb MBPP\uff08\u8bad\u7ec3\u548c\u6d4b\u8bd5\uff09\u6570\u636e\u96c6\u4e2d\u7684 974 \u4e2a\u521d\u59cb\u95ee\u9898\u4e2d\u9009\u51fa\u7684 378 \u4e2a\u826f\u597d\u683c\u5f0f\u7684\u95ee\u9898\uff08Liu \u7b49\u4eba\uff0c2024a\uff09\u3002\u8fd9\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u7684\u7ed3\u679c\u5982\u8868 18 \u6240\u793a\u3002\u5728\u8fd9\u4e9b Python \u53d8\u4f53\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cLlama 3 8B \u548c 70B \u8868\u73b0\u4f18\u4e8e\u76f8\u540c\u89c4\u6a21\u7684\u6a21\u578b\u8868\u73b0\u76f8\u4f3c\u3002\u5bf9\u4e8e\u6700\u5927\u7684\u6a21\u578b\uff0cLlama 3 405B\u3001Claude 3.5 Sonnet \u548c GPT-4o \u8868\u73b0\u7c7b\u4f3c\uff0c\u5176\u4e2d GPT-4o \u7684\u7ed3\u679c\u6700\u5f3a\u3002<\/p>\n<p><strong>\u6a21\u578b:<\/strong>\u00a0\u6211\u4eec\u5c06 Llama 3 \u4e0e\u5176\u4ed6\u7c7b\u4f3c\u89c4\u6a21\u7684\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u5bf9\u4e8e\u6700\u5927\u6a21\u578b Llama 3 405B\uff0cClaude 3.5 Sonnet \u548c GPT-4o \u7684\u8868\u73b0\u76f8\u4f3c\uff0c\u5176\u4e2d GPT-4o \u663e\u793a\u51fa\u6700\u4f73\u7ed3\u679c\u3002<\/p>\n<p><strong>\u591a\u7f16\u7a0b\u8bed\u8a00\u4ee3\u7801\u751f\u6210\uff1a<\/strong>\u00a0\u4e3a\u8bc4\u4f30\u9664 Python \u4e4b\u5916\u7684\u4ee3\u7801\u751f\u6210\u80fd\u529b\uff0c\u6211\u4eec\u62a5\u544a\u4e86\u57fa\u4e8e HumanEval \u548c MBPP \u95ee\u9898\u7ffb\u8bd1\u7684 MultiPL-E\uff08Cassano \u7b49\u4eba\uff0c2023\uff09\u57fa\u51c6\u6d4b\u8bd5\u7684\u7ed3\u679c\u3002\u8868 19 \u663e\u793a\u4e86\u4e00\u90e8\u5206\u6d41\u884c\u7f16\u7a0b\u8bed\u8a00\u7684\u7ed3\u679c\u3002<\/p>\n<p>\u8bf7\u6ce8\u610f\uff0c\u4e0e\u8868 18 \u4e2d\u7684 Python \u5bf9\u5e94\u9879\u76f8\u6bd4\uff0c\u6027\u80fd\u6709\u663e\u8457\u4e0b\u964d\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>5.2.4 \u591a\u8bed\u8a00\u57fa\u51c6\u6d4b\u8bd5<\/h4>\n<p>Llama 3 \u652f\u6301 8 \u79cd\u8bed\u8a00 &#8211; \u82f1\u8bed\u3001\u5fb7\u8bed\u3001\u6cd5\u8bed\u3001\u610f\u5927\u5229\u8bed\u3001\u8461\u8404\u7259\u8bed\u3001\u5370\u5730\u8bed\u3001\u897f\u73ed\u7259\u8bed\u548c\u6cf0\u8bed\uff0c\u5c3d\u7ba1\u57fa\u7840\u6a21\u578b\u7684\u8bad\u7ec3\u4f7f\u7528\u4e86\u66f4\u5e7f\u6cdb\u7684\u8bed\u8a00\u96c6\u5408\u3002\u5728 \u886820\u4e2d \u663e\u793a\u4e86\u6211\u4eec\u5728\u591a\u8bed\u8a00 MMLU (Hendrycks \u7b49\u4eba, 2021a) \u548c\u591a\u8bed\u8a00\u5c0f\u5b66\u6570\u5b66 (MGSM) (Shi \u7b49\u4eba, 2022) \u57fa\u51c6\u6d4b\u8bd5\u4e0a\u5bf9 Llama 3 \u8fdb\u884c\u8bc4\u4f30\u7684\u7ed3\u679c\u3002<\/p>\n<ul>\n<li><strong>\u591a\u8bed\u8a00 MMLU<\/strong>: \u6211\u4eec\u4f7f\u7528 Google \u7ffb\u8bd1\u5c06 MMLU \u7684\u95ee\u9898\u3001\u5c11\u4f8b\u793a\u4f8b\u548c\u7b54\u6848\u7ffb\u8bd1\u6210\u4e0d\u540c\u8bed\u8a00\u3002\u6211\u4eec\u5c06\u4efb\u52a1\u8bf4\u660e\u4fdd\u7559\u4e3a\u82f1\u8bed\uff0c\u5e76\u5728 5-shot \u8bbe\u7f6e\u4e0b\u8fdb\u884c\u8bc4\u4f30\u3002<\/li>\n<li><strong>MGSM (Shi \u7b49\u4eba, 2022)<\/strong>\uff1a\u5bf9\u4e8e\u6211\u4eec\u7684 Llama 3 \u6a21\u578b\uff0c\u6211\u4eec\u62a5\u544a\u4e86 MGSM \u7684 0-shot CoT \u7ed3\u679c\u3002\u591a\u8bed\u8a00 MMLU \u662f\u4e00\u4e2a\u5185\u90e8\u57fa\u51c6\u6d4b\u8bd5\uff0c\u5176\u4e2d\u5305\u542b\u5c06 MMLU (Hendrycks \u7b49\u4eba, 2021a) \u7684\u95ee\u9898\u548c\u7b54\u6848\u7ffb\u8bd1\u6210 7 \u79cd\u8bed\u8a00\u2014\u2014\u6211\u4eec\u62a5\u544a\u7684 5-shot \u7ed3\u679c\u662f\u8de8\u8fd9\u4e9b\u8bed\u8a00\u7684\u5e73\u5747\u503c\u3002<\/li>\n<\/ul>\n<p>\u5bf9\u4e8e MGSM (Shi \u7b49\u4eba, 2022)\uff0c\u6211\u4eec\u4f7f\u7528\u4e0e simple-evals (OpenAI, 2024) \u4e2d\u76f8\u540c\u7684\u539f\u751f\u63d0\u793a\u6765\u6d4b\u8bd5\u6211\u4eec\u7684\u6a21\u578b\uff0c\u5e76\u5c06\u5176\u7f6e\u4e8e 0-shot CoT \u73af\u5883\u4e2d\u3002\u5728\u8868 20 \u4e2d\uff0c\u6211\u4eec\u62a5\u544a\u4e86 MGSM \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5305\u542b\u7684\u6240\u6709\u8bed\u8a00\u7684\u5e73\u5747\u7ed3\u679c\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3467\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/38a75edfddc8b0d.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"339\" height=\"341\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/38a75edfddc8b0d.png 514w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/38a75edfddc8b0d-298x300.png 298w\" sizes=\"auto, (max-width: 339px) 100vw, 339px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u53d1\u73b0 Llama 3 405B \u5728 MGSM \u4e0a\u4f18\u4e8e\u5927\u591a\u6570\u5176\u4ed6\u6a21\u578b\uff0c\u5e73\u5747\u5f97\u5206\u8fbe\u5230 91.6%\u3002\u5728 MMLU \u4e0a\uff0c\u4e0e\u4e0a\u8ff0\u82f1\u8bed MMLU \u7ed3\u679c\u4e00\u81f4\uff0cLlama 3 405B \u843d\u540e\u4e8e GPT-4o 2%\u3002\u53e6\u4e00\u65b9\u9762\uff0cLlama 3 \u7684 70B \u548c 8B \u6a21\u578b\u90fd\u8868\u73b0\u51fa\u8272\uff0c\u5728\u4e24\u4e2a\u4efb\u52a1\u4e0a\u5747\u4ee5\u8f83\u5927\u4f18\u52bf\u9886\u5148\u4e8e\u7ade\u4e89\u5bf9\u624b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>5.2.5 \u6570\u5b66\u548c\u63a8\u7406\u57fa\u51c6<\/h4>\n<p>\u6211\u4eec\u7684\u6570\u5b66\u548c\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c\u5982\u88682\u6240\u793a\u3002Llama 3 8B\u6a21\u578b\u5728GSM8K\u3001MATH\u548cGPQA\u4e0a\u90fd\u4f18\u4e8e\u5176\u4ed6\u540c\u7b49\u89c4\u6a21\u7684\u6a21\u578b\u3002\u6211\u4eec\u768470B\u6a21\u578b\u5728\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5747\u8868\u73b0\u51fa\u663e\u8457\u4f18\u4e8e\u540c\u7c7b\u6a21\u578b\u7684\u6027\u80fd\u3002\u6700\u540e\uff0cLlama 3 405B\u6a21\u578b\u5728\u5176\u7c7b\u522b\u4e2d\u662fGSM8K\u548cARC-C\u6700\u4f73\u6a21\u578b\uff0c\u800c\u5728MATH\u4e0a\uff0c\u5b83\u662f\u7b2c\u4e8c\u597d\u7684\u6a21\u578b\u3002\u5728GPQA\u4e0a\uff0c\u5b83\u4e0eGPT-4 4o\u7ade\u4e89\u6fc0\u70c8\uff0c\u800cClaude 3.5 Sonnet\u4ee5\u663e\u8457\u4f18\u52bf\u4f4d\u5217\u699c\u9996\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>5.2.6 \u957f\u4e0a\u4e0b\u6587\u57fa\u51c6\u6d4b\u8bd5<\/h4>\n<p>\u6211\u4eec\u8003\u8651\u4e86\u4e00\u7cfb\u5217\u8de8\u8d8a\u4e0d\u540c\u9886\u57df\u548c\u6587\u672c\u7c7b\u578b\u4efb\u52a1\u3002\u5728\u4e0b\u9762\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u4f7f\u7528\u65e0\u504f\u8bc4\u4f30\u534f\u8bae\u7684\u5b50\u4efb\u52a1\uff0c\u5373\u57fa\u4e8e\u51c6\u786e\u6027\u7684\u6307\u6807\u800c\u4e0d\u662fn-gram\u91cd\u53e0\u6307\u6807\u3002\u6211\u4eec\u8fd8\u4f18\u5148\u8003\u8651\u53d1\u73b0\u65b9\u5dee\u8f83\u4f4e\u7684\u4efb\u52a1\u3002<\/p>\n<ul>\n<li><strong>Needle-in-a-Haystack (Kamradt, 2023)<\/strong>\u00a0\u8861\u91cf\u6a21\u578b\u68c0\u7d22\u9690\u85cf\u5728\u957f\u6587\u6863\u968f\u673a\u90e8\u5206\u4fe1\u606f\u7684\u80fd\u529b\u3002\u6211\u4eec\u7684Llama 3\u6a21\u578b\u8868\u73b0\u51fa\u5b8c\u7f8e\u7684\u9488\u68c0\u7d22\u6027\u80fd\uff0c\u6210\u529f\u5730\u5728\u6240\u6709\u6587\u6863\u6df1\u5ea6\u548c\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0b\u68c0\u7d22\u5230100%\u7684\u201c\u9488\u201d\u3002\u6211\u4eec\u8fd8\u6d4b\u91cf\u4e86Multi-needle\uff08\u8868 21\uff09\u6027\u80fd\uff0c\u8fd9\u662f\u4e00\u4e2aNeedle-in-a-Haystack\u7684\u53d8\u5316\uff0c\u5176\u4e2d\u6211\u4eec\u5c06\u56db\u6839\u201c\u9488\u201d\u63d2\u5165\u4e0a\u4e0b\u6587\u5e76\u6d4b\u8bd5\u6a21\u578b\u662f\u5426\u53ef\u4ee5\u68c0\u7d22\u5176\u4e2d\u7684\u4e24\u6839\u3002\u6211\u4eec\u7684Llama 3\u6a21\u578b\u5b9e\u73b0\u4e86\u63a5\u8fd1\u5b8c\u7f8e\u7684\u68c0\u7d22\u7ed3\u679c\u3002<\/li>\n<li><strong>ZeroSCROLLS (Shaham et al., 2023)<\/strong>\u662f\u4e00\u4e2a\u9488\u5bf9\u957f\u6587\u672c\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u7684\u96f6\u6837\u672c\u57fa\u51c6\u6d4b\u8bd5\u3002\u7531\u4e8e\u771f\u5b9e\u7b54\u6848\u672a\u516c\u5f00\u63d0\u4f9b\uff0c\u6211\u4eec\u62a5\u544a\u4e86\u9a8c\u8bc1\u96c6\u4e0a\u7684\u6570\u5b57\u3002\u6211\u4eec\u7684Llama 3 405B\u548c70B\u6a21\u578b\u5728\u8be5\u57fa\u51c6\u6d4b\u8bd5\u7684\u5404\u79cd\u4efb\u52a1\u4e0a\u4e0e\u5176\u4ed6\u6a21\u578b\u6301\u5e73\u6216\u8d85\u8d8a\u5b83\u4eec\u3002<\/li>\n<li><strong>InfiniteBench (Zhang et al., 2024)<\/strong>\u00a0\u8981\u6c42\u6a21\u578b\u7406\u89e3\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e2d\u7684\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb\u3002\u6211\u4eec\u5728En.QA\uff08\u5c0f\u8bf4\u4e0a\u7684\u95ee\u7b54\uff09\u548cEn.MC\uff08\u5c0f\u8bf4\u4e0a\u7684\u591a\u9879\u9009\u62e9\u95ee\u7b54\uff09\u4e0a\u8bc4\u4f30Llama 3\uff0c\u5176\u4e2d\u6211\u4eec\u7684405B\u6a21\u578b\u4f18\u4e8e\u6240\u6709\u5176\u4ed6\u6a21\u578b\u3002\u589e\u76ca\u5728En.QA\u4e0a\u5c24\u4e3a\u663e\u7740\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3468\" style=\"width: 842px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3468\" class=\"wp-image-3468\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03b87870a9a747d.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"832\" height=\"340\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03b87870a9a747d.png 1039w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03b87870a9a747d-300x123.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03b87870a9a747d-1024x419.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/03b87870a9a747d-768x314.png 768w\" sizes=\"auto, (max-width: 832px) 100vw, 832px\" \/><p id=\"caption-attachment-3468\" class=\"wp-caption-text\">\u8868 21 \u957f\u6587\u672c\u57fa\u51c6\u6d4b\u8bd5\u3002\u5bf9\u4e8e ZeroSCROLLS\uff08Shaham \u7b49\u4eba\uff0c2023\uff09\uff0c\u6211\u4eec\u62a5\u544a\u7684\u662f\u9a8c\u8bc1\u96c6\u4e0a\u7684\u7ed3\u679c\u3002\u5bf9\u4e8e QuALITY \u6211\u4eec\u62a5\u544a\u7cbe\u786e\u5339\u914d\uff0c\u5bf9\u4e8e Qasper &#8211; f1\uff0c\u5bf9\u4e8e SQuALITY &#8211; rougeL\u3002\u6211\u4eec\u4e3a InfiniteBench\uff08\u5f20\u7b49\uff0c2024\uff09En.QA \u6307\u6807\u62a5\u544a f1\uff0c\u5e76\u4e3a En.MC \u62a5\u544a\u51c6\u786e\u7387\u3002\u5bf9\u4e8e Multi-needle\uff08Kamradt\uff0c2023\uff09\uff0c\u6211\u4eec\u5728\u4e0a\u4e0b\u6587\u4e2d\u63d2\u5165 4 \u4e2a\u9488\u5934\uff0c\u6d4b\u8bd5\u6a21\u578b\u662f\u5426\u80fd\u591f\u68c0\u7d22\u4e0d\u540c\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684 2 \u4e2a\u9488\u5934\uff0c\u6211\u4eec\u8ba1\u7b97\u6700\u591a 128k \u7684 10 \u4e2a\u5e8f\u5217\u957f\u5ea6\u7684\u5e73\u5747\u53ec\u56de\u7387\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<h4>5.2.7 \u5de5\u5177\u4f7f\u7528\u6027\u80fd<\/h4>\n<p>\u6211\u4eec\u5229\u7528\u4e00\u7cfb\u5217\u96f6\u6837\u672c\u5de5\u5177\u4f7f\u7528\uff08\u5373\u51fd\u6570\u8c03\u7528\uff09\u57fa\u51c6\u6d4b\u8bd5\u8bc4\u4f30\u4e86\u6211\u4eec\u7684\u6a21\u578b\uff1aNexus (Srinivasan \u7b49\u4eba\uff0c2023)\u3001API-Bank (Li \u7b49\u4eba\uff0c2023b)\u3001Gorilla API-Bench (Patil \u7b49\u4eba\uff0c2023) \u548c Berkeley \u51fd\u6570\u8c03\u7528\u6392\u884c\u699c (BFCL) (Yan \u7b49\u4eba\uff0c2024)\u3002\u7ed3\u679c\u5982\u8868 22 \u6240\u793a\u3002<\/p>\n<p>\u5728 Nexus \u4e0a\uff0c\u6211\u4eec\u7684 Llama 3 \u53d8\u4f53\u8868\u73b0\u6700\u4f73\uff0c\u4f18\u4e8e\u5176\u4ed6\u540c\u7c7b\u6a21\u578b\u3002 \u5728 API-Bank \u4e0a\uff0c\u6211\u4eec\u7684 Llama 3 8B \u548c 70B \u6a21\u578b\u5728\u76f8\u5e94\u7c7b\u522b\u4e2d\u663e\u7740\u8d85\u8d8a\u5176\u4ed6\u6a21\u578b\u3002 405B \u6a21\u578b\u4ec5\u843d\u540e\u4e8e Claude 3.5 Sonnet 0.6%\u3002\u6700\u540e\uff0c\u6211\u4eec\u7684 405B \u548c 70B \u6a21\u578b\u5728 BFCL \u4e0a\u8868\u73b0\u4f18\u5f02\uff0c\u5e76\u4e14\u5728\u5404\u81ea\u7684\u89c4\u6a21\u7c7b\u522b\u4e2d\u6392\u540d\u7b2c\u4e8c\u3002 Llama 3 8B \u5728\u5176\u7c7b\u522b\u4e2d\u8868\u73b0\u6700\u4f73\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u8fdb\u884c\u4eba\u5de5\u8bc4\u4f30\u4ee5\u6d4b\u8bd5\u6a21\u578b\u7684\u5de5\u5177\u4f7f\u7528\u80fd\u529b\uff0c\u91cd\u70b9\u5173\u6ce8\u4ee3\u7801\u6267\u884c\u4efb\u52a1\u3002 \u6211\u4eec\u6536\u96c6\u4e86 2000 \u4e2a\u4e0e\u4ee3\u7801\u6267\u884c\uff08\u4e0d\u5305\u62ec\u7ed8\u56fe\u6216\u6587\u4ef6\u4e0a\u4f20\uff09\u76f8\u5173\u7684\u7528\u6237\u63d0\u793a\u3001\u7ed8\u56fe\u751f\u6210\u548c\u6587\u4ef6\u4e0a\u4f20\u3002\u8fd9\u4e9b\u63d0\u793a\u6765\u81ea <a href=\"https:\/\/www.kdjingpai.com\/pt\/lmsys-org\/\">LMSys<\/a> \u6570\u636e\u96c6 (Chiang \u7b49\u4eba\uff0c2024)\u3001GAIA \u57fa\u51c6\u6d4b\u8bd5 (Mialon \u7b49\u4eba\uff0c2023b)\u3001\u4eba\u5de5\u6807\u6ce8\u8005\u4ee5\u53ca\u5408\u6210\u751f\u6210\u3002\u6211\u4eec\u5c06 Llama 3 405B \u4e0e GPT-4o \u8fdb\u884c\u6bd4\u8f83\uff0c\u4f7f\u7528 OpenAI \u7684 Assistants API10 \u3002\u7ed3\u679c\u5982\u56fe 16 \u6240\u793a\u3002\u5728\u4ec5\u9650\u6587\u672c\u7684\u4ee3\u7801\u6267\u884c\u4efb\u52a1\u548c\u7ed8\u56fe\u751f\u6210\u65b9\u9762\uff0cLlama 3 405B \u660e\u663e\u4f18\u4e8e GPT-4o\u3002\u7136\u800c\uff0c\u5728\u6587\u4ef6\u4e0a\u4f20\u7528\u4f8b\u4e0a\uff0c\u5b83\u843d\u540e\u4e8e GPT-4o\u3002<\/p>\n<p>&nbsp;<\/p>\n<div class=\"mceTemp\"><\/div>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3469\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/8911c26221ade2f.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"565\" height=\"397\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/8911c26221ade2f.png 699w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/8911c26221ade2f-300x211.png 300w\" sizes=\"auto, (max-width: 565px) 100vw, 565px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>5.3 \u4eba\u5de5\u8bc4\u4f30<\/h3>\n<p>\u9664\u4e86\u5728\u6807\u51c6\u57fa\u51c6\u6570\u636e\u96c6\u4e0a\u7684\u8bc4\u4f30\u5916\uff0c\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u4eba\u7c7b\u8bc4\u4f30\u3002\u8fd9\u4e9b\u8bc4\u4f30\u4f7f\u6211\u4eec\u80fd\u591f\u6d4b\u91cf\u548c\u4f18\u5316\u6a21\u578b\u6027\u80fd\u7684\u66f4\u7ec6\u5fae\u65b9\u9762\uff0c\u4f8b\u5982\u6a21\u578b\u7684\u8bed\u6c14\u3001\u5197\u957f\u7a0b\u5ea6\u4ee5\u53ca\u5bf9\u7ec6\u5fae\u5dee\u522b\u548c\u6587\u5316\u80cc\u666f\u7684\u7406\u89e3\u3002\u7cbe\u5fc3\u8bbe\u8ba1\u7684 rengren \u8bc4\u4f30\u4e0e\u7528\u6237\u4f53\u9a8c\u5bc6\u5207\u76f8\u5173\uff0c\u63d0\u4f9b\u4e86\u6709\u5173\u6a21\u578b\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u8868\u73b0\u7684\u89c1\u89e3\u3002<\/p>\n<blockquote><p>https:\/\/platform.openai.com\/docs\/assistants\/overview<\/p>\n<p>\u5bf9\u4e8e\u591a\u8f6e\u4eba\u7c7b\u8bc4\u4f30\uff0c\u6bcf\u4e2a\u63d0\u793a\u4e2d\u7684\u8f6e\u6570\u5728 2 \u5230 11 \u4e4b\u95f4\u3002\u6211\u4eec\u8bc4\u4f30\u6a21\u578b\u5728\u6700\u540e\u4e00\u8f6e\u7684\u54cd\u5e94\u3002<\/p><\/blockquote>\n<p><strong>\u63d0\u793a\u6536\u96c6<\/strong>\u3002 \u6211\u4eec\u6536\u96c6\u4e86\u9ad8\u8d28\u91cf\u7684\u63d0\u793a\uff0c\u6db5\u76d6\u8303\u56f4\u5e7f\u6cdb\u7684\u7c7b\u522b\u548c\u96be\u5ea6\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u9996\u5148\u5f00\u53d1\u4e86\u4e00\u4e2a\u5206\u7c7b\u6cd5\uff0c\u5176\u4e2d\u5305\u542b\u5c3d\u53ef\u80fd\u591a\u6a21\u578b\u80fd\u529b\u7684\u7c7b\u522b\u548c\u5b50\u7c7b\u522b\u3002\u6211\u4eec\u4f7f\u7528\u8fd9\u4e2a\u5206\u7c7b\u6cd5\u6536\u96c6\u4e86\u7ea6 7,000 \u4e2a\u63d0\u793a\uff0c\u6db5\u76d6\u516d\u4e2a\u5355\u6b21\u80fd\u529b\uff08\u82f1\u8bed\u3001\u63a8\u7406\u3001\u7f16\u7801\u3001\u5370\u5730\u8bed\u3001\u897f\u73ed\u7259\u8bed\u548c\u8461\u8404\u7259\u8bed\uff09\u548c\u4e09\u4e2a\u591a\u8f6e\u80fd\u529b<sup>11<\/sup>\u00a0(\u82f1\u8bed\u3001\u63a8\u7406\u548c\u7f16\u7801)\u3002\u6211\u4eec\u786e\u4fdd\u5728\u6bcf\u4e2a\u7c7b\u522b\u4e2d\uff0c\u63d0\u793a\u5728\u5b50\u7c7b\u522b\u4e4b\u95f4\u5747\u5300\u5206\u5e03\u3002\u6211\u4eec\u8fd8\u5c06\u6bcf\u4e2a\u63d0\u793a\u5206\u7c7b\u4e3a\u4e09\u4e2a\u96be\u5ea6\u7ea7\u522b\u4e4b\u4e00\uff0c\u5e76\u786e\u4fdd\u6211\u4eec\u7684\u63d0\u793a\u96c6\u5408\u5305\u542b\u5927\u7ea6 10% \u7684\u7b80\u5355\u63d0\u793a\u300130% \u7684\u4e2d\u7b49\u96be\u5ea6\u63d0\u793a\u548c 60% \u7684\u56f0\u96be\u63d0\u793a\u3002\u6240\u6709\u7684\u4eba\u7c7b\u8bc4\u4f30\u56fe 16 Llama 3 405B \u4e0e GPT-4o \u5728\u4ee3\u7801\u6267\u884c\u4efb\u52a1\uff08\u5305\u62ec\u7ed8\u56fe\u548c\u6587\u4ef6\u4e0a\u4f20\uff09\u4e0a\u7684\u4eba\u7c7b\u8bc4\u4f30\u7ed3\u679c\u3002 Llama 3 405B \u5728\u4ee3\u7801\u6267\u884c\uff08\u4e0d\u5305\u62ec\u7ed8\u56fe\u6216\u6587\u4ef6\u4e0a\u4f20\uff09\u4ee5\u53ca\u60c5\u8282\u751f\u6210\u65b9\u9762\u90fd\u4f18\u4e8e GPT-4o\uff0c\u4f46\u5728\u6587\u4ef6\u4e0a\u4f20\u7528\u4f8b\u65b9\u9762\u843d\u540e\u3002<\/p>\n<p>\u63d0\u793a\u96c6\u90fd\u7ecf\u8fc7\u4e86\u4e25\u683c\u7684\u8d28\u91cf\u4fdd\u8bc1\u6d41\u7a0b\u3002\u5efa\u6a21\u56e2\u961f\u65e0\u6cd5\u8bbf\u95ee\u6211\u4eec\u7684\u4eba\u7c7b\u8bc4\u4f30\u63d0\u793a\uff0c\u4ee5\u9632\u6b62\u610f\u5916\u6c61\u67d3\u6216\u5bf9\u6d4b\u8bd5\u96c6\u8fc7\u5ea6\u62df\u5408\u3002<\/p>\n<p><strong>\u8bc4\u4f30\u8fc7\u7a0b<\/strong>\u3002\u4e3a\u4e86\u5bf9\u4e24\u4e2a\u6a21\u578b\u8fdb\u884c\u6210\u5bf9\u7684\u4eba\u7c7b\u8bc4\u4f30\uff0c\u6211\u4eec\u4f1a\u8be2\u95ee\u4eba\u7c7b\u6807\u6ce8\u5458\u4ed6\u4eec\u66f4\u559c\u6b22\u54ea\u4e24\u4e2a\u6a21\u578b\u54cd\u5e94\uff08\u7531\u4e0d\u540c\u7684\u6a21\u578b\u4ea7\u751f\uff09\u3002\u6807\u6ce8\u5458\u4f7f\u7528 7 \u5206\u5236\u8bc4\u5206\uff0c\u4f7f\u4ed6\u4eec\u80fd\u591f\u8868\u660e\u4e00\u4e2a\u6a21\u578b\u54cd\u5e94\u662f\u5426\u6bd4\u53e6\u4e00\u4e2a\u6a21\u578b\u54cd\u5e94\u597d\u5f97\u591a\u3001\u66f4\u597d\u3001\u7565\u597d\u6216\u5927\u81f4\u76f8\u540c\u3002\u5f53\u6807\u6ce8\u5458\u8868\u793a\u4e00\u4e2a\u6a21\u578b\u54cd\u5e94\u6bd4\u53e6\u4e00\u4e2a\u6a21\u578b\u54cd\u5e94\u66f4\u597d\u6216\u597d\u5f97\u591a\u65f6\uff0c\u6211\u4eec\u5c31\u4f1a\u5c06\u5176\u89c6\u4e3a\u8be5\u6a21\u578b\u7684\u201c\u80dc\u5229\u201d\u3002\u6211\u4eec\u5c06\u5bf9\u6a21\u578b\u8fdb\u884c\u6210\u5bf9\u6bd4\u8f83\uff0c\u5e76\u5728\u63d0\u793a\u96c6\u4e2d\u62a5\u544a\u6bcf\u4e2a\u80fd\u529b\u7684\u83b7\u80dc\u7387\u3002<\/p>\n<p><strong>\u7ed3\u679c<\/strong>\u3002 \u6211\u4eec\u4f7f\u7528\u4eba\u7c7b\u8bc4\u4f30\u8fc7\u7a0b\u5c06 Llama 3 405B \u4e0e GPT-4\uff080125 API \u7248\u672c\uff09\u3001GPT-4o\uff08API \u7248\u672c\uff09\u548c Claude 3.5 Sonnet\uff08API \u7248\u672c\uff09\u8fdb\u884c\u6bd4\u8f83\u3002\u8fd9\u4e9b\u8bc4\u4f30\u7684\u7ed3\u679c\u5982\u56fe 17 \u6240\u793a\u3002\u6211\u4eec\u89c2\u5bdf\u5230 Llama 3 405B \u7684\u8868\u73b0\u4e0e GPT-4 \u7684 0125 API \u7248\u672c\u5927\u81f4\u76f8\u5f53\uff0c\u800c\u4e0e GPT-4o \u548c Claude 3.5 Sonnet \u76f8\u6bd4\uff0c\u7ed3\u679c\u53c2\u5dee\u4e0d\u9f50\uff08\u4e00\u4e9b\u80dc\u5229\u548c\u4e00\u4e9b\u5931\u8d25\uff09\u3002\u5728\u51e0\u4e4e\u6240\u6709\u80fd\u529b\u4e0a\uff0cLlama 3 \u548c GPT-4 \u7684\u83b7\u80dc\u7387\u90fd\u5728\u8bef\u5dee\u8303\u56f4\u5185\u3002\u5728\u591a\u8f6e\u63a8\u7406\u548c\u7f16\u7801\u4efb\u52a1\u4e2d\uff0cLlama 3 405B \u7684\u6027\u80fd\u4f18\u4e8e GPT-4\uff0c\u4f46\u5728\u591a\u8bed\u8a00\uff08\u5370\u5730\u8bed\u3001\u897f\u73ed\u7259\u8bed\u548c\u8461\u8404\u7259\u8bed\uff09\u63d0\u793a\u65b9\u9762\u5219\u4e0d\u5982 GPT-4\u3002Llama 3 \u5728\u82f1\u8bed\u63d0\u793a\u4e0a\u7684\u8868\u73b0\u4e0e GPT-4 \u76f8\u5f53\uff0c\u5728\u591a\u8bed\u8a00\u63d0\u793a\u4e0a\u4e0e Claude 3.5 Sonnet \u76f8\u5f53\uff0c\u5e76\u5728\u5355\u8f6e\u548c\u591a\u8f6e\u82f1\u8bed\u63d0\u793a\u4e0a\u4f18\u4e8e Claude 3.5 Sonnet\u3002\u7136\u800c\uff0c\u5b83\u5728\u7f16\u7801\u548c\u63a8\u7406\u7b49\u65b9\u9762\u4e0d\u53ca Claude 3.5 Sonnet\u3002\u4ece\u8d28\u4e0a\u8bb2\uff0c\u6211\u4eec\u53d1\u73b0\u6a21\u578b\u5728\u4eba\u7c7b\u8bc4\u4f30\u4e2d\u7684\u6027\u80fd\u53d7\u5230\u8bed\u6c14\u3001\u54cd\u5e94\u7ed3\u6784\u548c\u5197\u957f\u7a0b\u5ea6\u7b49\u7ec6\u5fae\u56e0\u7d20\u7684\u5f88\u5927\u5f71\u54cd\u2014\u2014\u8fd9\u4e9b\u90fd\u662f\u6211\u4eec\u5728\u540e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6b63\u5728\u4f18\u5316\u7684\u56e0\u7d20\u3002\u603b\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u7684\u4eba\u7c7b\u8bc4\u4f30\u7ed3\u679c\u4e0e\u6807\u51c6\u57fa\u51c6\u8bc4\u4f30\u7684\u7ed3\u679c\u4e00\u81f4\uff1aLlama 3 405B \u4e0e\u9886\u5148\u7684\u884c\u4e1a\u6a21\u578b\u7ade\u4e89\u975e\u5e38\u6fc0\u70c8\uff0c\u4f7f\u5176\u6210\u4e3a\u8868\u73b0\u6700\u4f73\u7684\u516c\u5f00\u53ef\u7528\u6a21\u578b\u3002<\/p>\n<p><strong>\u5c40\u9650\u6027<\/strong>\u3002 \u6240\u6709\u7684\u4eba\u7c7b\u8bc4\u4f30\u7ed3\u679c\u90fd\u7ecf\u8fc7\u4e86\u4e25\u683c\u7684\u6570\u636e\u8d28\u91cf\u4fdd\u8bc1\u6d41\u7a0b\u3002\u7136\u800c\uff0c\u7531\u4e8e\u5b9a\u4e49\u6a21\u578b\u54cd\u5e94\u5ba2\u89c2\u6807\u51c6\u5f88\u56f0\u96be\uff0c\u4eba\u7c7b\u8bc4\u4f30\u4ecd\u7136\u53ef\u80fd\u4f1a\u53d7\u5230\u4eba\u7c7b\u6807\u6ce8\u5458\u7684\u4e2a\u4eba\u504f\u89c1\u3001\u80cc\u666f\u548c\u504f\u597d\u7684\u5f71\u54cd\uff0c\u8fd9\u53ef\u80fd\u5bfc\u81f4\u4e0d\u4e00\u81f4\u6216\u4e0d\u53ef\u9760\u7684\u7ed3\u679c\u3002<\/p>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3470\" style=\"width: 1226px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3470\" class=\"wp-image-3470 size-full\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cea2f5fa129384b.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"1216\" height=\"511\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cea2f5fa129384b.png 1216w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cea2f5fa129384b-300x126.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cea2f5fa129384b-1024x430.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cea2f5fa129384b-768x323.png 768w\" sizes=\"auto, (max-width: 1216px) 100vw, 1216px\" \/><p id=\"caption-attachment-3470\" class=\"wp-caption-text\">\u56fe 16 Llama 3 405B \u4e0e GPT-4o \u5728\u4ee3\u7801\u6267\u884c\u4efb\u52a1\uff08\u5305\u62ec\u7ed8\u56fe\u548c\u6587\u4ef6\u4e0a\u4f20\uff09\u4e0a\u7684\u4eba\u7c7b\u8bc4\u4f30\u7ed3\u679c\u3002 Llama 3 405B \u5728\u4ee3\u7801\u6267\u884c\uff08\u4e0d\u5305\u62ec\u7ed8\u56fe\u548c\u6587\u4ef6\u4e0a\u4f20\uff09\u4ee5\u53ca\u7ed8\u56fe\u751f\u6210\u65b9\u9762\u4f18\u4e8e GPT-4o\uff0c\u4f46\u5728\u6587\u4ef6\u4e0a\u4f20\u7528\u4f8b\u65b9\u9762\u843d\u540e\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3471\" style=\"width: 880px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3471\" class=\"wp-image-3471 \" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cb9c7be351edfac.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"870\" height=\"377\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cb9c7be351edfac.png 1214w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cb9c7be351edfac-300x130.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cb9c7be351edfac-1024x444.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/cb9c7be351edfac-768x333.png 768w\" sizes=\"auto, (max-width: 870px) 100vw, 870px\" \/><p id=\"caption-attachment-3471\" class=\"wp-caption-text\">\u56fe 17 Llama 3 405B \u6a21\u578b\u7684\u4eba\u5de5\u8bc4\u4f30\u7ed3\u679c\u3002\u5de6\uff1a\u4e0e GPT-4 \u7684\u6bd4\u8f83\u3002\u4e2d\uff1a\u4e0e GPT-4o \u7684\u6bd4\u8f83\u3002\u53f3\uff1a\u4e0e Claude 3.5 Sonnet \u7684\u6bd4\u8f83\u3002\u6240\u6709\u7ed3\u679c\u5747\u5305\u542b 95% \u7f6e\u4fe1\u533a\u95f4\uff0c\u5e76\u6392\u9664\u5e73\u5c40\u3002<\/p><\/div>\n<h3><\/h3>\n<h3>5.4 \u5b89\u5168\u6027<\/h3>\n<p>\u5b89\u5168\u6027\u90e8\u5206\u6d89\u53ca\u654f\u611f\u8bcd\uff0c\u53ef\u8df3\u8fc7\u6216\u4e0b\u8f7dPDF\u67e5\u9605\uff0c\u611f\u8c22<br \/>\n<div class=\"huoduan_hide_box\" style=\"border:1px dashed #F60; padding:10px; margin:10px 0; line-height:200%; color:#F00; background-color:#FFF4FF; overflow:hidden; clear:both;\"><img loading=\"lazy\" decoding=\"async\" class=\"wxpic\" align=\"right\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/05\/d8668ed8023fbe2.jpg\" style=\"width:150px;height:150px;margin-left:20px;display:inline;border:none\" width=\"150\" height=\"150\"  alt=\"AI\u751f\u4ea7\u529b\u5e94\u7528\" \/><span style=\"font-size:18px;\">\u6b64\u5904\u5185\u5bb9\u5df2\u7ecf\u88ab\u4f5c\u8005\u9690\u85cf\uff0c\u8bf7\u8f93\u5165\u9a8c\u8bc1\u7801\u67e5\u770b\u5185\u5bb9<\/span><form method=\"post\" style=\"margin:10px 0;\"><span class=\"yzts\" style=\"font-size:18px;float:left;\">\u9a8c\u8bc1\u7801\uff1a<\/span><input name=\"huoduan_verifycode\" id=\"verifycode\" type=\"text\" value=\"\" style=\"border:none;float:left;width:80px; height:32px; line-height:30px; padding:0 5px; border:1px solid #FF6600;-moz-border-radius: 0px;  -webkit-border-radius: 0px;  border-radius:0px;\" \/><input id=\"verifybtn\" style=\"border:none;float:left;width:80px; height:32px; line-height:32px; padding:0 5px; background-color:#F60; text-align:center; border:none; cursor:pointer; color:#FFF;-moz-border-radius: 0px; font-size:14px;  -webkit-border-radius: 0px;  border-radius:0px;\" name=\"\" type=\"submit\" value=\"\u63d0\u4ea4\u67e5\u770b\" \/><\/form><div style=\"clear:left;\"><\/div><span style=\"color:#00BF30\">\u8bf7\u5173\u6ce8\u672c\u7ad9\u5fae\u4fe1\u516c\u4f17\u53f7\uff0c\u56de\u590d\u201c<span style=\"color:blue\">\u9a8c\u8bc1\u7801<\/span>\u201d\uff0c\u83b7\u53d6\u9a8c\u8bc1\u7801\u3002\u5728\u5fae\u4fe1\u91cc\u641c\u7d22\u201c<span style=\"color:blue\">AI\u751f\u4ea7\u529b\u5e94\u7528<\/span>\u201d\u6216\u8005\u201c<span style=\"color:blue\">Artificial9527<\/span>\u201d\u6216\u8005\u5fae\u4fe1\u626b\u63cf\u53f3\u4fa7\u4e8c\u7ef4\u7801\u90fd\u53ef\u4ee5\u5173\u6ce8\u672c\u7ad9\u5fae\u4fe1\u516c\u4f17\u53f7\u3002<\/span><div class=\"cl\"><\/div><\/div><\/p>\n<p>&nbsp;<\/p>\n<h2>6 \u63a8\u8bba<\/h2>\n<p>\u6211\u4eec\u7814\u7a76\u4e86\u4e24\u79cd\u4e3b\u8981\u6280\u672f\u6765\u63d0\u9ad8 Llama 3 405B \u6a21\u578b\u7684\u63a8\u7406\u6548\u7387\uff1a(1) \u7ba1\u9053\u5e76\u884c\u548c (2) FP8 \u91cf\u5316\u3002\u6211\u4eec\u5df2\u516c\u5f00\u53d1\u5e03\u4e86 FP8 \u91cf\u5316\u7684\u5b9e\u73b0\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>6.1 \u7ba1\u9053\u5e76\u884c (Pipeline Parallelism)<\/h3>\n<p>\u5f53\u4f7f\u7528 BF16 \u8868\u793a\u6a21\u578b\u53c2\u6570\u65f6\uff0cLlama 3 405B \u6a21\u578b\u65e0\u6cd5\u88c5\u5165\u5355\u4e2a\u914d\u5907 8 \u4e2a Nvidia H100 GPU \u7684\u673a\u5668\u7684 GPU \u5185\u5b58\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u4f7f\u7528\u4e86 BF16 \u7cbe\u5ea6\uff0c\u5c06\u6a21\u578b\u63a8\u7406\u5e76\u884c\u5316\u5230\u4e24\u53f0\u673a\u5668\u4e0a\u7684 16 \u4e2a GPU \u4e0a\u3002\u5728\u6bcf\u53f0\u673a\u5668\u5185\uff0c\u9ad8\u5e26\u5bbd\u7684 NVLink \u4f7f\u5f97\u80fd\u591f\u4f7f\u7528\u5f20\u91cf\u5e76\u884c (Shoeybi et al.\uff0c2019)\u3002\u7136\u800c\uff0c\u8de8\u8282\u70b9\u8fde\u63a5\u5177\u6709\u8f83\u4f4e\u7684\u5e26\u5bbd\u548c\u66f4\u9ad8\u7684\u5ef6\u8fdf\uff0c\u56e0\u6b64\u6211\u4eec\u4f7f\u7528\u7ba1\u9053\u5e76\u884c (Huang et al.\uff0c2019)\u3002<\/p>\n<p>\u5728\u4f7f\u7528\u7ba1\u9053\u5e76\u884c\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6c14\u6ce1\u662f\u4e00\u4e2a\u4e3b\u8981\u7684\u6548\u7387\u95ee\u9898\uff08\u53c2\u89c1\u7b2c 3.3 \u8282\uff09\u3002\u4f46\u662f\uff0c\u5b83\u4eec\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5e76\u4e0d\u662f\u4e00\u4e2a\u95ee\u9898\uff0c\u56e0\u4e3a\u63a8\u7406\u4e0d\u6d89\u53ca\u9700\u8981\u7ba1\u9053\u51b2\u5237\u7684\u53cd\u5411\u4f20\u9012\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4f7f\u7528\u5fae\u6279\u91cf\u5316\u6765\u63d0\u9ad8\u7ba1\u9053\u5e76\u884c\u63a8\u7406\u7684\u541e\u5410\u91cf\u3002<\/p>\n<p>\u6211\u4eec\u8bc4\u4f30\u4e86\u5728 4,096 \u4e2a\u8f93\u5165\u6807\u8bb0\u548c 256 \u4e2a\u8f93\u51fa\u6807\u8bb0\u7684\u63a8\u7406\u5de5\u4f5c\u8d1f\u8f7d\u4e2d\u4f7f\u7528\u4e24\u4e2a\u5fae\u6279\u91cf\u7684\u6548\u679c\uff0c\u8fd9\u4e24\u4e2a\u5fae\u6279\u91cf\u5206\u522b\u7528\u4e8e\u63a8\u7406\u7684\u5173\u952e\u503c\u7f13\u5b58\u9884\u586b\u5145\u9636\u6bb5\u548c\u89e3\u7801\u9636\u6bb5\u3002\u6211\u4eec\u53d1\u73b0\u5fae\u6279\u91cf\u5316\u53ef\u4ee5\u63d0\u9ad8\u76f8\u540c\u672c\u5730\u6279\u91cf\u5927\u5c0f\u7684\u63a8\u7406\u541e\u5410\u91cf\uff1b\u53c2\u89c1\u56fe 24\u3002\u8fd9\u4e9b\u6539\u8fdb\u6765\u81ea\u5fae\u6279\u91cf\u5316\u5728\u4e24\u4e2a\u9636\u6bb5\u90fd\u80fd\u591f\u5e76\u53d1\u6267\u884c\u5fae\u6279\u91cf\u3002\u7531\u4e8e\u5fae\u6279\u91cf\u5316\u5bfc\u81f4\u989d\u5916\u7684\u540c\u6b65\u70b9\uff0c\u4e5f\u4f1a\u589e\u52a0\u5ef6\u8fdf\uff0c\u4f46\u603b\u4f53\u800c\u8a00\uff0c\u5fae\u6279\u91cf\u5316\u4ecd\u7136\u4f1a\u5bfc\u81f4\u66f4\u597d\u7684\u541e\u5410\u91cf-\u5ef6\u8fdf\u6743\u8861\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3482\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/67609bf5635f97b.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"845\" height=\"168\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/67609bf5635f97b.png 1239w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/67609bf5635f97b-300x60.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/67609bf5635f97b-1024x203.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/67609bf5635f97b-768x152.png 768w\" sizes=\"auto, (max-width: 845px) 100vw, 845px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3483\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/523ec08b4c6108a.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"763\" height=\"297\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/523ec08b4c6108a.png 1216w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/523ec08b4c6108a-300x117.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/523ec08b4c6108a-1024x399.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/523ec08b4c6108a-768x299.png 768w\" sizes=\"auto, (max-width: 763px) 100vw, 763px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>6.2 FP8 \u91cf\u5316<\/h3>\n<p>\u6211\u4eec\u5229\u7528 H100 GPU \u672c\u8eab\u7684 FP8 \u652f\u6301\u8fdb\u884c\u4f4e\u7cbe\u5ea6\u63a8\u7406\u5b9e\u9a8c\u3002\u4e3a\u4e86\u5b9e\u73b0\u4f4e\u7cbe\u5ea6\u63a8\u7406\uff0c\u6211\u4eec\u5c06\u6a21\u578b\u5185\u90e8\u7684\u5927\u591a\u6570\u77e9\u9635\u4e58\u6cd5\u5e94\u7528\u4e86 FP8 \u91cf\u5316\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u6a21\u578b\u4e2d\u524d\u9988\u7f51\u7edc\u5c42\u7684\u7edd\u5927\u591a\u6570\u53c2\u6570\u548c\u6fc0\u6d3b\u503c\u8fdb\u884c\u4e86\u91cf\u5316\uff0c\u8fd9\u4e9b\u5c42\u5360\u63a8\u7406\u8ba1\u7b97\u65f6\u95f4\u7684\u7ea6 50%\u3002\u6211\u4eec\u6ca1\u6709\u5bf9\u6a21\u578b\u7684\u81ea\u6ce8\u610f\u529b\u5c42\u4e2d\u7684\u53c2\u6570\u8fdb\u884c\u91cf\u5316\u3002\u6211\u4eec\u5229\u7528\u52a8\u6001\u7f29\u653e\u56e0\u5b50\u6765\u63d0\u9ad8\u7cbe\u5ea6 (Xiao \u7b49\u4eba\uff0c2024b)\uff0c\u5e76\u4f18\u5316\u6211\u4eec\u7684 CUDA \u5185\u683815 \u4ee5\u51cf\u5c11\u8ba1\u7b97\u6bd4\u4f8b\u7684\u5f00\u9500\u3002<\/p>\n<p>\u6211\u4eec\u53d1\u73b0 Llama 3 405B \u7684\u8d28\u91cf\u5bf9\u67d0\u4e9b\u7c7b\u578b\u7684\u91cf\u5316\u5f88\u654f\u611f\uff0c\u5e76\u505a\u4e86\u4e00\u4e9b\u989d\u5916\u7684\u6539\u52a8\u4ee5\u63d0\u9ad8\u6a21\u578b\u8f93\u51fa\u8d28\u91cf\uff1a<\/p>\n<ol>\n<li>\u4e0e Zhang \u7b49\u4eba (2021) \u76f8\u4f3c\uff0c\u6211\u4eec\u6ca1\u6709\u5bf9\u7b2c\u4e00\u4e2a\u548c\u6700\u540e\u4e00\u4e2a Transformer \u5c42\u8fdb\u884c\u91cf\u5316\u3002<\/li>\n<li>\u9ad8\u6392\u5217\u5ea6\u6807\u8bb0\uff08\u4f8b\u5982\u65e5\u671f\uff09\u53ef\u80fd\u5bfc\u81f4\u8f83\u5927\u7684\u6fc0\u6d3b\u503c\u3002\u53cd\u8fc7\u6765\uff0c\u8fd9\u4f1a\u5bfc\u81f4 FP8 \u4e2d\u8f83\u9ad8\u7684\u52a8\u6001\u7f29\u653e\u56e0\u5b50\uff0c\u5e76\u4ea7\u751f\u4e0d\u53ef\u5ffd\u89c6\u6570\u91cf\u7684\u6d6e\u70b9\u4e0b\u6ea2\uff0c\u4ece\u800c\u5bfc\u81f4\u89e3\u7801\u9519\u8bef\u3002\u56fe 26 \u663e\u793a\u4e86\u4f7f\u7528 BF16 \u548c FP8 \u63a8\u7406\u7684 Llama 3 405B \u7684\u5956\u52b1\u5f97\u5206\u5206\u5e03\u3002\u6211\u4eec\u7684 FP8 \u91cf\u5316\u65b9\u6cd5\u5bf9\u6a21\u578b\u7684\u54cd\u5e94\u51e0\u4e4e\u6ca1\u6709\u5f71\u54cd\u3002<\/li>\n<\/ol>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5c06\u52a8\u6001\u7f29\u653e\u56e0\u5b50\u4e0a\u9650\u8bbe\u7f6e\u4e3a 1200\u3002<\/p>\n<ol start=\"3\">\n<li>\u6211\u4eec\u4f7f\u7528\u9010\u884c\u91cf\u5316\uff0c\u4e3a\u53c2\u6570\u548c\u6fc0\u6d3b\u77e9\u9635\u8ba1\u7b97\u8de8\u884c\u7684\u7f29\u653e\u56e0\u5b50\uff08\u89c1\u56fe 25\uff09\u3002\u6211\u4eec\u53d1\u73b0\u8fd9\u6bd4\u5f20\u91cf\u7ea7\u91cf\u5316\u65b9\u6cd5\u6548\u679c\u66f4\u597d\u3002<\/li>\n<\/ol>\n<p>\u91cf\u5316\u9519\u8bef\u7684\u5f71\u54cd\u3002\u6807\u51c6\u57fa\u51c6\u7684\u8bc4\u4f30\u901a\u5e38\u8868\u660e\uff0c\u5373\u4f7f\u6ca1\u6709\u8fd9\u4e9b\u7f13\u89e3\u63aa\u65bd\uff0cFP8 \u63a8\u7406\u7684\u6027\u80fd\u4e5f\u4e0e BF16 \u63a8\u7406\u76f8\u5f53\u3002\u4f46\u662f\uff0c\u6211\u4eec\u53d1\u73b0\u8fd9\u6837\u7684\u57fa\u51c6\u4e0d\u80fd\u5145\u5206\u53cd\u6620 FP8 \u91cf\u5316\u7684\u5f71\u54cd\u3002\u5f53\u7f29\u653e\u56e0\u5b50\u6ca1\u6709\u88ab\u4e0a\u9650\u9650\u5236\u65f6\uff0c\u6a21\u578b\u5076\u5c14\u4f1a\u4ea7\u751f\u635f\u574f\u7684\u54cd\u5e94\uff0c\u5373\u4f7f\u57fa\u51c6\u6027\u80fd\u5f88\u5f3a\u3002<\/p>\n<p>\u4e0e\u5176\u4f9d\u9760\u57fa\u51c6\u6765\u6d4b\u91cf\u7531\u4e8e\u91cf\u5316\u5bfc\u81f4\u7684\u5206\u5e03\u53d8\u5316\uff0c\u4e0d\u5982\u5206\u6790\u4f7f\u7528 BF16 \u548c FP8 \u751f\u6210\u7684 100,000 \u4e2a\u54cd\u5e94\u7684\u5956\u52b1\u6a21\u578b\u5f97\u5206\u7684\u5206\u5e03\u3002\u56fe 26 \u663e\u793a\u4e86\u6211\u4eec\u91cf\u5316\u65b9\u6cd5\u5f97\u5230\u7684\u5956\u52b1\u5206\u5e03\u3002\u7ed3\u679c\u8868\u660e\uff0c\u6211\u4eec\u7684 FP8 \u91cf\u5316\u65b9\u6cd5\u5bf9\u6a21\u578b\u7684\u54cd\u5e94\u5f71\u54cd\u975e\u5e38\u6709\u9650\u3002<\/p>\n<p>\u6548\u7387\u5b9e\u9a8c\u8bc4\u4f30\u3002\u56fe 27 \u63cf\u8ff0\u4e86\u4f7f\u7528 Llama 3 405B \u5728\u9884\u586b\u5145\u548c\u89e3\u7801\u9636\u6bb5\u6267\u884c FP8 \u63a8\u7406\u7684\u541e\u5410\u91cf-\u5ef6\u8fdf\u6743\u8861\u5173\u7cfb\uff0c\u4f7f\u7528 4,096 \u4e2a\u8f93\u5165\u6807\u8bb0\u548c 256 \u4e2a\u8f93\u51fa\u6807\u8bb0\u3002\u8be5\u56fe\u5c06 FP8 \u63a8\u7406\u7684\u6548\u7387\u4e0e\u7b2c 6.1 \u8282\u4e2d\u63cf\u8ff0\u7684\u4e24\u53f0\u673a\u5668 BF16 \u63a8\u7406\u65b9\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u7ed3\u679c\u8868\u660e\uff0c\u4f7f\u7528 FP8 \u63a8\u7406\u53ef\u4ee5\u4f7f\u9884\u586b\u5145\u9636\u6bb5\u7684\u541e\u5410\u91cf\u63d0\u9ad8\u9ad8\u8fbe 50%\uff0c\u5e76\u4e14\u5728\u89e3\u7801\u671f\u95f4\u5927\u5e45\u63d0\u9ad8\u4e86\u541e\u5410\u91cf-\u5ef6\u8fdf\u6743\u8861\u5173\u7cfb\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3484\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7eb806b8b545704.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"818\" height=\"398\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7eb806b8b545704.png 1212w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7eb806b8b545704-300x146.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7eb806b8b545704-1024x498.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/7eb806b8b545704-768x374.png 768w\" sizes=\"auto, (max-width: 818px) 100vw, 818px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3485\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aaa27ece3e349f5.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"791\" height=\"222\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aaa27ece3e349f5.png 1212w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aaa27ece3e349f5-300x84.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aaa27ece3e349f5-1024x287.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/aaa27ece3e349f5-768x215.png 768w\" sizes=\"auto, (max-width: 791px) 100vw, 791px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3486\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1aadd83b7a99e7e.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"790\" height=\"306\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1aadd83b7a99e7e.png 1204w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1aadd83b7a99e7e-300x116.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1aadd83b7a99e7e-1024x396.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1aadd83b7a99e7e-768x297.png 768w\" sizes=\"auto, (max-width: 790px) 100vw, 790px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>7 \u89c6\u89c9\u5b9e\u9a8c<\/h2>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u5b9e\u9a8c\uff0c\u901a\u8fc7\u4e00\u79cd\u7ec4\u5408\u65b9\u6cd5\u5c06\u89c6\u89c9\u8bc6\u522b\u80fd\u529b\u6574\u5408\u5230 Llama 3 \u4e2d\u3002\u8be5\u65b9\u6cd5\u4e3b\u8981\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a<\/p>\n<p><strong>\u7b2c\u4e00\u4e2a\u9636\u6bb5:<\/strong>\u00a0\u6211\u4eec\u5c06\u9884\u8bad\u7ec3\u7684\u56fe\u50cf\u7f16\u7801\u5668 (Xu et al., 2023) \u4e0e\u9884\u8bad\u7ec3\u7684\u8bed\u8a00\u6a21\u578b\u7ed3\u5408\u8d77\u6765\uff0c\u5e76\u5728\u5927\u91cf\u7684\u56fe\u50cf\u6587\u672c\u5bf9\u4e0a\u5f15\u5165\u548c\u8bad\u7ec3\u4e00\u7ec4\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42 (Alayrac et al., 2022)\u3002\u8fd9\u5bfc\u81f4\u4e86\u5982\u56fe 28 \u6240\u793a\u7684\u6a21\u578b\u3002<\/p>\n<p><strong>\u7b2c\u4e8c\u4e2a\u9636\u6bb5:<\/strong>\u00a0\u6211\u4eec\u5f15\u5165\u4e86\u65f6\u95f4\u805a\u5408\u5c42\u548c\u989d\u5916\u7684\u89c6\u9891\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\uff0c\u8fd9\u4e9b\u5c42\u4f5c\u7528\u4e8e\u5927\u91cf\u89c6\u9891\u6587\u672c\u5bf9\u4e0a\uff0c\u4ee5\u5b66\u4e60\u6a21\u578b\u8bc6\u522b\u548c\u5904\u7406\u6765\u81ea\u89c6\u9891\u7684\u65f6\u95f4\u4fe1\u606f\u3002<\/p>\n<p>\u7ec4\u5408\u5f0f\u65b9\u6cd5\u6784\u5efa\u57fa\u7840\u6a21\u578b\u6709\u51e0\u4e2a\u4f18\u52bf:<\/p>\n<p>(1) \u5b83\u5141\u8bb8\u6211\u4eec\u5e76\u884c\u5f00\u53d1\u89c6\u89c9\u548c\u8bed\u8a00\u5efa\u6a21\u529f\u80fd\uff1b<\/p>\n<p>(2) \u5b83\u907f\u514d\u4e86\u8054\u5408\u9884\u8bad\u7ec3\u89c6\u89c9\u548c\u8bed\u8a00\u6570\u636e\u5e26\u6765\u7684\u590d\u6742\u6027\uff0c\u8fd9\u4e9b\u590d\u6742\u6027\u6e90\u4e8e\u89c6\u89c9\u6570\u636e\u7684\u6807\u8bb0\u5316\u3001\u4e0d\u540c\u6a21\u5f0f\u7684\u80cc\u666f\u56f0\u60d1\u5ea6\u7684\u5dee\u5f02\u4ee5\u53ca\u6a21\u5f0f\u4e4b\u95f4\u7684\u7ade\u4e89\uff1b<\/p>\n<p>(3) \u5b83\u4fdd\u8bc1\u4e86\u5f15\u5165\u89c6\u89c9\u8bc6\u522b\u80fd\u529b\u4e0d\u4f1a\u5f71\u54cd\u6a21\u578b\u5728\u4ec5\u6587\u672c\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\uff1b<\/p>\n<p>(4) \u4ea4\u53c9\u6ce8\u610f\u529b\u67b6\u6784\u786e\u4fdd\u6211\u4eec\u4e0d\u9700\u8981\u5c06\u5168\u5206\u8fa8\u7387\u56fe\u50cf\u4f20\u9012\u7ed9\u4e0d\u65ad\u589e\u5927\u7684 LLM \u4e3b\u5e72 (\u7279\u522b\u662f\u5728\u6bcf\u4e2a Transformer \u5c42\u4e2d\u7684\u524d\u9988\u7f51\u7edc)\uff0c\u4ece\u800c\u63d0\u9ad8\u63a8\u7406\u6548\u7387\u3002<\/p>\n<p>\u8bf7\u6ce8\u610f\uff0c\u6211\u4eec\u7684\u591a\u6a21\u6001\u6a21\u578b\u4ecd\u5728\u5f00\u53d1\u4e2d\uff0c\u5c1a\u672a\u51c6\u5907\u597d\u53d1\u5e03\u3002<\/p>\n<p>\u5728\u7b2c 7.6 \u548c 7.7 \u8282\u4e2d\u4ecb\u7ecd\u5b9e\u9a8c\u7ed3\u679c\u4e4b\u524d\uff0c\u6211\u4eec\u5c06\u63cf\u8ff0\u7528\u4e8e\u8bad\u7ec3\u89c6\u89c9\u8bc6\u522b\u80fd\u529b\u7684\u6570\u636e\u3001\u89c6\u89c9\u7ec4\u4ef6\u7684\u6a21\u578b\u67b6\u6784\u3001\u6211\u4eec\u5982\u4f55\u6269\u5c55\u8fd9\u4e9b\u7ec4\u4ef6\u7684\u8bad\u7ec3\u4ee5\u53ca\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u548c\u540e\u8bad\u7ec3\u914d\u65b9\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>7.1 \u6570\u636e<\/h3>\n<p>\u6211\u4eec\u5206\u522b\u63cf\u8ff0\u4e86\u56fe\u50cf\u548c\u89c6\u9891\u6570\u636e\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>7.1.1 \u56fe\u50cf\u6570\u636e<\/h4>\n<p>\u6211\u4eec\u7684\u56fe\u50cf\u7f16\u7801\u5668\u548c\u9002\u914d\u5668\u5728\u56fe\u50cf-\u6587\u672c\u5bf9\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u6211\u4eec\u901a\u8fc7\u4e00\u4e2a\u590d\u6742\u7684\u7684\u6570\u636e\u5904\u7406\u7ba1\u9053\u6784\u5efa\u8fd9\u4e2a\u6570\u636e\u96c6\uff0c\u8be5\u7ba1\u9053\u5305\u62ec\u56db\u4e2a\u4e3b\u8981\u9636\u6bb5\uff1a<\/p>\n<p><strong>(1) \u8d28\u91cf\u8fc7\u6ee4 (2) \u611f\u77e5\u53bb\u91cd (3) \u91cd\u91c7\u6837 (4) \u5149\u5b66\u5b57\u7b26\u8bc6\u522b<\/strong>\u00a0\u3002\u6211\u4eec\u8fd8\u5e94\u7528\u4e86\u4e00\u7cfb\u5217\u7684\u5b89\u5168\u63aa\u65bd\u3002<\/p>\n<ul>\n<li><strong>\u8d28\u91cf\u8fc7\u6ee4:<\/strong>\u00a0\u6211\u4eec\u5b9e\u65bd\u8d28\u91cf\u8fc7\u6ee4\u5668\uff0c\u901a\u8fc7\u8bf8\u5982\uff08Radford \u7b49\u4eba\uff0c2021\uff09\u4ea7\u751f\u7684\u4f4e\u5bf9\u9f50\u5206\u6570\u7b49\u542f\u53d1\u5f0f\u65b9\u6cd5\u53bb\u9664\u975e\u82f1\u8bed\u6807\u9898\u548c\u4f4e\u8d28\u91cf\u6807\u9898\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u4f4e\u4e8e\u7279\u5b9a CLIP \u5206\u6570\u7684\u6240\u6709\u56fe\u50cf-\u6587\u672c\u5bf9\u90fd\u5220\u9664\u3002<\/li>\n<li><strong>\u53bb\u91cd:<\/strong>\u00a0\u53bb\u9664\u5927\u89c4\u6a21\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u91cd\u590d\u6570\u636e\u53ef\u4ee5\u63d0\u9ad8\u6a21\u578b\u6027\u80fd\uff0c\u56e0\u4e3a\u5b83\u51cf\u5c11\u4e86\u7528\u4e8e\u5197\u4f59\u6570\u636e\u7684\u8bad\u7ec3\u8ba1\u7b97\uff08Esser \u7b49\u4eba\uff0c2024\uff1bLee \u7b49\u4eba\uff0c2021\uff1bAbbas \u7b49\u4eba\uff0c2023\uff09\uff0c\u5e76\u51cf\u5c11\u4e86\u6a21\u578b\u8bb0\u5fc6\u5316\u98ce\u9669\uff08Carlini \u7b49\u4eba\uff0c2023\uff1bSomepalli \u7b49\u4eba\uff0c2023\uff09\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u51fa\u4e8e\u6548\u7387\u548c\u9690\u79c1\u539f\u56e0\u5bf9\u8bad\u7ec3\u6570\u636e\u8fdb\u884c\u53bb\u91cd\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u4f7f\u7528\u6700\u65b0\u7684 SSCD \u590d\u5236\u68c0\u6d4b\u6a21\u578b\uff08Pizzi \u7b49\u4eba\uff0c2022\uff09\u7684\u5185\u90e8\u7248\u672c\u6765\u5927\u89c4\u6a21\u5730\u5bf9\u56fe\u50cf\u8fdb\u884c\u53bb\u91cd\u3002\u5bf9\u4e8e\u6240\u6709\u56fe\u50cf\uff0c\u6211\u4eec\u9996\u5148\u4f7f\u7528 SSCD \u6a21\u578b\u8ba1\u7b97\u4e00\u4e2a 512 \u7ef4\u7684\u8868\u793a\u3002\u7136\u540e\uff0c\u6211\u4eec\u4f7f\u7528\u8fd9\u4e9b\u5d4c\u5165\u9488\u5bf9\u6570\u636e\u96c6\u4e2d\u7684\u6240\u6709\u56fe\u50cf\u6267\u884c\u6700\u8fd1\u90bb (NN) \u641c\u7d22\uff0c\u4f7f\u7528\u4f59\u5f26\u76f8\u4f3c\u6027\u5ea6\u91cf\u3002\u6211\u4eec\u5c06\u9ad8\u4e8e\u7279\u5b9a\u76f8\u4f3c\u5ea6\u9608\u503c\u7684\u793a\u4f8b\u5b9a\u4e49\u4e3a\u91cd\u590d\u9879\u3002\u6211\u4eec\u4f7f\u7528\u8fde\u901a\u5206\u91cf\u7b97\u6cd5\u5bf9\u8fd9\u4e9b\u91cd\u590d\u9879\u8fdb\u884c\u5206\u7ec4\uff0c\u5e76\u4ec5\u4fdd\u7559\u6bcf\u4e2a\u8fde\u901a\u5206\u91cf\u7684\u5355\u4e2a\u56fe\u50cf-\u6587\u672c\u5bf9\u3002\u6211\u4eec\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u63d0\u9ad8\u53bb\u91cd\u7ba1\u9053\u7684\u6548\u7387\uff1a(1) \u4f7f\u7528 k \u5747\u503c\u805a\u7c7b\u9884\u5148\u5bf9\u6570\u636e\u8fdb\u884c\u805a\u7c7b (2) \u5bf9 NN \u641c\u7d22\u548c\u805a\u7c7b\u4f7f\u7528 FAISS\uff08Johnson \u7b49\u4eba\uff0c2019\uff09\u3002<\/li>\n<li><strong>\u91cd\u91c7\u6837:<\/strong> \u6211\u4eec\u786e\u4fdd\u56fe\u50cf-\u6587\u672c\u5bf9\u7684\u591a\u6837\u6027\uff0c\u7c7b\u4f3c\u4e8e Xu \u7b49\u4eba\uff082023\uff09\uff1bMahajan \u7b49\u4eba\uff082018\uff09\uff1bMikolov \u7b49\u4eba\uff082013\uff09\u3002\u9996\u5148\uff0c\u6211\u4eec\u901a\u8fc7\u89e3\u6790\u9ad8\u8d28\u91cf\u7684\u6587\u672c\u6e90\u6784\u5efa\u4e00\u4e2a n \u5143\u8bed\u6cd5\u8bcd\u6c47\u8868\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u8ba1\u7b97\u6570\u636e\u96c6\u4e2d\u6bcf\u4e2a\u8bcd\u6c47\u8868 n \u5143\u8bed\u6cd5\u7684\u9891\u7387\u3002\u7136\u540e\uff0c\u6211\u4eec\u4ee5\u5982\u4e0b\u65b9\u5f0f\u5bf9\u6570\u636e\u8fdb\u884c\u91cd\u91c7\u6837\uff1a\u5982\u679c\u6807\u9898\u4e2d\u7684\u4efb\u4f55 n \u5143\u8bed\u6cd5\u5728\u8bcd\u6c47\u8868\u4e2d\u51fa\u73b0\u7684\u6b21\u6570\u5c11\u4e8e T \u6b21\uff0c\u6211\u4eec\u5c31\u4fdd\u7559\u76f8\u5e94\u7684\u56fe\u50cf-\u6587\u672c\u5bf9\u3002\u5426\u5219\uff0c\u6211\u4eec\u72ec\u7acb\u5730\u5bf9\u6807\u9898\u4e2d\u7684\u6bcf\u4e2a n \u5143\u8bed\u6cd5 n i \u4ee5\u6982\u7387 T \/ f i \u8fdb\u884c\u62bd\u6837\uff0c\u5176\u4e2d f i \u8868\u793a n \u5143\u8bed\u6cd5 n i \u7684\u9891\u7387\uff1b\u5982\u679c\u4efb\u4f55 n \u5143\u8bed\u6cd5\u90fd\u88ab\u62bd\u53d6\u5230\u4e86\uff0c\u6211\u4eec\u5c31\u4fdd\u7559\u56fe\u50cf-\u6587\u672c\u5bf9\u3002\u8fd9\u79cd\u91cd\u91c7\u6837\u6709\u52a9\u4e8e\u63d0\u9ad8\u4f4e\u9891\u7c7b\u522b\u548c\u7ec6\u7c92\u5ea6\u8bc6\u522b\u4efb\u52a1\u7684\u6027\u80fd\u3002<\/li>\n<li><strong>\u5149\u5b66\u5b57\u7b26\u8bc6\u522b:<\/strong>\u00a0\u6211\u4eec\u901a\u8fc7\u63d0\u53d6\u56fe\u50cf\u4e2d\u7684\u6587\u5b57\u5e76\u5c06\u5176\u4e0e\u6807\u9898\u4e32\u8054\u8d77\u6765\uff0c\u8fdb\u4e00\u6b65\u6539\u8fdb\u4e86\u6211\u4eec\u7684\u56fe\u50cf\u6587\u672c\u6570\u636e\u3002\u4f7f\u7528\u4e13\u6709\u5149\u5b66\u5b57\u7b26\u8bc6\u522b (OCR) \u7ba1\u9053\u63d0\u53d6\u4e66\u5199\u6587\u672c\u3002\u6211\u4eec\u89c2\u5bdf\u5230\u5c06 OCR \u6570\u636e\u6dfb\u52a0\u5230\u8bad\u7ec3\u6570\u636e\u4e2d\u53ef\u4ee5\u6781\u5927\u5730\u63d0\u9ad8\u9700\u8981 OCR \u529f\u80fd\u7684\u4efb\u52a1\u7684\u6027\u80fd\uff0c\u4f8b\u5982\u6587\u6863\u7406\u89e3\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p>\u4e3a\u4e86\u63d0\u9ad8\u6a21\u578b\u5728\u6587\u6863\u7406\u89e3\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\uff0c\u6211\u4eec\u5c06\u6587\u6863\u9875\u9762\u6e32\u67d3\u4e3a\u56fe\u50cf\uff0c\u5e76\u5c06\u56fe\u50cf\u4e0e\u5176\u5404\u81ea\u7684\u6587\u672c\u914d\u5bf9\u3002\u6587\u6863\u6587\u672c\u662f\u76f4\u63a5\u4ece\u6e90\u83b7\u53d6\u7684\uff0c\u6216\u8005\u901a\u8fc7\u6587\u6863\u89e3\u6790\u7ba1\u9053\u83b7\u53d6\u7684\u3002<\/p>\n<p><strong>\u5b89\u5168\uff1a<\/strong>\u00a0\u6211\u4eec\u4e3b\u8981\u81f4\u529b\u4e8e\u786e\u4fdd\u56fe\u50cf\u8bc6\u522b\u9884\u8bad\u7ec3\u6570\u636e\u96c6\u4e0d\u5305\u542b\u4e0d\u5b89\u5168\u5185\u5bb9\uff0c\u4f8b\u5982\u6027\u8650\u5f85\u6750\u6599 (CSAM) (Thiel, 2023)\u3002\u6211\u4eec\u4f7f\u7528\u611f\u77e5\u54c8\u5e0c\u65b9\u6cd5\uff0c\u5982 PhotoDNA (Farid, 2021)\uff0c\u4ee5\u53ca\u5185\u90e8\u4e13\u6709\u5206\u7c7b\u5668\uff0c\u626b\u63cf\u6240\u6709\u8bad\u7ec3\u56fe\u50cf\u4ee5\u67e5\u627e CSAM\u3002\u6211\u4eec\u8fd8\u4f7f\u7528\u4e13\u6709\u7684\u5a92\u4f53\u98ce\u9669\u68c0\u7d22\u7ba1\u9053\u6765\u8bc6\u522b\u548c\u5220\u9664\u6211\u4eec\u8ba4\u4e3a\u662f NSFW \u7684\u56fe\u50cf\u6587\u672c\u5bf9\uff0c\u4f8b\u5982\u56e0\u4e3a\u5b83\u4eec\u5305\u542b\u6027\u6216\u66b4\u529b\u5185\u5bb9\u3002\u6211\u4eec\u76f8\u4fe1\u6700\u5c0f\u5316\u8bad\u7ec3\u6570\u636e\u96c6\u4e2d\u7684\u6b64\u7c7b\u6750\u6599\u7684\u6d41\u884c\u5ea6\u53ef\u4ee5\u63d0\u9ad8\u6700\u7ec8\u6a21\u578b\u7684\u5b89\u5168\u6027\u548c\u5e2e\u52a9\u6027\uff0c\u800c\u4e0d\u4f1a\u5f71\u54cd\u5176\u6709\u7528\u6027\u3002\u6700\u540e\uff0c\u6211\u4eec\u5bf9\u8bad\u7ec3\u96c6\u4e2d\u7684\u6240\u6709\u56fe\u50cf\u6267\u884c\u9762\u90e8\u6a21\u7cca\u5904\u7406\u3002\u6211\u4eec\u6d4b\u8bd5\u4e86\u8be5\u6a21\u578b\u9488\u5bf9\u4eba\u7c7b\u751f\u6210\u7684\u5f15\u7528\u9644\u52a0\u56fe\u50cf\u7684\u63d0\u793a\u3002<\/p>\n<p><strong>\u9000\u706b\u6570\u636e\uff1a<\/strong>\u00a0\u6211\u4eec\u901a\u8fc7\u4f7f\u7528 n-gram \u5bf9\u56fe\u50cf\u6807\u9898\u5bf9\u8fdb\u884c\u91cd\u91c7\u6837\uff0c\u521b\u5efa\u4e86\u4e00\u4e2a\u5305\u542b\u7ea6 3.5 \u4ebf\u4e2a\u793a\u4f8b\u7684\u9000\u706b\u6570\u636e\u96c6\u3002\u7531\u4e8e n-gram \u91cd\u91c7\u6837\u6709\u5229\u4e8e\u66f4\u4e30\u5bcc\u7684\u6587\u672c\u63cf\u8ff0\uff0c\u56e0\u6b64\u5b83\u9009\u62e9\u4e86\u4e00\u4e2a\u66f4\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u5b50\u96c6\u3002\u6211\u4eec\u8fd8\u7528\u6765\u81ea\u4e94\u4e2a\u989d\u5916\u6765\u6e90\u7684\u7ea6 1.5 \u4ebf\u4e2a\u793a\u4f8b\u589e\u5f3a\u4e86\u7ed3\u679c\u6570\u636e\uff1a<\/p>\n<ul>\n<li style=\"list-style-type: none;\">\n<ul>\n<li><strong>\u89c6\u89c9\u5b9a\u4f4d:<\/strong>\u00a0\u6211\u4eec\u5c06\u6587\u672c\u4e2d\u7684\u540d\u8bcd\u77ed\u8bed\u4e0e\u56fe\u50cf\u4e2d\u7684\u8fb9\u754c\u6846\u6216\u8499\u677f\u5173\u8054\u8d77\u6765\u3002\u5b9a\u4f4d\u4fe1\u606f\uff08\u8fb9\u754c\u6846\u548c\u8499\u677f\uff09\u4ee5\u4e24\u79cd\u65b9\u5f0f\u5728\u56fe\u50cf\u6587\u672c\u5bf9\u4e2d\u6307\u5b9a\uff1a(1) \u6211\u4eec\u5728\u56fe\u50cf\u4e0a\u8986\u76d6\u6846\u6216\u8499\u677f\uff0c\u5e76\u5728\u6587\u672c\u4e2d\u4f7f\u7528\u6807\u8bb0\u4f5c\u4e3a\u53c2\u8003\uff0c\u7c7b\u4f3c\u4e8e\u6807\u8bb0\u96c6 (Yang et al., 2023a)\u3002(2) \u6211\u4eec\u76f4\u63a5\u5c06\u5f52\u4e00\u5316 (x min, y min, x max, y max) \u5750\u6807\u63d2\u5165\u6587\u672c\uff0c\u5e76\u7528\u7279\u6b8a\u6807\u8bb0\u5c06\u5176\u5206\u9694\u5f00\u3002<\/li>\n<li><strong>\u622a\u56fe\u89e3\u6790:<\/strong>\u00a0\u6211\u4eec\u4ece HTML \u4ee3\u7801\u4e2d\u6e32\u67d3\u622a\u56fe\uff0c\u5e76\u8ba9\u6a21\u578b\u9884\u6d4b\u751f\u6210\u7279\u5b9a\u622a\u56fe\u5143\u7d20\u7684\u4ee3\u7801\uff0c\u7c7b\u4f3c\u4e8e Lee \u7b49\u4eba (2023)\u3002\u611f\u5174\u8da3\u7684\u5143\u7d20\u901a\u8fc7\u8fb9\u754c\u6846\u5728\u622a\u56fe\u4e2d\u6307\u793a\u3002<\/li>\n<li><strong>\u95ee\u7b54\u5bf9:<\/strong>\u00a0\u6211\u4eec\u5305\u62ec\u95ee\u7b54\u5bf9\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u4f7f\u7528\u592a\u5927\u800c\u65e0\u6cd5\u7528\u4e8e\u6a21\u578b\u5fae\u8c03\u7684\u5927\u91cf\u95ee\u7b54\u6570\u636e\u3002<\/li>\n<li><strong>\u5408\u6210\u6807\u9898:<\/strong>\u00a0\u6211\u4eec\u5305\u62ec\u5e26\u6709\u7531\u65e9\u671f\u6a21\u578b\u7248\u672c\u751f\u6210\u7684\u5408\u6210\u6807\u9898\u7684\u56fe\u50cf\u3002\u4e0e\u539f\u59cb\u6807\u9898\u76f8\u6bd4\uff0c\u6211\u4eec\u53d1\u73b0\u5408\u6210\u6807\u9898\u63d0\u4f9b\u4e86\u6bd4\u539f\u59cb\u6807\u9898\u66f4\u5168\u9762\u7684\u56fe\u50cf\u63cf\u8ff0\u3002<\/li>\n<li><strong>\u5408\u6210\u7ed3\u6784\u5316\u56fe\u50cf\u3002<\/strong>\u00a0\u6211\u4eec\u8fd8\u5305\u542b\u4e86\u5404\u79cd\u9886\u57df\uff08\u4f8b\u5982\u56fe\u8868\u3001\u8868\u683c\u3001\u6d41\u7a0b\u56fe\u3001\u6570\u5b66\u516c\u5f0f\u548c\u6587\u672c\u6570\u636e\uff09\u7684\u5408\u6210\u751f\u6210\u56fe\u50cf\u3002\u8fd9\u4e9b\u56fe\u50cf\u90fd\u9644\u5e26\u76f8\u5e94\u7684\u7ed3\u6784\u5316\u8868\u793a\uff0c\u4f8b\u5982\u5bf9\u5e94\u7684 Markdown \u6216 LaTeX \u7b26\u53f7\u3002 \u9664\u4e86\u63d0\u9ad8\u6a21\u578b\u5728\u8fd9\u4e9b\u9886\u57df\u7684\u8bc6\u522b\u80fd\u529b\u5916\uff0c\u6211\u4eec\u53d1\u73b0\u8fd9\u4e9b\u6570\u636e\u5bf9\u4e8e\u901a\u8fc7\u6587\u672c\u6a21\u578b\u751f\u6210\u7528\u4e8e\u5fae\u8c03\u7684\u95ee\u7b54\u5bf9\u4e5f\u5f88\u6709\u7528\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<div id=\"attachment_3487\" style=\"width: 1330px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-3487\" class=\"wp-image-3487 size-full\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9723618173bd21d.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"1320\" height=\"680\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9723618173bd21d.png 1320w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9723618173bd21d-300x155.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9723618173bd21d-1024x528.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/9723618173bd21d-768x396.png 768w\" sizes=\"auto, (max-width: 1320px) 100vw, 1320px\" \/><p id=\"caption-attachment-3487\" class=\"wp-caption-text\">\u56fe 28 \u672c\u6587\u7814\u7a76\u7684\u5c06\u591a\u6a21\u6001\u80fd\u529b\u6dfb\u52a0\u5230 Llama 3 \u7684\u7ec4\u5408\u65b9\u6cd5\u793a\u610f\u56fe\u3002 \u8fd9\u79cd\u65b9\u6cd5\u5bfc\u81f4\u4e00\u4e2a\u591a\u6a21\u6001\u6a21\u578b\uff0c\u7ecf\u8fc7\u4e94\u4e2a\u9636\u6bb5\u8bad\u7ec3\uff1a\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u3001\u591a\u6a21\u6001\u7f16\u7801\u5668\u9884\u8bad\u7ec3\u3001\u89c6\u89c9\u9002\u914d\u5668\u8bad\u7ec3\u3001\u6a21\u578b\u5fae\u8c03\u548c\u8bed\u97f3\u9002\u914d\u5668\u8bad\u7ec3\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<h4>7.1.2 \u89c6\u9891\u6570\u636e<\/h4>\n<p>\u4e3a\u4e86\u8fdb\u884c\u89c6\u9891\u9884\u8bad\u7ec3\uff0c\u6211\u4eec\u4f7f\u7528\u4e86\u5927\u578b\u7684\u89c6\u9891-\u6587\u672c\u5bf9\u6570\u636e\u96c6\u3002\u6211\u4eec\u7684\u6570\u636e\u96c6\u901a\u8fc7\u591a\u9636\u6bb5\u8fc7\u7a0b\u8fdb\u884c\u6574\u7406\u3002 \u6211\u4eec\u4f7f\u7528\u57fa\u4e8e\u89c4\u5219\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\u7b5b\u9009\u548c\u6e05\u7406\u76f8\u5173\u7684\u6587\u672c\uff0c\u4f8b\u5982\u786e\u4fdd\u6700\u5c0f\u957f\u5ea6\u5e76\u4fee\u590d\u5927\u5199\u5b57\u6bcd\u3002\u7136\u540e\uff0c\u6211\u4eec\u8fd0\u884c\u8bed\u8a00\u8bc6\u522b\u6a21\u578b\u4ee5\u8fc7\u6ee4\u6389\u975e\u82f1\u8bed\u6587\u672c\u3002<\/p>\n<p>\u6211\u4eec\u8fd0\u884cOCR\u68c0\u6d4b\u6a21\u578b\u4ee5\u8fc7\u6ee4\u6389\u8fc7\u5ea6\u53e0\u52a0\u6587\u672c\u7684\u89c6\u9891\u3002\u4e3a\u4e86\u786e\u4fdd\u89c6\u9891-\u6587\u672c\u5bf9\u4e4b\u95f4\u7684\u5408\u7406\u5bf9\u9f50\uff0c\u6211\u4eec\u4f7f\u7528CLIP\uff08Radford\u7b49\uff0c2021\uff09\u98ce\u683c\u7684\u56fe\u50cf-\u6587\u672c\u548c\u89c6\u9891-\u6587\u672c\u5bf9\u6bd4\u6a21\u578b\u3002 \u6211\u4eec\u9996\u5148\u4f7f\u7528\u89c6\u9891\u4e2d\u7684\u5355\u5e27\u8ba1\u7b97\u56fe\u50cf-\u6587\u672c\u76f8\u4f3c\u5ea6\uff0c\u5e76\u8fc7\u6ee4\u6389\u76f8\u4f3c\u5ea6\u4f4e\u7684\u5bf9\uff0c\u7136\u540e\u540e\u7eed\u8fc7\u6ee4\u6389\u89c6\u9891-\u6587\u672c\u5bf9\u9f50\u8f83\u5dee\u7684\u5bf9\u3002\u6211\u4eec\u7684\u67d0\u4e9b\u6570\u636e\u5305\u542b\u9759\u6001\u6216\u4f4e\u8fd0\u52a8\u89c6\u9891\uff1b\u6211\u4eec\u4f7f\u7528\u57fa\u4e8e\u8fd0\u52a8\u8bc4\u5206\u7684\u8fc7\u6ee4\uff08Girdhar\u7b49\uff0c2023\uff09\u6765\u8fc7\u6ee4\u8fd9\u4e9b\u6570\u636e\u3002\u6211\u4eec\u6ca1\u6709\u5bf9\u89c6\u9891\u7684\u89c6\u89c9\u8d28\u91cf\u8fdb\u884c\u4efb\u4f55\u8fc7\u6ee4\uff0c\u4f8b\u5982\u7f8e\u5b66\u8bc4\u5206\u6216\u5206\u8fa8\u7387\u8fc7\u6ee4\u3002<\/p>\n<p>\u6211\u4eec\u7684\u6570\u636e\u96c6\u5305\u542b\u5e73\u5747\u6301\u7eed\u65f6\u95f4\u4e3a 21 \u79d2\u7684\u4e2d\u503c\u6301\u7eed\u65f6\u95f4\u4e3a 16 \u79d2\u7684\u89c6\u9891\uff0c\u8d85\u8fc799% \u7684\u89c6\u9891\u90fd\u5728\u4e00\u5206\u949f\u4ee5\u5185\u3002\u7a7a\u95f4\u5206\u8fa8\u7387\u5728 320p \u548c 4K \u89c6\u9891\u4e4b\u95f4\u53d8\u5316\u5f88\u5927\uff0c\u5176\u4e2d\u8d85\u8fc7 70% \u7684\u89c6\u9891\u7684\u77ed\u8fb9\u5927\u4e8e 720 \u50cf\u7d20\u3002\u89c6\u9891\u5177\u6709\u4e0d\u540c\u7684\u7eb5\u6a2a\u6bd4\uff0c\u51e0\u4e4e\u6240\u6709\u89c6\u9891\u7684\u7eb5\u6a2a\u6bd4\u90fd\u5728 1:2 \u548c 2:1 \u4e4b\u95f4\uff0c\u4e2d\u503c\u4e3a 1:1\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>7.2 \u6a21\u578b\u67b6\u6784<\/h3>\n<p>\u6211\u4eec\u7684\u89c6\u89c9\u8bc6\u522b\u6a21\u578b\u7531\u4e09\u4e2a\u4e3b\u8981\u7ec4\u4ef6\u7ec4\u6210\uff1a (1) \u56fe\u50cf\u7f16\u7801\u5668\uff0c(2) \u56fe\u50cf\u9002\u914d\u5668\uff0c\u4ee5\u53ca (3) \u89c6\u9891\u9002\u914d\u5668\u3002<\/p>\n<p><strong>\u56fe\u50cf\u7f16\u7801\u5668:<\/strong><\/p>\n<p>\u6211\u4eec\u7684\u56fe\u50cf\u7f16\u7801\u5668\u662f\u4e00\u4e2a\u6807\u51c6\u7684\u89c6\u89c9 Transformer\uff08ViT\uff1bDosovitskiy \u7b49\u4eba (2020))\uff0c\u5b83\u88ab\u8bad\u7ec3\u7528\u6765\u5bf9\u9f50\u56fe\u50cf\u548c\u6587\u672c (Xu \u7b49\u4eba, 2023)\u3002 \u6211\u4eec\u4f7f\u7528 ViT-H\/14 \u7248\u672c\u7684\u56fe\u50cf\u7f16\u7801\u5668\uff0c\u5b83\u62e5\u6709 6.3 \u4ebf\u4e2a\u53c2\u6570\uff0c\u5728 25 \u4ebf\u5f20\u56fe\u50cf-\u6587\u672c\u5bf9\u4e0a\u8bad\u7ec3\u4e86\u4e94\u4e2a epoch\u3002 \u56fe\u50cf\u7f16\u7801\u5668\u7684\u8f93\u5165\u56fe\u50cf\u5206\u8fa8\u7387\u4e3a 224 \u00d7 224\uff1b\u56fe\u50cf\u88ab\u62c6\u5206\u4e3a 16 \u00d7 16 \u4e2a\u5927\u5c0f\u76f8\u7b49\u7684\u5c0f\u5757\uff08\u5373 14\u00d714 \u50cf\u7d20\u7684\u5757\u5927\u5c0f\uff09\u3002\u6b63\u5982\u4e4b\u524d ViP-Llava (Cai \u7b49\u4eba, 2024) \u7b49\u5de5\u4f5c\u6240\u8868\u660e\u7684\u90a3\u6837\uff0c\u6211\u4eec\u53d1\u73b0\u901a\u8fc7\u5bf9\u6bd4\u6587\u672c\u5bf9\u9f50\u76ee\u6807\u8bad\u7ec3\u7684\u56fe\u50cf\u7f16\u7801\u5668\u65e0\u6cd5\u4fdd\u7559\u7cbe\u7ec6\u7684\u5b9a\u4f4d\u4fe1\u606f\u3002\u4e3a\u4e86\u7f13\u89e3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u591a\u5c42\u7279\u5f81\u63d0\u53d6\u65b9\u6cd5\uff0c\u9664\u4e86\u6700\u540e\u4e00\u5c42\u7279\u5f81\u5916\uff0c\u8fd8\u63d0\u4f9b\u4e86\u7b2c 4\u30018\u300116\u300124 \u548c 31 \u5c42\u7684\u7279\u5f81\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u5728\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u7684\u9884\u8bad\u7ec3\u4e4b\u524d\u8fdb\u4e00\u6b65\u63d2\u5165\u4e86 8 \u4e2a\u95e8\u63a7\u81ea\u6ce8\u610f\u529b\u5c42\uff08\u603b\u5171 40 \u4e2a Transformer \u5757\uff09\uff0c\u4ee5\u5b66\u4e60\u7279\u5b9a\u4e8e\u5bf9\u9f50\u7684\u7279\u5f81\u3002\u56e0\u6b64\uff0c\u56fe\u50cf\u7f16\u7801\u5668\u6700\u7ec8\u62e5\u6709 8.5 \u4ebf\u4e2a\u53c2\u6570\u548c\u989d\u5916\u7684\u5c42\u3002 \u901a\u8fc7\u591a\u5c42\u7279\u5f81\uff0c\u56fe\u50cf\u7f16\u7801\u5668\u4e3a\u751f\u6210\u7684 16 \u00d7 16 = 256 \u4e2a\u5c0f\u5757\u4e2d\u7684\u6bcf\u4e00\u4e2a\u4ea7\u751f\u4e00\u4e2a 7680 \u7ef4\u5ea6\u7684\u8868\u793a\u3002\u5728\u540e\u7eed\u8bad\u7ec3\u9636\u6bb5\u4e2d\uff0c\u6211\u4eec\u4e0d\u4f1a\u51bb\u7ed3\u56fe\u50cf\u7f16\u7801\u5668\u7684\u53c2\u6570\uff0c\u56e0\u4e3a\u6211\u4eec\u53d1\u73b0\u8fd9\u53ef\u4ee5\u63d0\u9ad8\u6027\u80fd\uff0c\u7279\u522b\u662f\u5728\u6587\u672c\u8bc6\u522b\u7b49\u9886\u57df\u3002<\/p>\n<p><strong>\u56fe\u50cf\u9002\u914d\u5668:<\/strong><\/p>\n<p>\u6211\u4eec\u5728\u56fe\u50cf\u7f16\u7801\u5668\u4ea7\u751f\u7684\u89c6\u89c9\u6807\u8bb0\u8868\u793a\u548c\u8bed\u8a00\u6a21\u578b\u4ea7\u751f\u7684\u6807\u8bb0\u8868\u793a\u4e4b\u95f4\u5f15\u5165\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42 (Alayrac \u7b49\u4eba, 2022)\u3002\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u5e94\u7528\u4e8e\u6838\u5fc3\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u6bcf\u4e2a\u7b2c\u56db\u4e2a\u81ea\u6ce8\u610f\u529b\u5c42\u4e4b\u540e\u3002 \u4e0e\u8bed\u8a00\u6a21\u578b\u672c\u8eab\u4e00\u6837\uff0c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u4f7f\u7528\u5e7f\u4e49\u67e5\u8be2\u6ce8\u610f\u529b (GQA) \u6765\u63d0\u9ad8\u6548\u7387\u3002<\/p>\n<p>\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u4e3a\u6a21\u578b\u5f15\u5165\u4e86\u5927\u91cf\u53ef\u8bad\u7ec3\u53c2\u6570\uff1a\u5bf9\u4e8e Llama 3 405B\uff0c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u7684\u53c2\u6570\u7ea6\u4e3a 1000 \u4ebf\u3002 \u6211\u4eec\u5bf9\u56fe\u50cf\u9002\u914d\u5668\u8fdb\u884c\u4e86\u4e24\u4e2a\u9636\u6bb5\u7684\u9884\u8bad\u7ec3: (1) \u521d\u59cb\u9884\u8bad\u7ec3\u548c (2) \u9000\u706b\uff1a *\u00a0<strong>\u521d\u59cb\u9884\u8bad\u7ec3:<\/strong>\u00a0\u6211\u4eec\u5728\u4e0a\u8ff0\u7ea6 60 \u4ebf\u5f20\u56fe\u50cf-\u6587\u672c\u5bf9\u7684\u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u6211\u4eec\u7684\u56fe\u50cf\u9002\u914d\u5668\u3002\u4e3a\u4e86\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\uff0c\u6211\u4eec\u5c06\u6240\u6709\u56fe\u50cf\u7684\u5927\u5c0f\u8c03\u6574\u4e3a\u6700\u591a\u9002\u5408\u56db\u4e2a 336 \u00d7 336 \u50cf\u7d20\u7684\u56fe\u5757\uff0c\u5176\u4e2d\u6211\u4eec\u5b89\u6392\u56fe\u5757\u4ee5\u652f\u6301\u4e0d\u540c\u7684\u5bbd\u9ad8\u6bd4\uff0c\u4f8b\u5982 672 \u00d7 672\u3001672 \u00d7 336 \u548c 1344 \u00d7 336\u3002 *\u00a0<strong>\u9000\u706b:<\/strong>\u00a0\u6211\u4eec\u7ee7\u7eed\u4f7f\u7528\u4e0a\u8ff0\u9000\u706b\u6570\u636e\u96c6\u4e2d\u7684\u7ea6 5 \u4ebf\u5f20\u56fe\u50cf\u6765\u8bad\u7ec3\u56fe\u50cf\u9002\u914d\u5668\u3002\u5728\u9000\u706b\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u589e\u52a0\u6bcf\u4e2a\u56fe\u5757\u7684\u56fe\u50cf\u5206\u8fa8\u7387\u4ee5\u63d0\u9ad8\u5bf9\u9700\u8981\u66f4\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684\u4efb\u52a1\u7684\u6027\u80fd\uff0c\u4f8b\u5982\u4fe1\u606f\u56fe\u8868\u7406\u89e3\u3002<\/p>\n<p><strong>\u89c6\u9891\u9002\u914d\u5668:<\/strong><\/p>\n<p>\u6211\u4eec\u7684\u6a21\u578b\u6700\u591a\u53ef\u4ee5\u63a5\u53d7 64 \u5e27\u7684\u8f93\u5165\uff08\u5747\u5300\u91c7\u6837\u81ea\u5b8c\u6574\u89c6\u9891\uff09\uff0c\u6bcf\u4e2a\u5e27\u90fd\u88ab\u56fe\u50cf\u7f16\u7801\u5668\u5904\u7406\u3002\u6211\u4eec\u901a\u8fc7\u4e24\u4e2a\u7ec4\u4ef6\u5bf9\u89c6\u9891\u4e2d\u7684\u65f6\u95f4\u7ed3\u6784\u8fdb\u884c\u5efa\u6a21: (i) \u7f16\u7801\u7684\u89c6\u9891\u5e27\u901a\u8fc7\u4e00\u4e2a\u65f6\u95f4\u805a\u5408\u5668\u5408\u5e76\u6210\u4e00\u4e2a\uff0c\u8be5\u805a\u5408\u5668\u5c06 32 \u4e2a\u8fde\u7eed\u7684\u5e27\u5408\u5e76\u4e3a\u4e00\u4e2a\uff1b(ii) \u5728\u6bcf\u4e2a\u7b2c\u56db\u4e2a\u56fe\u50cf\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u4e4b\u524d\u6dfb\u52a0\u989d\u5916\u7684\u89c6\u9891\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u3002\u65f6\u95f4\u805a\u5408\u5668\u88ab\u5b9e\u73b0\u4e3a\u611f\u77e5\u5668\u91cd\u91c7\u6837\u5668 (Jaegle \u7b49\u4eba, 2021\uff1bAlayrac \u7b49\u4eba, 2022)\u3002 \u6211\u4eec\u4f7f\u7528\u6bcf\u89c6\u9891 16 \u5e27\uff08\u805a\u5408\u6210 1 \u5e27\uff09\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u4f46\u5728\u76d1\u7763\u5fae\u8c03\u671f\u95f4\u5c06\u8f93\u5165\u5e27\u6570\u589e\u52a0\u5230 64\u3002 \u89c6\u9891\u805a\u5408\u5668\u548c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u5728 Llama 3 7B \u548c 70B \u4e2d\u5206\u522b\u6709 0.6 \u4ebf\u548c 4.6 \u4ebf\u4e2a\u53c2\u6570\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>7.3 \u6a21\u578b\u89c4\u6a21<\/h3>\n<p>\u5728\u5c06\u89c6\u89c9\u8bc6\u522b\u7ec4\u4ef6\u6dfb\u52a0\u5230 Llama 3 \u540e\uff0c\u6a21\u578b\u5305\u542b\u81ea\u6ce8\u610f\u529b\u5c42\u3001\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u548c\u4e00\u4e2a ViT \u56fe\u50cf\u7f16\u7801\u5668\u3002\u6211\u4eec\u53d1\u73b0\u5728\u8bad\u7ec3\u8f83\u5c0f\uff0880 \u4ebf\u548c 700 \u4ebf\u53c2\u6570\uff09\u6a21\u578b\u7684\u9002\u914d\u5668\u65f6\uff0c\u6570\u636e\u5e76\u884c\u548c\u5f20\u91cf\u5e76\u884c\u662f\u6700\u6709\u6548\u7684\u7ec4\u5408\u3002\u5728\u8fd9\u4e9b\u89c4\u6a21\u4e0b\uff0c\u6a21\u578b\u6216\u6d41\u6c34\u7ebf\u5e76\u884c\u4e0d\u4f1a\u63d0\u9ad8\u6548\u7387\uff0c\u56e0\u4e3a\u6536\u96c6\u6a21\u578b\u53c2\u6570\u5c06\u4e3b\u5bfc\u8ba1\u7b97\u3002\u7136\u800c\uff0c\u5728\u8bad\u7ec3 4050 \u4ebf\u53c2\u6570\u6a21\u578b\u7684\u9002\u914d\u5668\u65f6\uff0c\u6211\u4eec\u786e\u5b9e\u4f7f\u7528\u4e86\u6d41\u6c34\u7ebf\u5e76\u884c\uff08\u9664\u4e86\u6570\u636e\u548c\u5f20\u91cf\u5e76\u884c\uff09\u3002\u5728\u8fd9\u4e2a\u89c4\u6a21\u4e0a\u8fdb\u884c\u8bad\u7ec3\u5e26\u6765\u4e86\u4e09\u4e2a\u65b0\u7684\u6311\u6218\uff0c\u9664\u4e86\u7b2c 3.3 \u8282\u4e2d\u6982\u8ff0\u7684\u6311\u6218\u4e4b\u5916\uff1a\u6a21\u578b\u5f02\u8d28\u6027\u3001\u6570\u636e\u5f02\u8d28\u6027\u548c\u6570\u503c\u4e0d\u7a33\u5b9a\u6027\u3002<\/p>\n<p><strong>\u6a21\u578b\u5f02\u8d28\u6027<\/strong>\u3002\u6a21\u578b\u8ba1\u7b97\u662f\u5f02\u8d28\u7684\uff0c\u56e0\u4e3a\u67d0\u4e9b\u6807\u8bb0\u6bd4\u5176\u4ed6\u6807\u8bb0\u6267\u884c\u66f4\u591a\u7684\u8ba1\u7b97\u3002\u7279\u522b\u662f\uff0c\u56fe\u50cf\u6807\u8bb0\u901a\u8fc7\u56fe\u50cf\u7f16\u7801\u5668\u548c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u8fdb\u884c\u5904\u7406\uff0c\u800c\u6587\u672c\u6807\u8bb0\u4ec5\u901a\u8fc7\u8bed\u8a00\u9aa8\u5e72\u7f51\u7edc\u8fdb\u884c\u5904\u7406\u3002\u8fd9\u79cd\u5f02\u8d28\u6027\u4f1a\u5bfc\u81f4\u6d41\u6c34\u7ebf\u5e76\u884c\u8c03\u5ea6\u4e2d\u7684\u74f6\u9888\u3002\u6211\u4eec\u901a\u8fc7\u786e\u4fdd\u6bcf\u4e2a\u6d41\u6c34\u7ebf\u9636\u6bb5\u5305\u542b\u4e94\u4e2a\u5c42\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff1a\u5373\u8bed\u8a00\u9aa8\u5e72\u7f51\u7edc\u4e2d\u7684\u56db\u4e2a\u81ea\u6ce8\u610f\u529b\u5c42\u548c\u4e00\u4e2a\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u3002\uff08\u56de\u60f3\u4e00\u4e0b\uff0c\u6211\u4eec\u5728\u6bcf\u56db\u4e2a\u81ea\u6ce8\u610f\u529b\u5c42\u540e\u5f15\u5165\u4e86\u4e00\u4e2a\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u3002\uff09\u6b64\u5916\uff0c\u6211\u4eec\u5c06\u56fe\u50cf\u7f16\u7801\u5668\u590d\u5236\u5230\u6240\u6709\u6d41\u6c34\u7ebf\u9636\u6bb5\u3002\u7531\u4e8e\u6211\u4eec\u8bad\u7ec3\u7684\u662f\u914d\u5bf9\u7684\u56fe\u50cf\u6587\u672c\u6570\u636e\uff0c\u8fd9\u4f7f\u6211\u4eec\u80fd\u591f\u5728\u8ba1\u7b97\u7684\u56fe\u50cf\u548c\u6587\u672c\u90e8\u5206\u4e4b\u95f4\u8fdb\u884c\u8d1f\u8f7d\u5e73\u8861\u3002<\/p>\n<p><strong>\u6570\u636e\u5f02\u8d28\u6027<\/strong>\u3002\u6570\u636e\u662f\u5f02\u8d28\u7684\uff0c\u56e0\u4e3a\u5e73\u5747\u800c\u8a00\uff0c\u56fe\u50cf\u6bd4\u5173\u8054\u6587\u672c\u5177\u6709\u66f4\u591a\u6807\u8bb0\uff1a\u4e00\u5f20\u56fe\u50cf\u6709 2308 \u4e2a\u6807\u8bb0\uff0c\u800c\u5173\u8054\u6587\u672c\u5e73\u5747\u53ea\u6709 192 \u4e2a\u6807\u8bb0\u3002\u56e0\u6b64\uff0c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u7684\u8ba1\u7b97\u9700\u8981\u66f4\u957f\u7684\u65f6\u95f4\u548c\u66f4\u591a\u7684\u5185\u5b58\uff0c\u800c\u4e0d\u662f\u81ea\u6ce8\u610f\u529b\u5c42\u7684\u8ba1\u7b97\u3002\u6211\u4eec\u901a\u8fc7\u5728\u56fe\u50cf\u7f16\u7801\u5668\u4e2d\u5f15\u5165\u5e8f\u5217\u5e76\u884c\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u8fd9\u6837\u6bcf\u4e2a GPU \u5904\u7406\u7684\u6807\u8bb0\u6570\u91cf\u5927\u81f4\u76f8\u540c\u3002\u7531\u4e8e\u5e73\u5747\u6587\u672c\u5c3a\u5bf8\u76f8\u5bf9\u8f83\u5c0f\uff0c\u6211\u4eec\u8fd8\u4f7f\u7528\u4e86\u66f4\u5927\u7684\u5fae\u6279\u91cf\u5927\u5c0f\uff088 \u800c\u4e0d\u662f 1\uff09\u3002<\/p>\n<p><strong>\u6570\u503c\u4e0d\u7a33\u5b9a\u6027<\/strong>\u3002\u5728\u5c06\u56fe\u50cf\u7f16\u7801\u5668\u6dfb\u52a0\u5230\u6a21\u578b\u540e\uff0c\u6211\u4eec\u53d1\u73b0\u4f7f\u7528 bf16 \u8fdb\u884c\u68af\u5ea6\u7d2f\u79ef\u4f1a\u5bfc\u81f4\u6570\u503c\u4e0d\u7a33\u5b9a\u3002\u6700\u53ef\u80fd\u7684\u89e3\u91ca\u662f\u56fe\u50cf\u6807\u8bb0\u901a\u8fc7\u6240\u6709\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u5f15\u5165\u5230\u8bed\u8a00\u9aa8\u5e72\u7f51\u7edc\u4e2d\u3002\u8fd9\u610f\u5473\u7740\u56fe\u50cf\u6807\u8bb0\u8868\u793a\u4e2d\u7684\u6570\u503c\u504f\u5dee\u5bf9\u6574\u4f53\u8ba1\u7b97\u7684\u5f71\u54cd\u8fc7\u5927\uff0c\u56e0\u4e3a\u9519\u8bef\u88ab\u590d\u5408\u8d77\u6765\u3002\u6211\u4eec\u901a\u8fc7\u4f7f\u7528 FP32 \u6267\u884c\u68af\u5ea6\u7d2f\u79ef\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>7.4 \u9884\u8bad\u7ec3<\/h3>\n<p><strong>\u56fe\u50cf\u9884\u8bad\u7ec3:<\/strong>\u00a0\u6211\u4eec\u4ece\u9884\u8bad\u7ec3\u7684\u6587\u672c\u6a21\u578b\u548c\u89c6\u89c9\u7f16\u7801\u5668\u6743\u91cd\u5f00\u59cb\u521d\u59cb\u5316\u3002\u89c6\u89c9\u7f16\u7801\u5668\u88ab\u89e3\u51bb\uff0c\u800c\u6587\u672c\u6a21\u578b\u6743\u91cd\u4fdd\u6301\u51bb\u7ed3\uff0c\u5982\u4e0a\u6240\u8ff0\u3002\u9996\u5148\uff0c\u6211\u4eec\u4f7f\u7528 60 \u4ebf\u4e2a\u56fe\u50cf-\u6587\u672c\u5bf9\u8bad\u7ec3\u6a21\u578b\uff0c\u6bcf\u4e2a\u56fe\u50cf\u90fd\u88ab\u8c03\u6574\u5927\u5c0f\u4ee5\u9002\u5408\u56db\u4e2a 336 \u00d7 336 \u50cf\u7d20\u7684\u56fe\u5757\u3002\u6211\u4eec\u4f7f\u7528 16,384 \u7684\u5168\u5c40\u6279\u91cf\u5927\u5c0f\u548c\u4f59\u5f26\u5b66\u4e60\u7387\u8ba1\u5212\uff0c\u521d\u59cb\u5b66\u4e60\u7387\u4e3a 10 \u00d7 10 -4\uff0c\u6743\u91cd\u8870\u51cf\u4e3a 0.01\u3002\u521d\u59cb\u5b66\u4e60\u7387\u662f\u57fa\u4e8e\u5c0f\u89c4\u6a21\u5b9e\u9a8c\u786e\u5b9a\u7684\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u53d1\u73b0\u5e76\u4e0d\u80fd\u5f88\u597d\u5730\u63a8\u5e7f\u5230\u975e\u5e38\u957f\u7684\u8bad\u7ec3\u8ba1\u5212\u4e2d\uff0c\u5e76\u4e14\u5728\u635f\u5931\u503c\u505c\u6ede\u65f6\uff0c\u6211\u4eec\u4f1a\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u591a\u6b21\u964d\u4f4e\u5b66\u4e60\u7387\u3002\u5728\u57fa\u672c\u9884\u8bad\u7ec3\u4e4b\u540e\uff0c\u6211\u4eec\u5c06\u56fe\u50cf\u5206\u8fa8\u7387\u8fdb\u4e00\u6b65\u63d0\u9ad8\uff0c\u5e76\u5728\u9000\u706b\u6570\u636e\u96c6\u4e0a\u7ee7\u7eed\u8bad\u7ec3\u76f8\u540c\u7684\u6743\u91cd\u3002\u4f18\u5316\u5668\u901a\u8fc7\u70ed\u8eab\u91cd\u65b0\u521d\u59cb\u5316\u5230\u5b66\u4e60\u7387 2 \u00d7 10 -5\uff0c\u518d\u6b21\u9075\u5faa\u4f59\u5f26\u8ba1\u5212\u3002<\/p>\n<p><strong>\u89c6\u9891\u9884\u8bad\u7ec3:<\/strong>\u00a0\u5bf9\u4e8e\u89c6\u9891\u9884\u8bad\u7ec3\uff0c\u6211\u4eec\u4ece\u4e0a\u8ff0\u56fe\u50cf\u9884\u8bad\u7ec3\u548c\u9000\u706b\u6743\u91cd\u5f00\u59cb\u3002\u6211\u4eec\u5c06\u6dfb\u52a0\u89c6\u9891\u805a\u5408\u5668\u548c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\uff0c\u5982\u67b6\u6784\u4e2d\u6240\u8ff0\uff0c\u5e76\u968f\u673a\u521d\u59cb\u5316\u5b83\u4eec\u3002\u6211\u4eec\u51bb\u7ed3\u6a21\u578b\u4e2d\u7684\u6240\u6709\u53c2\u6570\uff0c\u9664\u4e86\u7279\u5b9a\u4e8e\u89c6\u9891\u7684\u53c2\u6570\uff08\u805a\u5408\u5668\u548c\u89c6\u9891\u4ea4\u53c9\u6ce8\u610f\u529b\uff09\uff0c\u5e76\u5728\u89c6\u9891\u9884\u8bad\u7ec3\u6570\u636e\u4e0a\u8bad\u7ec3\u5b83\u4eec\u3002\u6211\u4eec\u4f7f\u7528\u4e0e\u56fe\u50cf\u9000\u706b\u9636\u6bb5\u76f8\u540c\u7684\u8bad\u7ec3\u8d85\u53c2\u6570\uff0c\u5b66\u4e60\u7387\u7565\u6709\u4e0d\u540c\u3002\u6211\u4eec\u4ece\u5b8c\u6574\u7684\u89c6\u9891\u4e2d\u5747\u5300\u91c7\u6837 16 \u5e27\uff0c\u5e76\u4f7f\u7528\u56db\u4e2a\u5927\u5c0f\u4e3a 448 \u00d7 448 \u50cf\u7d20\u7684\u5757\u6765\u8868\u793a\u6bcf\u4e00\u5e27\u3002\u6211\u4eec\u5728\u89c6\u9891\u805a\u5408\u5668\u4e2d\u4f7f\u7528 16 \u7684\u805a\u5408\u56e0\u5b50\uff0c\u4ece\u800c\u5f97\u5230\u4e00\u4e2a\u6709\u6548\u7684\u5e27\uff0c\u6587\u672c\u6807\u8bb0\u5c06\u4ea4\u53c9\u5173\u6ce8\u8be5\u5e27\u3002\u6211\u4eec\u4f7f\u7528 4,096 \u7684\u5168\u5c40\u6279\u91cf\u5927\u5c0f\u3001190 \u4e2a\u6807\u8bb0\u7684\u5e8f\u5217\u957f\u5ea6\u4ee5\u53ca 10 -4 \u7684\u5b66\u4e60\u7387\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>7.5 \u8bad\u7ec3\u540e\u5904\u7406<\/h3>\n<p>\u5728\u8fd9\u4e00\u7ae0\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u8be6\u7ec6\u63cf\u8ff0\u89c6\u89c9\u9002\u914d\u5668\u7684\u540e\u7eed\u8bad\u7ec3\u6b65\u9aa4\u3002<\/p>\n<p>\u5728\u9884\u8bad\u7ec3\u4e4b\u540e\uff0c\u6211\u4eec\u5c06\u6a21\u578b\u5fae\u8c03\u5230\u9ad8\u5ea6\u7cbe\u9009\u7684\u591a\u6a21\u6001\u5bf9\u8bdd\u6570\u636e\u4e0a\uff0c\u4ee5\u542f\u7528\u804a\u5929\u529f\u80fd\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5b9e\u73b0\u4e86\u76f4\u63a5\u504f\u597d\u4f18\u5316 (DPO) \u4ee5\u63d0\u9ad8\u4eba\u5de5\u8bc4\u4f30\u6027\u80fd\uff0c\u5e76\u91c7\u7528\u4e86\u62d2\u7edd\u91c7\u6837\u6765\u6539\u5584\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p>\u6700\u540e\uff0c\u6211\u4eec\u6dfb\u52a0\u4e86\u4e00\u4e2a\u8d28\u91cf\u8c03\u6574\u9636\u6bb5\uff0c\u5728\u8fd9\u4e2a\u9636\u6bb5\u6211\u4eec\u5c06\u7ee7\u7eed\u5728\u975e\u5e38\u5c0f\u7684\u9ad8\u8d28\u91cf\u5bf9\u8bdd\u6570\u636e\u96c6\u4e0a\u5fae\u8c03\u6a21\u578b\uff0c\u8fd9\u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86\u4eba\u5de5\u8bc4\u4f30\u7ed3\u679c\uff0c\u540c\u65f6\u4fdd\u7559\u4e86\u57fa\u51c6\u6d4b\u8bd5\u7684\u6027\u80fd\u3002<\/p>\n<p>\u4ee5\u4e0b\u5c06\u63d0\u4f9b\u6bcf\u4e2a\u6b65\u9aa4\u7684\u8be6\u7ec6\u4fe1\u606f\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>7.5.1 \u76d1\u7763\u5fae\u8c03\u6570\u636e<\/h4>\n<p>\u6211\u4eec\u5728\u4e0b\u9762\u5206\u522b\u63cf\u8ff0\u7528\u4e8e\u56fe\u50cf\u548c\u89c6\u9891\u529f\u80fd\u7684\u76d1\u7763\u5fae\u8c03 (SFT) \u6570\u636e\u3002<\/p>\n<p><strong>\u56fe\u50cf:<\/strong>\u00a0\u6211\u4eec\u4f7f\u7528\u4e0d\u540c\u6570\u636e\u96c6\u7684\u6df7\u5408\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\u3002<\/p>\n<ul>\n<li>\u5b66\u672f\u6570\u636e\u96c6\uff1a\u6211\u4eec\u4f7f\u7528\u6a21\u677f\u6216\u901a\u8fc7\u5927\u8bed\u8a00\u6a21\u578b (LLM) \u91cd\u5199\uff0c\u5c06\u9ad8\u5ea6\u8fc7\u6ee4\u7684\u73b0\u6709\u5b66\u672f\u6570\u636e\u96c6\u8f6c\u6362\u4e3a\u95ee\u7b54\u5bf9\u3002LLM \u91cd\u5199\u7684\u76ee\u7684\u662f\u7528\u4e0d\u540c\u7684\u6307\u4ee4\u6765\u6269\u5145\u6570\u636e\uff0c\u5e76\u63d0\u9ad8\u7b54\u6848\u7684\u8bed\u8a00\u8d28\u91cf\u3002<\/li>\n<li>\u4eba\u5de5\u6807\u6ce8\uff1a\u6211\u4eec\u901a\u8fc7\u4eba\u5de5\u6807\u6ce8\u5458\u4e3a\u5404\u79cd\u4efb\u52a1\uff08\u5f00\u653e\u5f0f\u95ee\u7b54\u3001\u5b57\u5e55\u3001\u5b9e\u9645\u7528\u4f8b\u7b49\uff09\u548c\u9886\u57df\uff08\u4f8b\u5982\uff0c\u81ea\u7136\u56fe\u50cf\u548c\u7ed3\u6784\u5316\u56fe\u50cf\uff09\u6536\u96c6\u591a\u6a21\u6001\u5bf9\u8bdd\u6570\u636e\u3002\u6807\u6ce8\u5458\u5c06\u6536\u5230\u56fe\u50cf\u5e76\u88ab\u8981\u6c42\u7f16\u5199\u5bf9\u8bdd\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u786e\u4fdd\u591a\u6837\u6027\uff0c\u6211\u4eec\u5c06\u5927\u89c4\u6a21\u6570\u636e\u96c6\u8fdb\u884c\u805a\u7c7b\uff0c\u5e76\u5728\u4e0d\u540c\u7684\u805a\u7c7b\u4e2d\u5747\u5300\u5730\u91c7\u6837\u56fe\u50cf\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u901a\u8fc7\u4f7f\u7528 k \u8fd1\u90bb\u6269\u5c55\u79cd\u5b50\u6765\u83b7\u53d6\u4e00\u4e9b\u7279\u5b9a\u9886\u57df\u7684\u9644\u52a0\u56fe\u50cf\u3002\u6807\u6ce8\u5458\u8fd8\u5c06\u83b7\u5f97\u73b0\u6709\u6a21\u578b\u7684\u4e2d\u95f4\u68c0\u67e5\u70b9\uff0c\u4ee5\u4fc3\u8fdb\u6a21\u578b\u5728\u5faa\u73af\u4e2d\u7684\u6837\u5f0f\u6807\u6ce8\uff0c\u4ee5\u4fbf\u53ef\u4ee5\u4f7f\u7528\u6a21\u578b\u751f\u6210\u4f5c\u4e3a\u6807\u6ce8\u5458\u63d0\u4f9b\u989d\u5916\u4eba\u7c7b\u7f16\u8f91\u7684\u8d77\u70b9\u3002\u8fd9\u662f\u4e00\u4e2a\u8fed\u4ee3\u8fc7\u7a0b\uff0c\u5176\u4e2d\u6a21\u578b\u68c0\u67e5\u70b9\u5c06\u5b9a\u671f\u66f4\u65b0\u4e3a\u6027\u80fd\u66f4\u597d\u7684\u7248\u672c\uff0c\u8fd9\u4e9b\u7248\u672c\u662f\u5728\u6700\u65b0\u6570\u636e\u4e0a\u8bad\u7ec3\u7684\u3002\u8fd9\u589e\u52a0\u4e86\u4eba\u5de5\u6807\u6ce8\u7684\u91cf\u548c\u6548\u7387\uff0c\u540c\u65f6\u4e5f\u63d0\u9ad8\u4e86\u8d28\u91cf\u3002<\/p>\n<ul>\n<li>\u5408\u6210\u6570\u636e\uff1a\u6211\u4eec\u63a2\u7d22\u4e86\u901a\u8fc7\u4f7f\u7528\u56fe\u50cf\u7684\u6587\u672c\u8868\u793a\u548c\u6587\u672c\u8f93\u5165 LLM \u751f\u6210\u5408\u6210\u591a\u6a21\u6001\u6570\u636e\u7684\u4e0d\u540c\u65b9\u6cd5\u3002\u57fa\u672c\u601d\u8def\u662f\u5229\u7528\u6587\u672c\u8f93\u5165 LLM \u7684\u63a8\u7406\u80fd\u529b\u5728\u6587\u672c\u57df\u4e2d\u751f\u6210\u95ee\u7b54\u5bf9\uff0c\u5e76\u7528\u5176\u5bf9\u5e94\u7684\u56fe\u50cf\u66ff\u6362\u6587\u672c\u8868\u793a\u4ee5\u4ea7\u751f\u5408\u6210\u7684\u591a\u6a21\u6001\u6570\u636e\u3002\u793a\u4f8b\u5305\u62ec\u5c06\u6765\u81ea\u95ee\u7b54\u6570\u636e\u96c6\u7684\u6587\u672c\u5448\u73b0\u4e3a\u56fe\u50cf\u6216\u5c06\u8868\u683c\u6570\u636e\u5448\u73b0\u4e3a\u5408\u6210\u8868\u683c\u548c\u56fe\u8868\u56fe\u50cf\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u4f7f\u7528\u73b0\u6709\u56fe\u50cf\u7684\u5b57\u5e55\u548c OCR \u63d0\u53d6\u6765\u751f\u6210\u4e0e\u56fe\u50cf\u76f8\u5173\u7684\u4e00\u822c\u5bf9\u8bdd\u6216\u95ee\u7b54\u6570\u636e\u3002<\/li>\n<\/ul>\n<p><strong>\u89c6\u9891:<\/strong>\u00a0\u4e0e\u56fe\u50cf\u9002\u914d\u5668\u7c7b\u4f3c\uff0c\u6211\u4eec\u5c06\u9884\u5148\u5b58\u5728\u6ce8\u91ca\u7684\u5b66\u672f\u6570\u636e\u96c6\u7528\u4e8e\u8f6c\u6362\uff0c\u5c06\u5176\u8f6c\u6362\u4e3a\u9002\u5f53\u7684\u6587\u672c\u6307\u4ee4\u548c\u76ee\u6807\u54cd\u5e94\u3002\u76ee\u6807\u5c06\u8f6c\u6362\u4e3a\u5f00\u653e\u5f0f\u56de\u7b54\u6216\u591a\u9879\u9009\u62e9\u9898\uff0c\u89c6\u60c5\u51b5\u800c\u5b9a\u3002\u6211\u4eec\u8bf7\u4eba\u5de5\u6807\u6ce8\u5458\u4e3a\u89c6\u9891\u6dfb\u52a0\u95ee\u9898\u548c\u76f8\u5e94\u7684\u7b54\u6848\u3002\u6211\u4eec\u8981\u6c42\u6807\u6ce8\u5458\u5173\u6ce8\u90a3\u4e9b\u4e0d\u80fd\u57fa\u4e8e\u5355\u4e2a\u5e27\u6765\u56de\u7b54\u7684\u95ee\u9898\uff0c\u4ee5\u5f15\u5bfc\u6807\u6ce8\u5458\u671d\u7740\u9700\u8981\u65f6\u95f4\u7406\u89e3\u7684\u95ee\u9898\u65b9\u5411\u524d\u8fdb\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>7.5.2 \u76d1\u7763\u5fae\u8c03\u65b9\u6848<\/h4>\n<p>\u6211\u4eec\u5206\u522b\u4ecb\u7ecd\u4e86\u56fe\u50cf\u548c\u89c6\u9891\u80fd\u529b\u7684\u53d7\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u65b9\u6848\uff1a<\/p>\n<p><strong>\u56fe\u50cf:<\/strong>\u00a0\u6211\u4eec\u4ece\u9884\u8bad\u7ec3\u7684\u56fe\u50cf\u9002\u914d\u5668\u521d\u59cb\u5316\u6a21\u578b\uff0c\u4f46\u5c06\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u6743\u91cd\u66ff\u6362\u4e3a\u6307\u4ee4\u8c03\u4f18\u8bed\u8a00\u6a21\u578b\u7684\u6743\u91cd\u3002\u4e3a\u4e86\u4fdd\u6301\u4ec5\u6587\u672c\u6027\u80fd\uff0c\u8bed\u8a00\u6a21\u578b\u6743\u91cd\u4fdd\u6301\u51bb\u7ed3\uff0c\u5373\u6211\u4eec\u53ea\u4f1a\u66f4\u65b0\u89c6\u89c9\u7f16\u7801\u5668\u548c\u56fe\u50cf\u9002\u914d\u5668\u6743\u91cd\u3002<\/p>\n<p>\u6211\u4eec\u7684\u5fae\u8c03\u65b9\u6cd5\u7c7b\u4f3c\u4e8e Wortsman \u7b49\u4eba\uff082022\uff09\u3002\u9996\u5148\uff0c\u6211\u4eec\u4f7f\u7528\u591a\u4e2a\u6570\u636e\u968f\u673a\u5b50\u96c6\u3001\u5b66\u4e60\u7387\u548c\u91cd\u91cf\u8870\u51cf\u503c\u8fdb\u884c\u8d85\u53c2\u6570\u626b\u63cf\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u6839\u636e\u6a21\u578b\u6027\u80fd\u5bf9\u5b83\u4eec\u8fdb\u884c\u6392\u540d\u3002\u6700\u540e\uff0c\u6211\u4eec\u5c06\u524d K \u4e2a\u6a21\u578b\u7684\u6743\u91cd\u53d6\u5e73\u5747\u4ee5\u83b7\u5f97\u6700\u7ec8\u6a21\u578b\u3002K \u7684\u503c\u901a\u8fc7\u8bc4\u4f30\u5e73\u5747\u6a21\u578b\u5e76\u9009\u62e9\u6027\u80fd\u6700\u9ad8\u5b9e\u4f8b\u6765\u786e\u5b9a\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u4e0e\u901a\u8fc7\u7f51\u683c\u641c\u7d22\u627e\u5230\u7684\u6700\u4f73\u5355\u4e2a\u6a21\u578b\u76f8\u6bd4\uff0c\u5e73\u5747\u6a21\u578b\u59cb\u7ec8\u4ea7\u751f\u66f4\u597d\u7684\u7ed3\u679c\u3002\u6b64\u5916\uff0c\u8fd9\u79cd\u7b56\u7565\u51cf\u5c11\u4e86\u5bf9\u8d85\u53c2\u6570\u7684\u654f\u611f\u6027\u3002<\/p>\n<p><strong>\u89c6\u9891:<\/strong>\u00a0\u5bf9\u4e8e\u89c6\u9891 SFT\uff0c\u6211\u4eec\u4f7f\u7528\u9884\u8bad\u7ec3\u6743\u91cd\u521d\u59cb\u5316\u89c6\u9891\u805a\u5408\u5668\u548c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u3002\u6a21\u578b\u7684\u5176\u4f59\u53c2\u6570\uff08\u56fe\u50cf\u6743\u91cd\u548c LLM\uff09\u5219\u4ece\u76f8\u5e94\u7684\u6a21\u578b\u4e2d\u521d\u59cb\u5316\uff0c\u5e76\u6309\u7167\u5b83\u4eec\u7684\u5fae\u8c03\u9636\u6bb5\u8fdb\u884c\u3002\u7c7b\u4f3c\u4e8e\u89c6\u9891\u9884\u8bad\u7ec3\uff0c\u7136\u540e\u53ea\u5bf9\u89c6\u9891 SFT \u6570\u636e\u4e0a\u7684\u89c6\u9891\u53c2\u6570\u8fdb\u884c\u5fae\u8c03\u3002\u5728\u6b64\u9636\u6bb5\uff0c\u6211\u4eec\u5c06\u89c6\u9891\u957f\u5ea6\u589e\u52a0\u5230 64 \u5e27\uff0c\u5e76\u4f7f\u7528 32 \u7684\u805a\u5408\u56e0\u5b50\u83b7\u5f97\u4e24\u4e2a\u6709\u6548\u5e27\u3002 ayn\u0131 zamanda,\u5757\u7684\u5206\u8fa8\u7387\u4e5f\u76f8\u5e94\u63d0\u9ad8\uff0c\u4ee5\u4e0e\u5bf9\u5e94\u7684\u56fe\u50cf\u8d85\u53c2\u6570\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>7.5.3 \u504f\u597d\u6570\u636e<\/h4>\n<p>\u4e3a\u4e86\u5956\u52b1\u5efa\u6a21\u548c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u591a\u6a21\u6001\u6210\u5bf9\u504f\u597d\u6570\u636e\u96c6\u3002<\/p>\n<ul>\n<li><strong>\u4eba\u5de5\u6807\u6ce8:<\/strong>\u00a0\u4eba\u5de5\u6807\u6ce8\u7684\u504f\u597d\u6570\u636e\u5305\u542b\u4e24\u4e2a\u4e0d\u540c\u6a21\u578b\u8f93\u51fa\u7684\u6bd4\u8f83\uff0c\u6807\u8bb0\u4e3a\u201c\u9009\u62e9\u201d\u548c\u201c\u62d2\u7edd\u201d\uff0c\u5e76\u4f7f\u7528 7 \u7ea7\u8bc4\u5206\u3002\u7528\u4e8e\u751f\u6210\u54cd\u5e94\u7684\u6a21\u578b\u662f\u6bcf\u5468\u4ece\u4e00\u6c60\u6700\u4f73\u8fd1\u671f\u6a21\u578b\u4e2d\u968f\u673a\u91c7\u6837\u7684\uff0c\u6bcf\u4e2a\u6a21\u578b\u90fd\u6709\u4e0d\u540c\u7684\u7279\u5f81\u3002\u9664\u4e86\u504f\u597d\u6807\u7b7e\u5916\uff0c\u6211\u4eec\u8fd8\u8981\u6c42\u6807\u6ce8\u5458\u63d0\u4f9b\u53ef\u9009\u7684\u4eba\u5de5\u7f16\u8f91\u4ee5\u66f4\u6b63\u201c\u9009\u62e9\u201d\u54cd\u5e94\u4e2d\u7684\u4e0d\u51c6\u786e\u6027\uff0c\u56e0\u4e3a\u89c6\u89c9\u4efb\u52a1\u5bf9\u4e0d\u51c6\u786e\u6027\u7684\u5bb9\u5fcd\u5ea6\u8f83\u4f4e\u3002\u8bf7\u6ce8\u610f\uff0c\u4eba\u5de5\u7f16\u8f91\u662f\u4e00\u4e2a\u53ef\u9009\u6b65\u9aa4\uff0c\u56e0\u4e3a\u5b9e\u8df5\u4e2d\u5b58\u5728\u6570\u91cf\u548c\u8d28\u91cf\u4e4b\u95f4\u7684\u6743\u8861\u3002<\/li>\n<li><strong>\u5408\u6210\u6570\u636e:<\/strong>\u00a0\u901a\u8fc7\u4f7f\u7528\u4ec5\u6587\u672c\u7684 LLM \u7f16\u8f91\u5e76\u6545\u610f\u5728\u76d1\u7763\u5fae\u8c03\u6570\u636e\u96c6\u5f15\u5165\u9519\u8bef\u4e5f\u53ef\u4ee5\u751f\u6210\u5408\u6210\u504f\u597d\u5bf9\u3002\u6211\u4eec\u53d6\u5bf9\u8bdd\u6570\u636e\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u4f7f\u7528 LLM \u5f15\u5165\u5fae\u5999\u4f46\u6709\u610f\u4e49\u7684\u9519\u8bef\uff08\u4f8b\u5982\uff0c\u66f4\u6539\u5bf9\u8c61\u3001\u66f4\u6539\u5c5e\u6027\u3001\u6dfb\u52a0\u8ba1\u7b97\u9519\u8bef\u7b49\uff09\u3002\u8fd9\u4e9b\u7f16\u8f91\u540e\u7684\u54cd\u5e94\u88ab\u7528\u4f5c\u8d1f\u9762\u201c\u62d2\u7edd\u201d\u6837\u672c\uff0c\u5e76\u4e0e\u201c\u9009\u62e9\u201d\u7684\u539f\u59cb\u76d1\u7763\u5fae\u8c03\u6570\u636e\u914d\u5bf9\u3002<\/li>\n<li><strong>\u62d2\u7edd\u91c7\u6837:<\/strong>\u00a0\u6b64\u5916\uff0c\u4e3a\u4e86\u521b\u5efa\u66f4\u591a\u7b56\u7565\u6027\u8d1f\u6837\u672c\uff0c\u6211\u4eec\u5229\u7528\u62d2\u7edd\u91c7\u6837\u7684\u8fed\u4ee3\u8fc7\u7a0b\u6765\u6536\u96c6\u989d\u5916\u7684\u504f\u597d\u6570\u636e\u3002\u6211\u4eec\u5728\u63a5\u4e0b\u6765\u7684\u90e8\u5206\u4e2d\u5c06\u66f4\u8be6\u7ec6\u5730\u8ba8\u8bba\u62d2\u7edd\u91c7\u6837\u7684\u4f7f\u7528\u65b9\u5f0f\u3002\u603b\u800c\u8a00\u4e4b\uff0c\u62d2\u7edd\u91c7\u6837\u7528\u4e8e\u4ece\u6a21\u578b\u4e2d\u8fed\u4ee3\u5730\u91c7\u6837\u9ad8\u8d28\u91cf\u7684\u751f\u6210\u7ed3\u679c\u3002\u56e0\u6b64\uff0c\u4f5c\u4e3a\u526f\u4ea7\u54c1\uff0c\u6240\u6709\u672a\u88ab\u9009\u4e2d\u7684\u751f\u6210\u7ed3\u679c\u90fd\u53ef\u7528\u4f5c\u8d1f\u9762\u62d2\u7edd\u6837\u672c\uff0c\u5e76\u7528\u4f5c\u989d\u5916\u7684\u504f\u597d\u6570\u636e\u5bf9\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h4>7.5.4 \u5956\u52b1\u6a21\u578b<\/h4>\n<p>\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u89c6\u89c9\u5956\u52b1\u6a21\u578b (RM)\uff0c\u5b83\u57fa\u4e8e\u89c6\u89c9 SFT \u6a21\u578b\u548c\u8bed\u8a00 RM\u3002\u89c6\u89c9\u7f16\u7801\u5668\u548c\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u4ece\u89c6\u89c9 SFT \u6a21\u578b\u521d\u59cb\u5316\uff0c\u5e76\u5728\u8bad\u7ec3\u671f\u95f4\u89e3\u51bb\uff0c\u800c\u81ea\u6ce8\u610f\u529b\u5c42\u5219\u4ece\u8bed\u8a00 RM \u521d\u59cb\u5316\u5e76\u4fdd\u6301\u51bb\u7ed3\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u51bb\u7ed3\u8bed\u8a00 RM \u90e8\u5206\u901a\u5e38\u4f1a\u5bfc\u81f4\u66f4\u597d\u7684\u7cbe\u5ea6\uff0c\u5c24\u5176\u662f\u5728\u9700\u8981 RM \u6839\u636e\u5176\u77e5\u8bc6\u6216\u8bed\u8a00\u8d28\u91cf\u8fdb\u884c\u5224\u65ad\u7684\u4efb\u52a1\u4e2d\u3002\u6211\u4eec\u91c7\u7528\u4e0e\u8bed\u8a00 RM \u76f8\u540c\u7684\u8bad\u7ec3\u76ee\u6807\uff0c\u4f46\u6dfb\u52a0\u4e86\u4e00\u4e2a\u52a0\u6743\u6b63\u5219\u5316\u9879\u6765\u5bf9\u6279\u91cf\u5e73\u5747\u7684\u5956\u52b1 logits \u5e73\u65b9\u8fdb\u884c\u5904\u7406\uff0c\u4ee5\u9632\u6b62\u5956\u52b1\u5206\u6570\u6f02\u79fb\u3002<\/p>\n<p>\u7b2c 7.5.3 \u8282\u4e2d\u7684\u4eba\u7c7b\u504f\u597d\u6ce8\u91ca\u7528\u4e8e\u8bad\u7ec3\u89c6\u89c9 RM\u3002\u6211\u4eec\u9075\u5faa\u4e0e\u8bed\u8a00\u504f\u597d\u6570\u636e\uff08\u7b2c 4.2.1 \u8282\uff09\u76f8\u540c\u7684\u505a\u6cd5\uff0c\u521b\u5efa\u4e24\u5230\u4e09\u4e2a\u5177\u6709\u6e05\u6670\u6392\u540d\u7684\u5bf9\uff08\u7f16\u8f91\u7248\u672c &gt; \u9009\u62e9\u7248\u672c &gt; \u62d2\u7edd\u7248\u672c\uff09\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u901a\u8fc7\u6270\u52a8\u4e0e\u56fe\u50cf\u4fe1\u606f\u76f8\u5173\u7684\u8bcd\u8bed\u6216\u77ed\u8bed\uff08\u4f8b\u5982\u6570\u5b57\u6216\u89c6\u89c9\u6587\u672c\uff09\u6765\u5408\u6210\u589e\u5f3a\u8d1f\u9762\u56de\u590d\u3002\u8fd9\u9f13\u52b1\u89c6\u89c9 RM \u57fa\u4e8e\u5b9e\u9645\u7684\u56fe\u50cf\u5185\u5bb9\u8fdb\u884c\u5224\u65ad\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>7.5.5 \u76f4\u63a5\u504f\u597d\u4f18\u5316<\/h4>\n<p>\u4e0e\u8bed\u8a00\u6a21\u578b\uff08\u7b2c 4.1.4 \u8282\uff09\u7c7b\u4f3c\uff0c\u6211\u4eec\u4f7f\u7528\u76f4\u63a5\u504f\u597d\u4f18\u5316 (DPO\uff1bRafailov \u7b49\u4eba (2023)) \u8fdb\u4e00\u6b65\u8bad\u7ec3\u89c6\u89c9\u9002\u914d\u5668\uff0c\u5e76\u4f7f\u7528\u7b2c 7.5.3 \u8282\u4e2d\u63cf\u8ff0\u7684\u504f\u597d\u6570\u636e\u3002\u4e3a\u4e86\u5bf9\u6297\u540e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5206\u5e03\u504f\u79fb\uff0c\u6211\u4eec\u53ea\u4fdd\u7559\u6700\u8fd1\u51e0\u6279\u4eba\u7c7b\u504f\u597d\u6ce8\u91ca\uff0c\u800c\u4e22\u5f03\u90a3\u4e9b\u4e0e\u7b56\u7565\u5dee\u8ddd\u8f83\u5927\u7684\u6279\u6b21\uff08\u4f8b\u5982\uff0c\u5982\u679c\u66f4\u6539\u4e86\u57fa\u672c\u9884\u8bad\u7ec3\u6a21\u578b\uff09\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u4e0e\u5176\u59cb\u7ec8\u51bb\u7ed3\u53c2\u8003\u6a21\u578b\uff0c\u4e0d\u5982\u6bcf k \u6b65\u4ee5\u6307\u6570\u79fb\u52a8\u5e73\u5747 (EMA) \u7684\u65b9\u5f0f\u66f4\u65b0\u5b83\uff0c\u6709\u52a9\u4e8e\u6a21\u578b\u4ece\u6570\u636e\u4e2d\u5b66\u4e60\u66f4\u591a\uff0c\u4ece\u800c\u5728\u4eba\u7c7b\u8bc4\u4f30\u4e2d\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002\u603b\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u89c6\u89c9 DPO \u6a21\u578b\u5728\u4eba\u7c7b\u8bc4\u4f30\u4e2d\u59cb\u7ec8\u4f18\u4e8e\u5176 SFT \u8d77\u59cb\u70b9\uff0c\u5e76\u4e14\u5728\u6bcf\u6b21\u5fae\u8c03\u8fed\u4ee3\u4e2d\u90fd\u8868\u73b0\u51fa\u8272\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>7.5.6 \u62d2\u7edd\u91c7\u6837<\/h4>\n<p>\u73b0\u6709\u7684\u5927\u591a\u6570\u95ee\u7b54\u5bf9\u4ec5\u5305\u542b\u6700\u7ec8\u7b54\u6848\uff0c\u800c\u7f3a\u4e4f\u63a8\u7406\u4efb\u52a1\u6cdb\u5316\u826f\u597d\u7684\u6a21\u578b\u6240\u9700\u7684\u601d\u7ef4\u94fe\u89e3\u91ca\u3002\u6211\u4eec\u4f7f\u7528\u62d2\u7edd\u91c7\u6837\u6765\u4e3a\u8fd9\u4e9b\u793a\u4f8b\u751f\u6210\u7f3a\u5c11\u7684\u89e3\u91ca\uff0c\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p>\u7ed9\u5b9a\u4e00\u4e2a\u95ee\u7b54\u5bf9\uff0c\u6211\u4eec\u901a\u8fc7\u4f7f\u7528\u4e0d\u540c\u7684\u7cfb\u7edf\u63d0\u793a\u6216\u6e29\u5ea6\u5bf9\u5fae\u8c03\u540e\u7684\u6a21\u578b\u8fdb\u884c\u91c7\u6837\u6765\u751f\u6210\u591a\u4e2a\u7b54\u6848\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u751f\u6210\u7684\u7b54\u6848\u4e0e\u771f\u5b9e\u7b54\u6848\u901a\u8fc7\u542f\u53d1\u5f0f\u65b9\u6cd5\u6216LLM\u88c1\u5224\u8fdb\u884c\u6bd4\u8f83\u3002\u6700\u540e\uff0c\u6211\u4eec\u5c06\u6b63\u786e\u7684\u7b54\u6848\u6dfb\u52a0\u5230\u5fae\u8c03\u6570\u636e\u4e2d\u91cd\u65b0\u8bad\u7ec3\u6a21\u578b\u3002\u6211\u4eec\u53d1\u73b0\u6bcf\u9053\u9898\u4fdd\u7559\u591a\u4e2a\u6b63\u786e\u7b54\u6848\u662f\u6709\u7528\u7684\u3002<\/p>\n<p>\u4e3a\u4e86\u786e\u4fdd\u53ea\u5c06\u9ad8\u8d28\u91cf\u7684\u793a\u4f8b\u6dfb\u52a0\u5230\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u5b9e\u65bd\u4e86\u4ee5\u4e0b\u4e24\u4e2a\u5b89\u5168\u63aa\u65bd\uff1a<\/p>\n<ol>\n<li>\u6211\u4eec\u53d1\u73b0\uff0c\u4e00\u4e9b\u793a\u4f8b\u5305\u542b\u9519\u8bef\u7684\u89e3\u91ca\uff0c\u5c3d\u7ba1\u6700\u7ec8\u7b54\u6848\u662f\u6b63\u786e\u7684\u3002\u6211\u4eec\u6ce8\u610f\u5230\u8fd9\u79cd\u6a21\u5f0f\u5728\u53ea\u6709\u5f88\u5c0f\u4e00\u90e8\u5206\u751f\u6210\u7684\u7b54\u6848\u6b63\u786e\u7684\u9898\u76ee\u4e2d\u66f4\u5e38\u89c1\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5c06\u6b63\u786e\u7b54\u6848\u6982\u7387\u4f4e\u4e8e\u7279\u5b9a\u9608\u503c\u7684\u9898\u76ee\u7684\u7b54\u6848\u4e22\u5f03\u3002<\/li>\n<li>\u8bc4\u5ba1\u5458\u7531\u4e8e\u8bed\u8a00\u6216\u98ce\u683c\u7684\u5dee\u5f02\u800c\u504f\u7231\u67d0\u4e9b\u7b54\u6848\u3002\u6211\u4eec\u4f7f\u7528\u5956\u52b1\u6a21\u578b\u6765\u9009\u62e9 K \u4e2a\u6700\u9ad8\u8d28\u91cf\u7684\u7b54\u6848\u5e76\u5c06\u5176\u6dfb\u52a0\u5230\u8bad\u7ec3\u4e2d\u3002<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h4>7.5.7 \u8d28\u91cf\u8c03\u4f18<\/h4>\n<p>\u6211\u4eec\u7cbe\u5fc3\u7b56\u5212\u4e86\u4e00\u4e2a\u89c4\u6a21\u867d\u5c0f\u4f46\u9ad8\u5ea6\u7cbe\u9009\u7684\u7cbe\u7ec6\u8c03\u4f18\uff08SFT\uff09\u6570\u636e\u96c6\uff0c\u6240\u6709\u6837\u672c\u90fd\u7ecf\u8fc7\u4eba\u5de5\u6216\u6211\u4eec\u6700\u4f73\u6a21\u578b\u7684\u91cd\u5199\u548c\u9a8c\u8bc1\uff0c\u4ee5\u6ee1\u8db3\u6700\u9ad8\u6807\u51c6\u3002 \u6211\u4eec\u7528\u8fd9\u4e9b\u6570\u636e\u8bad\u7ec3DPO\u6a21\u578b\uff0c\u4ee5\u63d0\u9ad8\u54cd\u5e94\u8d28\u91cf\uff0c\u5e76\u5c06\u6b64\u8fc7\u7a0b\u79f0\u4e3a\u8d28\u91cf\u8c03\u4f18\uff08QT\uff09\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u5f53QT\u6570\u636e\u96c6\u6db5\u76d6\u5e7f\u6cdb\u7684\u4efb\u52a1\u5e76\u4e14\u5e94\u7528\u9002\u5f53\u7684\u63d0\u524d\u505c\u6b62\u65f6\uff0cQT\u53ef\u4ee5\u663e\u7740\u63d0\u9ad8\u4eba\u7c7b\u8bc4\u4f30\u7ed3\u679c\uff0c\u800c\u4e0d\u4f1a\u5f71\u54cd\u57fa\u51c6\u6d4b\u8bd5\u9a8c\u8bc1\u7684\u4e00\u822c\u6027\u6027\u80fd\u3002\u5728\u6b64\u9636\u6bb5\uff0c\u6211\u4eec\u4ec5\u6839\u636e\u57fa\u51c6\u6d4b\u8bd5\u9009\u62e9\u68c0\u67e5\u70b9\uff0c\u4ee5\u786e\u4fdd\u80fd\u529b\u4fdd\u6301\u6216\u6539\u8fdb\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>7.6 \u56fe\u50cf\u8bc6\u522b\u7ed3\u679c<\/h3>\n<p>\u6211\u4eec\u8bc4\u4f30\u4e86 Llama 3 \u56fe\u50cf\u7406\u89e3\u80fd\u529b\u5728\u4e00\u7cfb\u5217\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\uff0c\u8fd9\u4e9b\u4efb\u52a1\u6db5\u76d6\u4e86\u81ea\u7136\u56fe\u50cf\u7406\u89e3\u3001\u6587\u672c\u7406\u89e3\u3001\u56fe\u8868\u7406\u89e3\u548c\u591a\u6a21\u6001\u63a8\u7406\uff1a<\/p>\n<ul>\n<li>MMMU (Yue \u7b49\u4eba\uff0c2024a) \u662f\u4e00\u4e2a\u5177\u6709\u6311\u6218\u6027\u7684\u591a\u6a21\u6001\u63a8\u7406\u6570\u636e\u96c6\uff0c\u6a21\u578b\u9700\u8981\u7406\u89e3\u56fe\u50cf\u5e76\u89e3\u51b3\u8de8\u8d8a 30 \u4e2a\u4e0d\u540c\u5b66\u79d1\u7684\u5927\u5b66\u6c34\u5e73\u95ee\u9898\u3002\u8fd9\u5305\u62ec\u591a\u9879\u9009\u62e9\u9898\u548c\u5f00\u653e\u5f0f\u95ee\u9898\u3002\u6211\u4eec\u5c06\u6a21\u578b\u5728\u5305\u542b 900 \u5f20\u56fe\u50cf\u7684\u9a8c\u8bc1\u96c6\u4e0a\u8fdb\u884c\u8bc4\u4f30\uff0c\u4e0e\u5176\u4ed6\u5de5\u4f5c\u4e00\u81f4\u3002<\/li>\n<li>VQAv2 (Antol \u7b49\u4eba\uff0c2015) \u6d4b\u8bd5\u4e86\u6a21\u578b\u7ed3\u5408\u56fe\u50cf\u7406\u89e3\u3001\u8bed\u8a00\u7406\u89e3\u548c\u5e38\u8bc6\u77e5\u8bc6\u6765\u56de\u7b54\u5173\u4e8e\u81ea\u7136\u56fe\u50cf\u7684\u4e00\u822c\u6027\u95ee\u9898\u7684\u80fd\u529b\u3002<\/li>\n<li>AI2 Diagram (Kembhavi \u7b49\u4eba\uff0c2016) \u8bc4\u4f30\u4e86\u6a21\u578b\u89e3\u6790\u79d1\u5b66\u56fe\u8868\u5e76\u56de\u7b54\u6709\u5173\u56fe\u8868\u95ee\u9898\u7684\u80fd\u529b\u3002\u6211\u4eec\u4f7f\u7528\u4e0e <a href=\"https:\/\/www.kdjingpai.com\/pt\/geminibardgubu\/\">Gemini<\/a> \u548c x.ai \u76f8\u540c\u7684\u8bc4\u4f30\u534f\u8bae\uff0c\u5e76\u4f7f\u7528\u900f\u660e\u8fb9\u754c\u6846\u62a5\u544a\u5206\u6570\u3002<\/li>\n<li>ChartQA (Masry \u7b49\u4eba\uff0c2022) \u662f\u4e00\u4e2a\u5177\u6709\u6311\u6218\u6027\u7684\u56fe\u8868\u7406\u89e3\u57fa\u51c6\u6d4b\u8bd5\u3002\u8fd9\u8981\u6c42\u6a21\u578b\u5728\u89c6\u89c9\u4e0a\u7406\u89e3\u4e0d\u540c\u7c7b\u578b\u7684\u56fe\u8868\u5e76\u56de\u7b54\u6709\u5173\u8fd9\u4e9b\u56fe\u8868\u903b\u8f91\u95ee\u9898\u3002<\/li>\n<li>TextVQA (Singh \u7b49\u4eba\uff0c2019) \u662f\u4e00\u4e2a\u6d41\u884c\u7684\u57fa\u51c6\u6570\u636e\u96c6\uff0c\u9700\u8981\u6a21\u578b\u9605\u8bfb\u548c\u63a8\u7406\u56fe\u50cf\u4e2d\u7684\u6587\u672c\u4ee5\u56de\u7b54\u6709\u5173\u5b83\u4eec\u7684\u7591\u95ee\u3002\u8fd9\u6d4b\u8bd5\u4e86\u6a21\u578b\u5bf9\u81ea\u7136\u56fe\u50cf\u4e2d OCR \u7406\u89e3\u80fd\u529b\u3002<\/li>\n<li>DocVQA (Mathew \u7b49\u4eba\uff0c2020) \u662f\u4e00\u4e2a\u4e13\u6ce8\u4e8e\u6587\u6863\u5206\u6790\u548c\u8bc6\u522b\u7684\u57fa\u51c6\u6570\u636e\u96c6\u3002\u5b83\u5305\u542b\u5404\u79cd\u6587\u6863\u7684\u56fe\u50cf\uff0c\u8bc4\u4f30\u4e86\u6a21\u578b\u6267\u884c OCR \u7406\u89e3\u5e76\u63a8\u7406\u6587\u6863\u5185\u5bb9\u4ee5\u56de\u7b54\u6709\u5173\u5b83\u4eec\u7684\u95ee\u9898\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>\u8868 29 \u5c55\u793a\u4e86\u6211\u4eec\u5b9e\u9a8c\u7684\u7ed3\u679c\u3002\u8868\u4e2d\u7684\u7ed3\u679c\u8868\u660e\uff0c\u9644\u52a0\u5230 Llama 3 \u4e0a\u7684\u89c6\u89c9\u6a21\u5757\u5728\u4e0d\u540c\u6a21\u578b\u5bb9\u91cf\u7684\u5404\u79cd\u56fe\u50cf\u8bc6\u522b\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u90fd\u5177\u6709\u7ade\u4e89\u529b\u3002\u4f7f\u7528 resulting Llama 3-V 405B \u6a21\u578b\uff0c\u6211\u4eec\u5728\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u5747\u4f18\u4e8e GPT-4V\uff0c\u4f46\u7565\u900a\u4e8e Gemini 1.5 Pro \u548c Claude 3.5 Sonnet\u3002Llama 3 405B \u5728\u6587\u6863\u7406\u89e3\u4efb\u52a1\u4e0a\u8868\u73b0\u5c24\u4e3a\u51fa\u8272\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3488\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4df98bc4d7137f4.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"790\" height=\"204\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4df98bc4d7137f4.png 1411w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4df98bc4d7137f4-300x78.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4df98bc4d7137f4-1024x265.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/4df98bc4d7137f4-768x199.png 768w\" sizes=\"auto, (max-width: 790px) 100vw, 790px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>7.7 \u89c6\u9891\u8bc6\u522b\u7ed3\u679c<\/h3>\n<p>\u6211\u4eec\u5bf9 Llama 3 \u7684\u89c6\u9891\u9002\u914d\u5668\u5728\u4e09\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff1a<\/p>\n<ul>\n<li><strong>PerceptionTest (Lin et al., 2023)<\/strong>\uff1a\u8be5\u57fa\u51c6\u6d4b\u8bd5\u6a21\u578b\u5bf9\u77ed\u89c6\u9891\u7247\u6bb5\u8fdb\u884c\u7406\u89e3\u548c\u9884\u6d4b\u7684\u80fd\u529b\u3002\u5b83\u5305\u542b\u4e86\u5404\u79cd\u7c7b\u578b\u7684\u95ee\u9898\uff0c\u4f8b\u5982\u8bc6\u522b\u7269\u4f53\u3001\u52a8\u4f5c\u3001\u573a\u666f\u7b49\u3002\u6211\u4eec\u6839\u636e\u5b98\u65b9\u63d0\u4f9b\u7684\u4ee3\u7801\u548c\u8bc4\u4f30\u6307\u6807\uff08\u51c6\u786e\u7387\uff09\u62a5\u544a\u7ed3\u679c\u3002<\/li>\n<li><strong>TVQA (Lei et al., 2018)<\/strong>\uff1a\u8be5\u57fa\u51c6\u8bc4\u4f30\u6a21\u578b\u7684\u590d\u5408\u63a8\u7406\u80fd\u529b\uff0c\u9700\u8981\u8fdb\u884c\u7a7a\u95f4\u65f6\u95f4\u5b9a\u4f4d\u3001\u8bc6\u522b\u89c6\u89c9\u6982\u5ff5\u4ee5\u53ca\u4e0e\u5b57\u5e55\u5bf9\u8bdd\u8054\u5408\u63a8\u7406\u3002\u7531\u4e8e\u8be5\u6570\u636e\u96c6\u6765\u6e90\u4e8e\u6d41\u884c\u7684\u7535\u89c6\u8282\u76ee\uff0c\u56e0\u6b64\u5b83\u8fd8\u6d4b\u8bd5\u4e86\u6a21\u578b\u5229\u7528\u8fd9\u4e9b\u7535\u89c6\u8282\u76ee\u7684\u5916\u90e8\u77e5\u8bc6\u6765\u56de\u7b54\u95ee\u9898\u7684\u80fd\u529b\u3002\u5b83\u5305\u542b\u8d85\u8fc7 1.5 \u4e07\u4e2a\u9a8c\u8bc1 QA \u5bf9\uff0c\u6bcf\u4e2a\u5bf9\u5e94\u7684\u89c6\u9891\u7247\u6bb5\u5e73\u5747\u957f\u5ea6\u4e3a 76 \u79d2\u3002\u5b83\u91c7\u7528\u591a\u9879\u9009\u62e9\u683c\u5f0f\uff0c\u6bcf\u4e2a\u95ee\u9898\u6709\u4e94\u4e2a\u9009\u9879\uff0c\u6211\u4eec\u6839\u636e\u5148\u524d\u5de5\u4f5c\uff08OpenAI, 2023b\uff09\u5728\u9a8c\u8bc1\u96c6\u4e0a\u62a5\u544a\u6027\u80fd\u3002<\/li>\n<li><strong>ActivityNet-QA (Yu et al., 2019)<\/strong>\uff1a\u8be5\u57fa\u51c6\u8bc4\u4f30\u6a21\u578b\u5bf9\u957f\u89c6\u9891\u7247\u6bb5\u8fdb\u884c\u7406\u89e3\u4ee5\u4e86\u89e3\u52a8\u4f5c\u3001\u7a7a\u95f4\u5173\u7cfb\u3001\u65f6\u95f4\u5173\u7cfb\u3001\u8ba1\u6570\u7b49\u7684\u80fd\u529b\u3002\u5b83\u5305\u542b 800 \u4e2a\u89c6\u9891\u4e2d\u7684 8,000 \u4e2a\u6d4b\u8bd5 QA \u5bf9\uff0c\u6bcf\u4e2a\u89c6\u9891\u5e73\u5747\u957f\u5ea6\u4e3a 3 \u5206\u949f\u3002\u5bf9\u4e8e\u8bc4\u4f30\uff0c\u6211\u4eec\u9075\u5faa\u5148\u524d\u5de5\u4f5c\uff08Google, 2023\uff1bLin et al., 2023\uff1bMaaz et al., 2024\uff09\u7684\u534f\u8bae\uff0c\u5176\u4e2d\u6a21\u578b\u751f\u6210\u77ed\u7684\u5355\u8bcd\u6216\u77ed\u8bed\u56de\u7b54\uff0c\u5e76\u4e14\u4f7f\u7528 GPT-3.5 API \u5c06\u5176\u4e0e\u771f\u5b9e\u7b54\u6848\u8fdb\u884c\u6bd4\u8f83\u4ee5\u8bc4\u4f30\u8f93\u51fa\u7684\u6b63\u786e\u6027\u3002\u6211\u4eec\u62a5\u544a API \u8ba1\u7b97\u51fa\u7684\u5e73\u5747\u51c6\u786e\u7387\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><strong>\u63a8\u7406\u8fc7\u7a0b<\/strong><\/p>\n<p>\u5728\u6267\u884c\u63a8\u7406\u65f6\uff0c\u6211\u4eec\u4ece\u5b8c\u6574\u7684\u89c6\u9891\u7247\u6bb5\u4e2d\u5747\u5300\u91c7\u6837\u5e27\u5e76\u5c06\u5176\u4e0e\u7b80\u77ed\u7684\u6587\u672c\u63d0\u793a\u4e00\u8d77\u4f20\u9012\u7ed9\u6a21\u578b\u3002\u7531\u4e8e\u5927\u591a\u6570\u57fa\u51c6\u90fd\u6d89\u53ca\u56de\u7b54\u591a\u9879\u9009\u62e9\u9898\uff0c\u56e0\u6b64\u6211\u4eec\u4f7f\u7528\u4ee5\u4e0b\u63d0\u793a\uff1a<\/p>\n<ul>\n<li>\u4ece\u4ee5\u4e0b\u9009\u9879\u4e2d\u9009\u62e9\u6b63\u786e\u7684\u7b54\u6848\uff1a{\u95ee\u9898}\u3002\u53ea\u7528\u6b63\u786e\u9009\u9879\u5b57\u6bcd\u56de\u7b54\uff0c\u4e0d\u8981\u5199\u5176\u4ed6\u4efb\u4f55\u4e1c\u897f\u3002<\/li>\n<\/ul>\n<p>\u5bf9\u4e8e\u9700\u8981\u751f\u6210\u7b80\u77ed\u7b54\u6848\uff08\u4f8b\u5982 ActivityNet-QA \u548c NExT-QA\uff09\u7684\u57fa\u51c6\uff0c\u6211\u4eec\u4f7f\u7528\u4ee5\u4e0b\u63d0\u793a\uff1a<\/p>\n<ul>\n<li>\u4f7f\u7528\u4e00\u4e2a\u5355\u8bcd\u6216\u77ed\u8bed\u56de\u7b54\u95ee\u9898\uff1a{\u95ee\u9898}\u3002<\/li>\n<\/ul>\n<p>\u5bf9\u4e8e NExT-QA\uff0c\u7531\u4e8e\u8bc4\u4f30\u6307\u6807\uff08WUPS\uff09\u5bf9\u957f\u5ea6\u548c\u4f7f\u7528\u7684\u7279\u5b9a\u5355\u8bcd\u5f88\u654f\u611f\uff0c\u56e0\u6b64\u6211\u4eec\u8fd8\u63d0\u793a\u6a21\u578b\u8981\u5177\u4f53\u5316\u5e76\u56de\u590d\u6700\u7a81\u51fa\u7684\u7b54\u6848\uff0c\u4f8b\u5982\u5728\u88ab\u95ee\u53ca\u4f4d\u7f6e\u95ee\u9898\u65f6\u6307\u5b9a\u201c\u5ba2\u5385\u201d\u800c\u4e0d\u662f\u7b80\u5355\u5730\u56de\u7b54\u201c\u623f\u5b50\u201d\u3002\u5bf9\u4e8e\u5305\u542b\u5b57\u5e55\uff08\u5373 TVQA\uff09\u7684\u57fa\u51c6\uff0c\u6211\u4eec\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5c06\u7247\u6bb5\u5bf9\u5e94\u7684\u5b57\u5e55\u5305\u542b\u5728\u63d0\u793a\u4e2d\u3002<\/p>\n<p><strong>\u7ed3\u679c<\/strong><\/p>\n<p>\u8868 30 \u5c55\u793a\u4e86 Llama 3 8B \u548c 70B \u6a21\u578b\u7684\u6027\u80fd\u3002\u6211\u4eec\u5c06\u5176\u4e0e\u4e24\u4e2a Gemini \u6a21\u578b\u548c\u4e24\u4e2a GPT-4 \u6a21\u578b\u7684\u6027\u80fd\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u8bf7\u6ce8\u610f\uff0c\u6240\u6709\u7ed3\u679c\u90fd\u662f\u96f6\u6837\u672c\u7684\u7ed3\u679c\uff0c\u56e0\u4e3a\u6211\u4eec\u5728\u8bad\u7ec3\u6216\u5fae\u8c03\u6570\u636e\u4e2d\u6ca1\u6709\u5305\u542b\u8fd9\u4e9b\u57fa\u51c6\u7684\u4efb\u4f55\u90e8\u5206\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u6211\u4eec\u7684 Llama 3 \u6a21\u578b\u5728\u540e\u5904\u7406\u671f\u95f4\u8bad\u7ec3\u5c0f\u578b\u89c6\u9891\u9002\u914d\u5668\u975e\u5e38\u6709\u7ade\u4e89\u529b\uff0c\u5e76\u4e14\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u751a\u81f3\u6bd4\u5176\u4ed6\u53ef\u80fd\u4ece\u9884\u8bad\u7ec3\u5f00\u59cb\u5c31\u5229\u7528\u539f\u751f\u591a\u6a21\u6001\u5904\u7406\u7684\u6a21\u578b\u66f4\u597d\u3002Llama 3 \u5728\u89c6\u9891\u8bc6\u522b\u65b9\u9762\u8868\u73b0\u7279\u522b\u51fa\u8272\uff0c\u56e0\u4e3a\u6211\u4eec\u4ec5\u8bc4\u4f30\u4e86 8B \u548c 70B \u53c2\u6570\u6a21\u578b\u3002Llama 3 \u5728 PerceptionTest \u4e0a\u53d6\u5f97\u4e86\u6700\u4f73\u6210\u7ee9\uff0c\u8fd9\u8868\u660e\u8be5\u6a21\u578b\u5177\u6709\u5f88\u5f3a\u7684\u8fdb\u884c\u590d\u6742\u65f6\u95f4\u63a8\u7406\u7684\u80fd\u529b\u3002\u5728\u50cf ActivityNet-QA \u8fd9\u6837\u7684\u957f\u7247\u6bb5\u6d3b\u52a8\u7406\u89e3\u4efb\u52a1\u4e2d\uff0c\u5373\u4f7f Llama 3 \u53ea\u5904\u7406\u9ad8\u8fbe 64 \u5e27\uff0c\u5b83\u4e5f\u80fd\u83b7\u5f97\u5f3a\u52b2\u7684\u7ed3\u679c\uff08\u5bf9\u4e8e 3 \u5206\u949f\u7684\u89c6\u9891\uff0c\u6a21\u578b\u6bcf 3 \u79d2\u53ea\u5904\u7406\u4e00\u5e27\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3489\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b49fa0ca37d1d31.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"812\" height=\"188\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b49fa0ca37d1d31.png 1410w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b49fa0ca37d1d31-300x70.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b49fa0ca37d1d31-1024x237.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/b49fa0ca37d1d31-768x178.png 768w\" sizes=\"auto, (max-width: 812px) 100vw, 812px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3490\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1cd86c8488594b5.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"809\" height=\"227\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1cd86c8488594b5.png 1387w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1cd86c8488594b5-300x84.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1cd86c8488594b5-1024x287.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/1cd86c8488594b5-768x215.png 768w\" sizes=\"auto, (max-width: 809px) 100vw, 809px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>8 \u8bed\u97f3\u5b9e\u9a8c<\/h2>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u7814\u7a76\u5c06\u8bed\u97f3\u529f\u80fd\u6574\u5408\u5230 Llama 3 \u4e2d\u7684\u7ec4\u5408\u65b9\u6cd5\uff0c\u7c7b\u4f3c\u4e8e\u6211\u4eec\u7528\u4e8e\u89c6\u89c9\u8bc6\u522b\u7684\u65b9\u6848\u3002\u5728\u8f93\u5165\u7aef\uff0c\u6dfb\u52a0\u4e86\u7f16\u7801\u5668\u548c\u9002\u914d\u5668\u6765\u5904\u7406\u8bed\u97f3\u4fe1\u53f7\u3002\u6211\u4eec\u5229\u7528\u7cfb\u7edf\u63d0\u793a\uff08\u4ee5\u6587\u672c\u5f62\u5f0f\uff09\u6765\u4f7f Llama 3 \u652f\u6301\u4e0d\u540c\u7684\u8bed\u97f3\u7406\u89e3\u6a21\u5f0f\u3002\u5982\u679c\u6ca1\u6709\u63d0\u4f9b\u7cfb\u7edf\u63d0\u793a\uff0c\u8be5\u6a21\u578b\u5c06\u5145\u5f53\u901a\u7528\u8bed\u97f3\u5bf9\u8bdd\u6a21\u578b\uff0c\u80fd\u591f\u6709\u6548\u5730\u4ee5\u4e0e\u4ec5\u6587\u672c\u7248\u672c Llama 3 \u4e00\u81f4\u7684\u65b9\u5f0f\u54cd\u5e94\u7528\u6237\u8bed\u97f3\u3002\u5f15\u5165\u5bf9\u8bdd\u5386\u53f2\u4f5c\u4e3a\u63d0\u793a\u524d\u7f00\u53ef\u4ee5\u6539\u5584\u591a\u8f6e\u5bf9\u8bdd\u4f53\u9a8c\u3002\u6211\u4eec\u8fd8\u5c1d\u8bd5\u4e86\u4f7f\u7528\u7cfb\u7edf\u63d0\u793a\u6765\u5b9e\u73b0 Llama 3 \u7684\u81ea\u52a8\u8bed\u97f3\u8bc6\u522b\uff08ASR\uff09\u548c\u81ea\u52a8\u8bed\u97f3\u7ffb\u8bd1\uff08AST\uff09\u3002Llama 3 \u7684\u8bed\u97f3\u754c\u9762\u652f\u6301\u9ad8\u8fbe 34 \u79cd\u8bed\u8a00\u300218 \u5b83\u8fd8\u5141\u8bb8\u6587\u672c\u548c\u8bed\u97f3\u4ea4\u66ff\u8f93\u5165\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u89e3\u51b3\u9ad8\u7ea7\u97f3\u9891\u7406\u89e3\u4efb\u52a1\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u5c1d\u8bd5\u4e86\u4e00\u79cd\u8bed\u97f3\u751f\u6210\u65b9\u6cd5\uff0c\u5176\u4e2d\u6211\u4eec\u5b9e\u73b0\u4e86\u6d41\u5f0f\u6587\u672c\u5230\u8bed\u97f3\uff08TTS\uff09\u7cfb\u7edf\uff0c\u8be5\u7cfb\u7edf\u5728\u8bed\u8a00\u6a21\u578b\u89e3\u7801\u8fc7\u7a0b\u4e2d\u52a8\u6001\u751f\u6210\u8bed\u97f3\u6ce2\u5f62\u3002\u6211\u4eec\u57fa\u4e8e\u4e13\u6709\u7684 TTS \u7cfb\u7edf\u8bbe\u8ba1\u4e86 Llama 3 \u7684\u8bed\u97f3\u751f\u6210\u5668\uff0c\u5e76\u4e14\u6ca1\u6709\u9488\u5bf9\u8bed\u97f3\u751f\u6210\u5fae\u8c03\u8bed\u8a00\u6a21\u578b\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u901a\u8fc7\u5728\u63a8\u7406\u65f6\u5229\u7528 Llama 3 \u8bcd\u5d4c\u5165\u6765\u4e13\u6ce8\u4e8e\u6539\u8fdb\u8bed\u97f3\u5408\u6210\u7684\u5ef6\u8fdf\u3001\u51c6\u786e\u6027\u548c\u81ea\u7136\u5ea6\u3002\u8bed\u97f3\u754c\u9762\u5982\u56fe 28 \u548c 29 \u6240\u793a\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>8.1 \u6570\u636e<\/h3>\n<h4>8.1.1 \u8bed\u97f3\u7406\u89e3<\/h4>\n<p>\u8bad\u7ec3\u6570\u636e\u53ef\u4ee5\u5206\u4e3a\u4e24\u7c7b\u3002\u9884\u8bad\u7ec3\u6570\u636e\u5305\u62ec\u5927\u91cf\u7684\u672a\u6807\u8bb0\u8bed\u97f3\uff0c\u7528\u4e8e\u4ee5\u81ea\u6211\u76d1\u7763\u7684\u65b9\u5f0f\u521d\u59cb\u5316\u8bed\u97f3\u7f16\u7801\u5668\u3002\u6709\u76d1\u7763\u7684\u5fae\u8c03\u6570\u636e\u5305\u62ec\u8bed\u97f3\u8bc6\u522b\u3001\u8bed\u97f3\u7ffb\u8bd1\u548c\u53e3\u8bed\u5bf9\u8bdd\u6570\u636e\uff1b\u8fd9\u4e9b\u6570\u636e\u7528\u4e8e\u5728\u4e0e\u5927\u578b\u8bed\u8a00\u6a21\u578b\u96c6\u6210\u65f6\u89e3\u9501\u7279\u5b9a\u7684\u80fd\u529b\u3002<\/p>\n<p><strong>\u9884\u8bad\u7ec3\u6570\u636e:<\/strong>\u00a0\u4e3a\u4e86\u9884\u8bad\u7ec3\u8bed\u97f3\u7f16\u7801\u5668\uff0c\u6211\u4eec\u6574\u7406\u4e86\u4e00\u4e2a\u5305\u542b\u7ea61500\u4e07\u5c0f\u65f6\u8bed\u97f3\u8bb0\u5f55\u7684\u6570\u636e\u96c6\uff0c\u6db5\u76d6\u591a\u79cd\u8bed\u8a00\u3002 \u6211\u4eec\u4f7f\u7528\u8bed\u97f3\u6d3b\u52a8\u68c0\u6d4b (VAD) \u6a21\u578b\u8fc7\u6ee4\u97f3\u9891\u6570\u636e\uff0c\u5e76\u9009\u62e9 VAD \u9608\u503c\u9ad8\u4e8e 0.7 \u7684\u97f3\u9891\u6837\u672c\u8fdb\u884c\u9884\u8bad\u7ec3\u3002 \u5728\u8bed\u97f3\u9884\u8bad\u7ec3\u6570\u636e\u4e2d\uff0c\u6211\u4eec\u8fd8\u91cd\u70b9\u786e\u4fdd\u4e0d\u5b58\u5728\u4e2a\u4eba\u8eab\u4efd\u4fe1\u606f (PII)\u3002 \u6211\u4eec\u4f7f\u7528 Presidio Analyzer \u6765\u8bc6\u522b\u6b64\u7c7b PII\u3002<\/p>\n<p><strong>\u8bed\u97f3\u8bc6\u522b\u548c\u7ffb\u8bd1\u6570\u636e:<\/strong>\u00a0\u6211\u4eec\u7684 ASR \u8bad\u7ec3\u6570\u636e\u5305\u542b 23 \u4e07\u5c0f\u65f6\u7684\u624b\u5199\u8f6c\u5f55\u8bed\u97f3\u8bb0\u5f55\uff0c\u6db5\u76d6 34 \u79cd\u8bed\u8a00\u3002 \u6211\u4eec\u7684 AST \u8bad\u7ec3\u6570\u636e\u5305\u542b 90,000 \u5c0f\u65f6\u7684\u53cc\u5411\u7ffb\u8bd1\uff1a\u4ece 33 \u79cd\u8bed\u8a00\u5230\u82f1\u8bed\u4ee5\u53ca\u4ece\u82f1\u8bed\u5230 33 \u79cd\u8bed\u8a00\u3002\u8fd9\u4e9b\u6570\u636e\u5305\u542b\u76d1\u7763\u548c\u4f7f\u7528 NLLB \u5de5\u5177\u5305 (NLLB Team \u7b49\u4eba\uff0c2022) \u751f\u6210\u7684\u5408\u6210\u6570\u636e\u3002 \u4f7f\u7528\u5408\u6210 AST \u6570\u636e\u53ef\u4ee5\u63d0\u9ad8\u4f4e\u8d44\u6e90\u8bed\u8a00\u7684\u6a21\u578b\u8d28\u91cf\u3002 \u6211\u4eec\u6570\u636e\u4e2d\u7684\u8bed\u97f3\u7247\u6bb5\u6700\u5927\u957f\u5ea6\u4e3a 60 \u79d2\u3002<\/p>\n<p><strong>\u53e3\u8bed\u5bf9\u8bdd\u6570\u636e:<\/strong>\u00a0\u4e3a\u4e86\u5fae\u8c03\u7528\u4e8e\u53e3\u8bed\u5bf9\u8bdd\u7684\u8bed\u97f3\u9002\u914d\u5668\uff0c\u6211\u4eec\u901a\u8fc7\u8be2\u95ee\u8bed\u8a00\u6a21\u578b\u5bf9\u8fd9\u4e9b\u63d0\u793a\u7684\u8f6c\u5f55\u505a\u51fa\u56de\u5e94\u6765\u5408\u6210\u8bed\u97f3\u63d0\u793a\u7684\u56de\u590d (Fathullah \u7b49\u4eba\uff0c2024)\u3002 \u6211\u4eec\u4f7f\u7528 ASR \u6570\u636e\u96c6\u7684\u5b50\u96c6\uff08\u5305\u542b 60,000 \u5c0f\u65f6\u7684\u8bed\u97f3\uff09\u4ee5\u8fd9\u79cd\u65b9\u5f0f\u751f\u6210\u5408\u6210\u6570\u636e\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u901a\u8fc7\u5728\u7528\u4e8e\u5fae\u8c03 Llama 3 \u7684\u6570\u636e\u5b50\u96c6\u4e0a\u8fd0\u884c Voicebox TTS \u7cfb\u7edf (Le \u7b49\u4eba\uff0c2024) \u751f\u6210\u4e86 25,000 \u5c0f\u65f6\u7684\u5408\u6210\u6570\u636e\u3002 \u6211\u4eec\u4f7f\u7528\u4e86\u51e0\u79cd\u542f\u53d1\u5f0f\u65b9\u6cd5\u6765\u9009\u62e9\u4e0e\u8bed\u97f3\u5206\u5e03\u76f8\u5339\u914d\u7684\u5fae\u8c03\u6570\u636e\u7684\u5b50\u96c6\u3002\u8fd9\u4e9b\u542f\u53d1\u5f0f\u65b9\u6cd5\u5305\u62ec\u5173\u6ce8\u76f8\u5bf9\u8f83\u77ed\u4e14\u7ed3\u6784\u7b80\u5355\u7684\u63d0\u793a\uff0c\u5e76\u4e14\u4e0d\u5305\u542b\u975e\u6587\u672c\u7b26\u53f7\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>8.1.2 \u8bed\u97f3\u751f\u6210<\/h4>\n<p id=\"pageNum-e791fe29-7d24-4eba-9060-871202afbe79\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\">\u8bed\u97f3\u751f\u6210\u6570\u636e\u96c6\u4e3b\u8981\u5305\u62ec\u7528\u4e8e\u8bad\u7ec3\u6587\u672c\u89c4\u8303\u5316\uff08TN\uff09\u6a21\u578b\u548c\u97f5\u5f8b\u6a21\u578b\uff08PM\uff09\u7684\u6570\u636e\u96c6\u3002\u4e24\u79cd\u8bad\u7ec3\u6570\u636e\u90fd\u901a\u8fc7\u6dfb\u52a0 Llama 3 \u8bcd\u5d4c\u5165\u4f5c\u4e3a\u989d\u5916\u7684\u8f93\u5165\u7279\u5f81\u8fdb\u884c\u589e\u5f3a\uff0c\u4ee5\u63d0\u4f9b\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/span><\/span><\/p>\n<p id=\"pageNum-3c504fb7-e3e5-48e4-9f07-45a28a5bbf59\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-wrapper\" lang=\"zh-CN\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\">\u6587\u672c\u89c4\u8303\u5316\u6570\u636e\u3002\u6211\u4eec\u7684 TN \u8bad\u7ec3\u6570\u636e\u96c6\u5305\u542b 5.5 \u4e07\u4e2a\u6837\u672c\uff0c\u6db5\u76d6\u4e86\u5e7f\u6cdb\u7684\u7b26\u53f7\u7c7b\u522b\uff08\u4f8b\u5982\uff0c\u6570\u5b57\u3001\u65e5\u671f\u3001\u65f6\u95f4\uff09\uff0c\u8fd9\u4e9b\u7c7b\u522b\u9700\u8981\u975e\u5e73\u51e1\u7684\u89c4\u8303\u5316\u3002\u6bcf\u4e2a\u6837\u672c\u7531\u4e66\u9762\u5f62\u5f0f\u6587\u672c\u548c\u76f8\u5e94\u7684\u89c4\u8303\u5316\u53e3\u8bed\u5f62\u5f0f\u6587\u672c\u7ec4\u6210\uff0c\u5e76\u5305\u542b\u4e00\u4e2a\u63a8\u65ad\u7684\u624b\u5de5\u5236\u4f5c\u7684 TN \u89c4\u5219\u5e8f\u5217\uff0c\u7528\u4e8e\u6267\u884c\u89c4\u8303\u5316\u3002<\/span><\/span><\/span><\/p>\n<p id=\"pageNum-a1044cfb-0252-4890-bfeb-e088c5854482\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\">\u58f0\u97f5\u6a21\u578b\u6570\u636e\u3002PM \u8bad\u7ec3\u6570\u636e\u5305\u62ec\u4ece\u4e00\u4e2a\u5305\u542b 50,000 \u5c0f\u65f6\u7684 TTS \u6570\u636e\u96c6\u63d0\u53d6\u7684\u8bed\u8a00\u548c\u58f0\u97f5\u7279\u5f81\uff0c\u8fd9\u4e9b\u7279\u5f81\u4e0e\u4e13\u4e1a\u914d\u97f3\u6f14\u5458\u5728\u5f55\u97f3\u5ba4\u73af\u5883\u4e2d\u5f55\u5236\u7684\u6587\u5b57\u7a3f\u4ef6\u548c\u97f3\u9891\u914d\u5bf9\u3002<\/span><\/span><\/p>\n<p id=\"pageNum-fc46a058-d035-4b86-940a-e6148d288771\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-wrapper\" lang=\"zh-CN\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\">Llama 3 \u5d4c\u5165\u3002Llama 3 \u5d4c\u5165\u53d6\u81ea\u7b2c 16 \u5c42\u89e3\u7801\u5668\u8f93\u51fa\u3002\u6211\u4eec\u4ec5\u4f7f\u7528 Llama 3 8B \u6a21\u578b\uff0c\u5e76\u63d0\u53d6\u7ed9\u5b9a\u6587\u672c\u7684\u5d4c\u5165\uff08\u5373 TN \u7684\u4e66\u9762\u8f93\u5165\u6587\u672c\u6216 PM \u7684\u97f3\u9891\u8f6c\u5f55\uff09\uff0c\u5c31\u50cf\u5b83\u4eec\u662f\u7531 Llama 3 \u6a21\u578b\u5728\u7a7a\u7528\u6237\u63d0\u793a\u4e0b\u751f\u6210\u7684\u3002\u5728\u4e00\u4e2a\u6837\u672c\u4e2d\uff0c\u6bcf\u4e2a Llama 3 \u6807\u8bb0\u5e8f\u5217\u5757\u90fd\u660e\u786e\u5730\u4e0e TN \u6216 PM \u672c\u5730\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u76f8\u5e94\u5757\u5bf9\u9f50\uff0c\u5373 TN \u7279\u5b9a\u7684\u6587\u672c\u6807\u8bb0\uff08\u7531 Unicode \u7c7b\u522b\u533a\u5206\uff09\u6216\u8bed\u97f3\u901f\u7387\u7279\u5f81\u3002\u8fd9\u5141\u8bb8\u4f7f\u7528 Llama 3 \u6807\u8bb0\u548c\u5d4c\u5165\u7684\u6d41\u5f0f\u8f93\u5165\u8bad\u7ec3 TN \u548c PM \u6a21\u5757\u3002<\/span><\/span><\/span><\/p>\n<h3 data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\">8.2 \u6a21\u578b\u67b6\u6784<\/h3>\n<h4>8.2.1 \u8bed\u97f3\u7406\u89e3<\/h4>\n<p>\u5728\u8f93\u5165\u65b9\u9762\uff0c\u8bed\u97f3\u6a21\u5757\u7531\u4e24\u4e2a\u8fde\u7eed\u7684\u6a21\u5757\u7ec4\u6210\uff1a\u8bed\u97f3\u7f16\u7801\u5668\u548c\u9002\u914d\u5668\u3002\u8bed\u97f3\u6a21\u5757\u7684\u8f93\u51fa\u76f4\u63a5\u4f5c\u4e3a\u6807\u8bb0\u8868\u793a\u8f93\u5165\u5230\u8bed\u8a00\u6a21\u578b\u4e2d\uff0c\u4f7f\u8bed\u97f3\u548c\u6587\u672c\u6807\u8bb0\u80fd\u591f\u76f4\u63a5\u4ea4\u4e92\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e24\u4e2a\u65b0\u7684\u7279\u6b8a\u6807\u8bb0\uff0c\u7528\u4e8e\u5305\u542b\u8bed\u97f3\u8868\u793a\u5e8f\u5217\u3002\u8bed\u97f3\u6a21\u5757\u4e0e\u89c6\u89c9\u6a21\u5757\uff08\u89c1\u7b2c7\u8282\uff09\u6709\u663e\u8457\u4e0d\u540c\uff0c\u540e\u8005\u901a\u8fc7\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u5c06\u591a\u6a21\u6001\u4fe1\u606f\u8f93\u5165\u5230\u8bed\u8a00\u6a21\u578b\u4e2d\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u8bed\u97f3\u6a21\u5757\u751f\u6210\u7684\u5d4c\u5165\u53ef\u4ee5\u65e0\u7f1d\u96c6\u6210\u5230\u6587\u672c\u6807\u8bb0\u4e2d\uff0c\u4f7f\u8bed\u97f3\u63a5\u53e3\u80fd\u591f\u5229\u7528Llama 3\u8bed\u8a00\u6a21\u578b\u7684\u6240\u6709\u529f\u80fd\u3002<\/p>\n<p><strong>\u8bed\u97f3\u7f16\u7801\u5668\uff1a<\/strong>\u6211\u4eec\u7684\u8bed\u97f3\u7f16\u7801\u5668\u662f\u4e00\u4e2a\u62e5\u670910\u4ebf\u53c2\u6570\u7684Conformer\u6a21\u578b\uff08Gulati\u7b49\uff0c2020\uff09\u3002\u6a21\u578b\u7684\u8f93\u5165\u753180\u7ef4\u7684\u6885\u5c14\u9891\u8c31\u56fe\u7279\u5f81\u7ec4\u6210\uff0c\u8fd9\u4e9b\u7279\u5f81\u9996\u5148\u901a\u8fc7\u4e00\u4e2a\u6b65\u5e45\u4e3a4\u7684\u5806\u53e0\u5c42\u5904\u7406\uff0c\u7136\u540e\u901a\u8fc7\u7ebf\u6027\u6295\u5f71\u5c06\u5e27\u957f\u5ea6\u7f29\u51cf\u523040\u6beb\u79d2\u3002\u5904\u7406\u540e\u7684\u7279\u5f81\u7531\u4e00\u4e2a\u5305\u542b24\u4e2aConformer\u5c42\u7684\u7f16\u7801\u5668\u5904\u7406\u3002\u6bcf\u4e2aConformer\u5c42\u5177\u67091536\u7684\u6f5c\u5728\u7ef4\u5ea6\uff0c\u5305\u62ec\u4e24\u4e2a\u7ef4\u5ea6\u4e3a4096\u7684Macron-net\u98ce\u683c\u7684\u524d\u9988\u7f51\u7edc\u3001\u4e00\u4e2a\u6838\u5927\u5c0f\u4e3a7\u7684\u5377\u79ef\u6a21\u5757\uff0c\u4ee5\u53ca\u4e00\u4e2a\u5177\u670924\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u65cb\u8f6c\u6ce8\u610f\u529b\u6a21\u5757\uff08Su\u7b49\uff0c2024\uff09\u3002<\/p>\n<p><strong>\u8bed\u97f3\u9002\u914d\u5668\uff1a<\/strong>\u8bed\u97f3\u9002\u914d\u5668\u5305\u542b\u7ea61\u4ebf\u53c2\u6570\u3002\u5b83\u7531\u4e00\u4e2a\u5377\u79ef\u5c42\u3001\u4e00\u4e2a\u65cb\u8f6cTransformer\u5c42\u548c\u4e00\u4e2a\u7ebf\u6027\u5c42\u7ec4\u6210\u3002\u5377\u79ef\u5c42\u7684\u6838\u5927\u5c0f\u4e3a3\uff0c\u6b65\u5e45\u4e3a2\uff0c\u65e8\u5728\u5c06\u8bed\u97f3\u5e27\u957f\u5ea6\u51cf\u5c11\u523080\u6beb\u79d2\u3002\u8fd9\u5141\u8bb8\u6a21\u578b\u5411\u8bed\u8a00\u6a21\u578b\u63d0\u4f9b\u66f4\u7c97\u7c92\u5ea6\u7684\u7279\u5f81\u3002Transformer\u5c42\u7684\u6f5c\u5728\u7ef4\u5ea6\u4e3a3072\uff0c\u524d\u9988\u7f51\u7edc\u7684\u7ef4\u5ea6\u4e3a4096\uff0c\u8fdb\u4e00\u6b65\u5904\u7406\u7ecf\u8fc7\u5377\u79ef\u4e0b\u91c7\u6837\u540e\u7684\u8bed\u97f3\u4fe1\u606f\u3002\u6700\u540e\uff0c\u7ebf\u6027\u5c42\u5c06\u8f93\u51fa\u7ef4\u5ea6\u6620\u5c04\u5230\u4e0e\u8bed\u8a00\u6a21\u578b\u5d4c\u5165\u5c42\u5339\u914d\u3002<\/p>\n<p>&nbsp;<\/p>\n<h4>8.2.2 \u8bed\u97f3\u751f\u6210<\/h4>\n<p>\u6211\u4eec\u5728\u8bed\u97f3\u751f\u6210\u7684\u4e24\u4e2a\u5173\u952e\u7ec4\u4ef6\u4e2d\u4f7f\u7528\u4e86Llama 3 8B\u7684\u5d4c\u5165\uff1a\u6587\u672c\u6807\u51c6\u5316\u548c\u97f5\u5f8b\u5efa\u6a21\u3002\u6587\u672c\u6807\u51c6\u5316\uff08TN\uff09\u6a21\u5757\u901a\u8fc7\u5c06\u4e66\u9762\u6587\u672c\u4e0a\u4e0b\u6587\u5730\u8f6c\u6362\u4e3a\u53e3\u8bed\u5f62\u5f0f\u6765\u786e\u4fdd\u8bed\u4e49\u6b63\u786e\u6027\u3002\u97f5\u5f8b\u5efa\u6a21\uff08PM\uff09\u6a21\u5757\u901a\u8fc7\u4f7f\u7528\u8fd9\u4e9b\u5d4c\u5165\u9884\u6d4b\u97f5\u5f8b\u7279\u5f81\u6765\u589e\u5f3a\u81ea\u7136\u6027\u548c\u8868\u73b0\u529b\u3002\u8fd9\u4e24\u4e2a\u7ec4\u4ef6\u534f\u540c\u5de5\u4f5c\uff0c\u5b9e\u73b0\u4e86\u51c6\u786e\u800c\u81ea\u7136\u7684\u8bed\u97f3\u751f\u6210\u3002<\/p>\n<p>**\u6587\u672c\u6807\u51c6\u5316**\uff1a\u4f5c\u4e3a\u751f\u6210\u8bed\u97f3\u8bed\u4e49\u6b63\u786e\u6027\u7684\u51b3\u5b9a\u56e0\u7d20\uff0c\u6587\u672c\u6807\u51c6\u5316\uff08TN\uff09\u6a21\u5757\u8fdb\u884c\u4ece\u4e66\u9762\u6587\u672c\u5230\u76f8\u5e94\u53e3\u8bed\u5f62\u5f0f\u7684\u4e0a\u4e0b\u6587\u611f\u77e5\u8f6c\u6362\uff0c\u8fd9\u4e9b\u53e3\u8bed\u5f62\u5f0f\u6700\u7ec8\u7531\u4e0b\u6e38\u7ec4\u4ef6\u8fdb\u884c\u53e3\u5934\u5316\u3002\u4f8b\u5982\uff0c\u6839\u636e\u8bed\u4e49\u4e0a\u4e0b\u6587\uff0c\u4e66\u9762\u5f62\u5f0f\u7684\u201c123\u201d\u53ef\u80fd\u88ab\u8bfb\u4f5c\u57fa\u6570\uff08one hundred twenty three\uff09\u6216\u9010\u4e2a\u6570\u5b57\u62fc\u8bfb\uff08one two three\uff09\u3002TN\u7cfb\u7edf\u7531\u4e00\u4e2a\u6d41\u5f0f\u7684\u57fa\u4e8eLSTM\u7684\u5e8f\u5217\u6807\u8bb0\u6a21\u578b\u7ec4\u6210\uff0c\u8be5\u6a21\u578b\u9884\u6d4b\u7528\u4e8e\u8f6c\u6362\u8f93\u5165\u6587\u672c\u7684\u624b\u5de5\u5236\u4f5c\u7684TN\u89c4\u5219\u5e8f\u5217\uff08Kang\u7b49\uff0c2024\uff09\u3002\u8be5\u795e\u7ecf\u6a21\u578b\u8fd8\u901a\u8fc7\u4ea4\u53c9\u6ce8\u610f\u529b\u63a5\u6536Llama 3\u5d4c\u5165\uff0c\u4ee5\u5229\u7528\u5176\u4e2d\u7f16\u7801\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u4f7f\u5f97\u6700\u5c11\u7684\u6587\u672c\u6807\u8bb0\u524d\u77bb\u548c\u6d41\u5f0f\u8f93\u5165\/\u8f93\u51fa\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n<p>**\u97f5\u5f8b\u5efa\u6a21**\uff1a\u4e3a\u4e86\u589e\u5f3a\u5408\u6210\u8bed\u97f3\u7684\u81ea\u7136\u6027\u548c\u8868\u73b0\u529b\uff0c\u6211\u4eec\u96c6\u6210\u4e86\u4e00\u4e2a\u4ec5\u89e3\u7801Transformer\u67b6\u6784\u7684\u97f5\u5f8b\u6a21\u578b\uff08PM\uff09\uff08Radford\u7b49\uff0c2021\uff09\uff0c\u8be5\u6a21\u578b\u5c06Llama 3\u5d4c\u5165\u4f5c\u4e3a\u989d\u5916\u8f93\u5165\u3002\u8fd9\u79cd\u96c6\u6210\u5229\u7528\u4e86Llama 3\u7684\u8bed\u8a00\u80fd\u529b\uff0c\u4f7f\u7528\u5176\u6587\u672c\u8f93\u51fa\u548c\u4e2d\u95f4\u5d4c\u5165\uff08Devlin\u7b49\uff0c2018\uff1bDong\u7b49\uff0c2019\uff1bRaffel\u7b49\uff0c2020\uff1bGuo\u7b49\uff0c2023\uff09\u6765\u589e\u5f3a\u97f5\u5f8b\u7279\u5f81\u7684\u9884\u6d4b\uff0c\u4ece\u800c\u51cf\u5c11\u4e86\u6a21\u578b\u6240\u9700\u7684\u524d\u77bb\u3002PM\u96c6\u6210\u4e86\u591a\u4e2a\u8f93\u5165\u7ec4\u4ef6\u6765\u751f\u6210\u5168\u9762\u7684\u97f5\u5f8b\u9884\u6d4b\uff1a\u4ece\u4e0a\u6587\u8be6\u7ec6\u4ecb\u7ecd\u7684\u6587\u672c\u6807\u51c6\u5316\u524d\u7aef\u5f97\u51fa\u7684\u8bed\u8a00\u7279\u5f81\u3001\u6807\u8bb0\u548c\u5d4c\u5165\u3002PM\u9884\u6d4b\u4e09\u4e2a\u5173\u952e\u7684\u97f5\u5f8b\u7279\u5f81\uff1a\u6bcf\u4e2a\u97f3\u7d20\u7684\u5bf9\u6570\u6301\u7eed\u65f6\u95f4\u3001\u5bf9\u6570\u57fa\u9891\uff08F0\uff09\u5e73\u5747\u503c\u548c\u97f3\u7d20\u6301\u7eed\u65f6\u95f4\u5185\u7684\u5bf9\u6570\u529f\u7387\u5e73\u5747\u503c\u3002\u6a21\u578b\u7531\u4e00\u4e2a\u5355\u5411Transformer\u548c\u516d\u4e2a\u6ce8\u610f\u529b\u5934\u7ec4\u6210\u3002\u6bcf\u4e2a\u5757\u5305\u62ec\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u548c\u5177\u6709864\u9690\u85cf\u7ef4\u5ea6\u7684\u53cc\u5168\u8fde\u63a5\u5c42\u3002PM\u7684\u4e00\u4e2a\u663e\u8457\u7279\u5f81\u662f\u5176\u53cc\u91cd\u4ea4\u53c9\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4e00\u5c42\u4e13\u7528\u4e8e\u8bed\u8a00\u8f93\u5165\uff0c\u53e6\u4e00\u5c42\u4e13\u7528\u4e8eLlama\u5d4c\u5165\u3002\u8be5\u8bbe\u7f6e\u6709\u6548\u5730\u7ba1\u7406\u4e86\u4e0d\u540c\u8f93\u5165\u901f\u7387\uff0c\u800c\u65e0\u9700\u663e\u5f0f\u5bf9\u9f50\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>8.3 \u8bad\u7ec3\u65b9\u6848<\/h3>\n<h4>8.3.1 \u8bed\u97f3\u7406\u89e3<\/h4>\n<p>\u8bed\u97f3\u6a21\u5757\u7684\u8bad\u7ec3\u5206\u4e24\u4e2a\u9636\u6bb5\u8fdb\u884c\u3002\u7b2c\u4e00\u9636\u6bb5\uff0c\u8bed\u97f3\u9884\u8bad\u7ec3\uff0c\u5229\u7528\u672a\u6807\u8bb0\u6570\u636e\u8bad\u7ec3\u4e00\u4e2a\u8bed\u97f3\u7f16\u7801\u5668\uff0c\u8be5\u7f16\u7801\u5668\u5728\u8bed\u8a00\u548c\u58f0\u5b66\u6761\u4ef6\u65b9\u9762\u8868\u73b0\u51fa\u5f3a\u5927\u7684\u6cdb\u5316\u80fd\u529b\u3002\u7b2c\u4e8c\u9636\u6bb5\uff0c\u76d1\u7763\u5fae\u8c03\uff0c\u9002\u914d\u5668\u548c\u9884\u8bad\u7ec3\u7f16\u7801\u5668\u4e0e\u8bed\u8a00\u6a21\u578b\u96c6\u6210\uff0c\u5e76\u4e0e\u5176\u8054\u5408\u8bad\u7ec3\uff0c\u800cLLM\u4fdd\u6301\u51bb\u7ed3\u72b6\u6001\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u54cd\u5e94\u8bed\u97f3\u8f93\u5165\u3002\u6b64\u9636\u6bb5\u4f7f\u7528\u5bf9\u5e94\u8bed\u97f3\u7406\u89e3\u80fd\u529b\u7684\u6807\u8bb0\u6570\u636e\u3002<\/p>\n<p>\u591a\u8bed\u8a00 ASR \u548c AST \u5efa\u6a21\u5e38\u5e38\u5bfc\u81f4\u8bed\u8a00\u6df7\u6dc6\/\u5e72\u6270\uff0c\u4ece\u800c\u964d\u4f4e\u6027\u80fd\u3002\u4e00\u79cd\u6d41\u884c\u7684\u7f13\u89e3\u65b9\u6cd5\u662f\u5728\u6e90\u7aef\u548c\u76ee\u6807\u7aef\u90fd\u52a0\u5165\u8bed\u8a00\u8bc6\u522b\uff08LID\uff09\u4fe1\u606f\u3002\u8fd9\u53ef\u4ee5\u5728\u9884\u5148\u786e\u5b9a\u7684\u65b9\u5411\u4e0a\u63d0\u9ad8\u6027\u80fd\uff0c\u4f46\u540c\u65f6\u4e5f\u53ef\u80fd\u5bfc\u81f4\u6cdb\u5316\u80fd\u529b\u4e0b\u964d\u3002\u4f8b\u5982\uff0c\u5982\u679c\u4e00\u4e2a\u7ffb\u8bd1\u7cfb\u7edf\u671f\u671b\u5728\u6e90\u7aef\u548c\u76ee\u6807\u7aef\u90fd\u63d0\u4f9b LID \u4fe1\u606f\uff0c\u90a3\u4e48\u8be5\u6a21\u578b\u4e0d\u592a\u53ef\u80fd\u5728\u8bad\u7ec3\u4e2d\u672a\u89c1\u8fc7\u7684\u65b9\u5411\u4e0a\u8868\u73b0\u51fa\u826f\u597d\u7684\u96f6\u6837\u672c\u6027\u80fd\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7684\u6311\u6218\u5728\u4e8e\u8bbe\u8ba1\u4e00\u4e2a\u5141\u8bb8\u4e00\u5b9a\u7a0b\u5ea6 LID \u4fe1\u606f\u7684\u7cfb\u7edf\uff0c\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u8db3\u591f\u901a\u7528\uff0c\u4ee5\u4fbf\u5728\u672a\u89c1\u7684\u65b9\u5411\u8fdb\u884c\u8bed\u97f3\u7ffb\u8bd1\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4ec5\u5305\u542b\u8981\u8f93\u51fa\u6587\u672c\uff08\u76ee\u6807\u7aef\uff09\u7684 LID \u4fe1\u606f\u7684\u7cfb\u7edf\u63d0\u793a\u3002\u8fd9\u4e9b\u63d0\u793a\u4e2d\u6ca1\u6709\u8bed\u97f3\u8f93\u5165\uff08\u6e90\u7aef\uff09\u7684 LID \u4fe1\u606f\uff0c\u8fd9\u4e5f\u53ef\u80fd\u4f7f\u5176\u80fd\u591f\u5904\u7406\u4ee3\u7801\u5207\u6362\u8bed\u97f3\u3002\u5bf9\u4e8e ASR\uff0c\u6211\u4eec\u4f7f\u7528\u4ee5\u4e0b\u7cfb\u7edf\u63d0\u793a\uff1a\u7528{\u8bed\u8a00}\u91cd\u590d\u6211\u7684\u8bdd\uff1a\uff0c\u5176\u4e2d{\u8bed\u8a00}\u6765\u81ea 34 \u79cd\u8bed\u8a00\u4e4b\u4e00\uff08\u82f1\u8bed\u3001\u6cd5\u8bed\u7b49\uff09\u3002\u5bf9\u4e8e\u8bed\u97f3\u7ffb\u8bd1\uff0c\u7cfb\u7edf\u63d0\u793a\u4e3a\uff1a\u201c\u5c06\u4ee5\u4e0b\u53e5\u5b50\u7ffb\u8bd1\u6210{\u8bed\u8a00}\uff1a\u201d\u3002\u8fd9\u79cd\u8bbe\u8ba1\u5df2\u88ab\u8bc1\u660e\u80fd\u591f\u6709\u6548\u5730\u63d0\u793a\u8bed\u8a00\u6a21\u578b\u4ee5\u671f\u671b\u7684\u8bed\u8a00\u8fdb\u884c\u54cd\u5e94\u3002\u6211\u4eec\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u76f8\u540c\u7684\u7cfb\u7edf\u63d0\u793a\u3002<\/p>\n<p>\u6211\u4eec\u4f7f\u7528\u81ea\u76d1\u7763\u5f0f BEST-RQ \u7b97\u6cd5\uff08Chiu \u7b49\uff0c2022\uff09\u5bf9\u8bed\u97f3\u8fdb\u884c\u9884\u8bad\u7ec3\u3002<\/p>\n<p id=\"pageNum-80a123d3-047c-4fe4-9e72-e2bc9fe07853\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\">\u7f16\u7801\u5668\u91c7\u7528\u957f\u5ea6\u4e3a 32 \u5e27\u7684\u63a9\u7801\uff0c\u5bf9\u8f93\u5165 mel \u8c31\u56fe\u7684\u6982\u7387\u4e3a 2.5%\u3002\u5982\u679c\u8bed\u97f3\u8bdd\u8bed\u8d85\u8fc7 60 \u79d2\uff0c\u6211\u4eec\u5c06\u968f\u673a\u88c1\u526a 6K \u5e27\uff0c\u5bf9\u5e94 60 \u79d2\u7684\u8bed\u97f3\u3002\u901a\u8fc7\u5806\u53e0 4 \u4e2a\u8fde\u7eed\u5e27\u3001\u5c06 320 \u7ef4\u5411\u91cf\u6295\u5f71\u5230 16 \u7ef4\u7a7a\u95f4\uff0c\u5e76\u5728 8192 \u4e2a\u5411\u91cf\u7684\u4ee3\u7801\u5e93\u5185\u4f7f\u7528\u4f59\u5f26\u76f8\u4f3c\u5ea6\u5ea6\u91cf\u8fdb\u884c\u6700\u8fd1\u90bb\u641c\u7d22\uff0c\u5bf9 mel \u8c31\u56fe\u7279\u5f81\u8fdb\u884c\u91cf\u5316\u3002\u4e3a\u4e86\u7a33\u5b9a\u9884\u8bad\u7ec3\uff0c\u6211\u4eec\u91c7\u7528 16 \u4e2a\u4e0d\u540c\u7684\u4ee3\u7801\u5e93\u3002\u6295\u5f71\u77e9\u9635\u548c\u4ee3\u7801\u5e93\u968f\u673a\u521d\u59cb\u5316\uff0c\u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e0d\u66f4\u65b0\u3002\u591a\u8f6f\u6700\u5927\u635f\u5931\u4ec5\u7528\u4e8e\u63a9\u7801\u5e27\uff0c\u4ee5\u63d0\u9ad8\u6548\u7387\u3002\u7f16\u7801\u5668\u7ecf\u8fc7 50 \u4e07\u6b65\u8bad\u7ec3\uff0c\u5168\u5c40\u6279\u5904\u7406\u5927\u5c0f\u4e3a 2048 \u4e2a\u8bed\u97f3\u3002<\/span><\/span><\/p>\n<p id=\"pageNum-d0c379d6-6ac9-48fd-9ea6-9c4874365197\" data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\"><span class=\"notranslate immersive-translate-target-wrapper\" lang=\"zh-CN\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-translation-pre-whitespace immersive-translate-target-translation-theme-none immersive-translate-target-translation-block-wrapper-theme-none immersive-translate-target-translation-block-wrapper\" data-immersive-translate-translation-element-mark=\"1\"><span class=\"notranslate immersive-translate-target-inner immersive-translate-target-translation-theme-none-inner\" data-immersive-translate-translation-element-mark=\"1\">\u76d1\u7763\u5fae\u8c03\u3002\u9884\u8bad\u7ec3\u8bed\u97f3\u7f16\u7801\u5668\u548c\u968f\u673a\u521d\u59cb\u5316\u7684\u9002\u914d\u5668\u5728\u76d1\u7763\u5fae\u8c03\u9636\u6bb5\u4e0e Llama 3 \u8054\u5408\u4f18\u5316\u3002\u8bed\u8a00\u6a21\u578b\u5728\u6b64\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u4e0d\u53d8\u3002\u8bad\u7ec3\u6570\u636e\u662f ASR\u3001AST \u548c\u5bf9\u8bdd\u6570\u636e\u7684\u6df7\u5408\u3002Llama 3 8B \u7684\u8bed\u97f3\u6a21\u578b\u7ecf\u8fc7 650K \u6b21\u66f4\u65b0\u8bad\u7ec3\uff0c\u4f7f\u7528\u5168\u5c40\u6279\u5927\u5c0f\u4e3a 512 \u4e2a\u8bdd\u8bed\u548c\u521d\u59cb\u5b66\u4e60\u7387\u4e3a 10\u3002Llama 3 70B \u7684\u8bed\u97f3\u6a21\u578b\u7ecf\u8fc7 600K \u6b21\u66f4\u65b0\u8bad\u7ec3\uff0c\u4f7f\u7528\u5168\u5c40\u6279\u5927\u5c0f\u4e3a 768 \u4e2a\u8bdd\u8bed\u548c\u521d\u59cb\u5b66\u4e60\u7387\u4e3a 4 \u00d7 10\u3002<\/span><\/span><\/span><\/p>\n<h4 data-immersive-translate-walked=\"a2eb285b-cee4-460c-a544-9c994d015215\">8.3.2 \u8bed\u97f3\u751f\u6210<\/h4>\n<p>\u4e3a\u4e86\u652f\u6301\u5b9e\u65f6\u5904\u7406\uff0c\u97f5\u5f8b\u6a21\u578b\u91c7\u7528\u4e86\u4e00\u4e2a\u524d\u77bb\u673a\u5236\uff0c\u8be5\u673a\u5236\u8003\u8651\u4e86\u56fa\u5b9a\u6570\u91cf\u7684\u672a\u6765\u97f3\u4f4d\u548c\u53ef\u53d8\u6570\u91cf\u7684\u672a\u6765\u6807\u8bb0\u3002 \u8fd9\u786e\u4fdd\u5728\u5904\u7406\u4f20\u5165\u6587\u672c\u65f6\u4fdd\u6301\u4e00\u81f4\u7684\u524d\u77bb\uff0c\u8fd9\u5bf9\u4e8e\u4f4e\u5ef6\u8fdf\u8bed\u97f3\u5408\u6210\u5e94\u7528\u7a0b\u5e8f\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p><strong>\u8bad\u7ec3:<\/strong>\u00a0\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u79cd\u5229\u7528\u56e0\u679c\u63a9\u7801\u7684\u52a8\u6001\u5bf9\u9f50\u7b56\u7565\uff0c\u4ee5\u4fc3\u8fdb\u8bed\u97f3\u5408\u6210\u7684\u6d41\u5f0f\u4f20\u8f93\u3002\u8be5\u7b56\u7565\u7ed3\u5408\u4e86\u97f5\u5f8b\u6a21\u578b\u4e2d\u7684\u524d\u77bb\u673a\u5236\uff0c\u7528\u4e8e\u56fa\u5b9a\u6570\u91cf\u7684\u672a\u6765\u97f3\u4f4d\u548c\u53ef\u53d8\u6570\u91cf\u7684\u672a\u6765\u6807\u8bb0\uff0c\u4e0e\u6587\u672c\u89c4\u8303\u5316\u8fc7\u7a0b\u4e2d\u7684\u5206\u5757\u8fc7\u7a0b\u76f8\u4e00\u81f4\uff08\u7b2c 8.1.2 \u8282\uff09\u3002<\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e2a\u97f3\u4f4d\uff0c\u6807\u8bb0\u524d\u77bb\u5305\u62ec\u5757\u5927\u5c0f\u5b9a\u4e49\u7684\u6700\u5927\u6807\u8bb0\u6570\uff0c\u5bfc\u81f4 Llama \u5d4c\u5165\u5177\u6709\u53ef\u53d8\u524d\u77bb\uff0c\u800c\u97f3\u4f4d\u5177\u6709\u56fa\u5b9a\u524d\u77bb\u3002<\/p>\n<p>Llama 3 \u5d4c\u5165\u6765\u81ea Llama 3 8B \u6a21\u578b\uff0c\u5728\u97f5\u5f8b\u6a21\u578b\u8bad\u7ec3\u671f\u95f4\u4fdd\u6301\u51bb\u7ed3\u3002\u8f93\u5165\u7535\u8bdd\u7387\u7279\u5f81\u5305\u62ec\u8bed\u8a00\u548c\u8bf4\u8bdd\u4eba\/\u98ce\u683c\u53ef\u63a7\u6027\u5143\u7d20\u3002\u6a21\u578b\u8bad\u7ec3\u4f7f\u7528\u6279\u91cf\u5927\u5c0f\u4e3a 1,024 \u4e2a\u8bed\u8c03\u7684 AdamW \u4f18\u5316\u5668\uff0c\u5b66\u4e60\u7387\u4e3a 9 \u00d7 10 -4\uff0c\u8bad\u7ec3\u8d85\u8fc7 100 \u4e07\u6b21\u66f4\u65b0\uff0c\u524d 3,000 \u6b21\u66f4\u65b0\u8fdb\u884c\u5b66\u4e60\u7387\u9884\u70ed\uff0c\u7136\u540e\u9075\u5faa\u4f59\u5f26\u8c03\u5ea6\u3002<\/p>\n<p><strong>\u63a8\u7406:<\/strong>\u00a0\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u4f7f\u7528\u76f8\u540c\u7684\u524d\u77bb\u673a\u5236\u548c\u56e0\u679c\u63a9\u7801\u7b56\u7565\uff0c\u4ee5\u786e\u4fdd\u8bad\u7ec3\u4e0e\u5b9e\u65f6\u5904\u7406\u4e4b\u95f4\u7684\u4e00\u81f4\u6027\u3002PM \u4ee5\u6d41\u5f0f\u65b9\u5f0f\u5904\u7406\u4f20\u5165\u6587\u672c\uff0c\u4e3a\u7535\u8bdd\u7387\u7279\u5f81\u9010\u4e2a\u66f4\u65b0\u8f93\u5165\u7535\u8bdd\uff0c\u4e3a\u6807\u8bb0\u7387\u7279\u5f81\u9010\u5757\u66f4\u65b0\u8f93\u5165\u3002\u53ea\u6709\u5f53\u8be5\u5757\u7684\u7b2c\u4e00\u4e2a\u7535\u8bdd\u5f53\u524d\u65f6\u624d\u66f4\u65b0\u65b0\u7684\u5757\u8f93\u5165\uff0c\u4ece\u800c\u5728\u8bad\u7ec3\u671f\u95f4\u4fdd\u6301\u5bf9\u9f50\u548c\u524d\u77bb\u3002<\/p>\n<p>\u4e3a\u4e86\u9884\u6d4b\u97f5\u5f8b\u76ee\u6807\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u4e00\u79cd\u5ef6\u8fdf\u6a21\u5f0f\u65b9\u6cd5\uff08Kharitonov \u7b49\u4eba\uff0c2021\uff09\uff0c\u8fd9\u589e\u5f3a\u4e86\u6a21\u578b\u6355\u83b7\u548c\u590d\u5236\u957f\u7a0b\u97f5\u5f8b\u4f9d\u8d56\u5173\u7cfb\u7684\u80fd\u529b\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u52a9\u4e8e\u5408\u6210\u7684\u8bed\u97f3\u7684\u81ea\u7136\u5ea6\u548c\u8868\u8fbe\u529b\uff0c\u786e\u4fdd\u4f4e\u5ef6\u8fdf\u548c\u9ad8\u8d28\u91cf\u8f93\u51fa\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>8.4 \u8bed\u97f3\u7406\u89e3\u7ed3\u679c<\/h3>\n<p>\u6211\u4eec\u8bc4\u4f30\u4e86 Llama 3 \u8bed\u97f3\u754c\u9762\u7684\u8bed\u97f3\u7406\u89e3\u80fd\u529b\uff0c\u5305\u62ec\u4e09\u4e2a\u4efb\u52a1\uff1a(1) \u81ea\u52a8\u8bed\u97f3\u8bc6\u522b (2) \u8bed\u97f3\u7ffb\u8bd1 (3) \u8bed\u97f3\u95ee\u7b54\u3002\u6211\u4eec\u5c06 Llama 3 \u7684\u8bed\u97f3\u754c\u9762\u6027\u80fd\u4e0e\u4e09\u79cd\u6700\u5148\u8fdb\u7684\u8bed\u97f3\u7406\u89e3\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\uff1aWhisper\uff08Radford \u7b49\u4eba\uff0c2023 \u5e74\uff09\u3001SeamlessM4T\uff08Barrault \u7b49\u4eba\uff0c2023 \u5e74\uff09\u548c Gemini\u3002\u5728\u6240\u6709\u8bc4\u4f30\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u8d2a\u5a6a\u641c\u7d22\u6765\u9884\u6d4b Llama 3 \u7684\u6807\u8bb0\u3002<\/p>\n<p><strong>\u8bed\u97f3\u8bc6\u522b\u3002<\/strong>\u00a0\u6211\u4eec\u5728 Multilingual LibriSpeech (MLS; Pratap \u7b49\u4eba\uff0c2020 \u5e74)\u3001LibriSpeech (Panayotov \u7b49\u4eba\uff0c2015 \u5e74)\u3001VoxPopuli (Wang \u7b49\u4eba\uff0c2021a) \u548c FLEURS \u591a\u8bed\u8a00\u6570\u636e\u96c6\u7684\u5b50\u96c6\uff08Conneau \u7b49\u4eba\uff0c2023 \u5e74\uff09\u4e0a\u7684\u82f1\u6587\u6570\u636e\u96c6\u4e0a\u8bc4\u4f30\u4e86 ASR \u6027\u80fd\u3002\u5728\u8bc4\u4f30\u4e2d\uff0c\u4f7f\u7528 <a href=\"https:\/\/www.kdjingpai.com\/pt\/crisperwhisper\/\">Whisper<\/a> \u6587\u672c\u6807\u51c6\u5316\u7a0b\u5e8f\u5bf9\u89e3\u7801\u7ed3\u679c\u8fdb\u884c\u540e\u5904\u7406\uff0c\u4ee5\u786e\u4fdd\u4e0e\u5176\u4ed6\u6a21\u578b\u62a5\u544a\u7684\u7ed3\u679c\u4e00\u81f4\u3002\u5728\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u6211\u4eec\u5728\u8fd9\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u7684\u6807\u51c6\u6d4b\u8bd5\u96c6\u4e0a\u6d4b\u91cf Llama 3 \u8bed\u97f3\u754c\u9762\u7684\u5355\u8bcd\u9519\u8bef\u7387\uff0c\u9664\u4e86\u4e2d\u6587\u3001\u65e5\u8bed\u3001\u97e9\u8bed\u548c\u6cf0\u8bed\uff0c\u6211\u4eec\u62a5\u544a\u4e86\u5b57\u7b26\u9519\u8bef\u7387\u3002<\/p>\n<p>\u8868 31 \u663e\u793a\u4e86 ASR \u8bc4\u4f30\u7ed3\u679c\u3002\u5b83\u8bc1\u660e\u4e86 Llama 3\uff08\u4ee5\u53ca\u66f4\u5e7f\u6cdb\u7684\u591a\u6a21\u6001\u57fa\u7840\u6a21\u578b\uff09\u5728\u8bed\u97f3\u8bc6\u522b\u4efb\u52a1\u4e0a\u7684\u5f3a\u52b2\u6027\u80fd\uff1a\u6211\u4eec\u7684\u6a21\u578b\u5728\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u90fd\u4f18\u4e8e Whisper20 \u548c SeamlessM4T \u7b49\u4e13\u95e8\u9488\u5bf9\u8bed\u97f3\u7684\u6a21\u578b\u3002\u5728 MLS \u82f1\u6587\u65b9\u9762\uff0cLlama 3 \u7684\u8868\u73b0\u4e0e Gemini \u76f8\u4f3c\u3002<\/p>\n<p><strong>\u8bed\u97f3\u7ffb\u8bd1\u3002<\/strong>\u00a0\u6211\u4eec\u8fd8\u8bc4\u4f30\u4e86\u6211\u4eec\u7684\u6a21\u578b\u5728\u8bed\u97f3\u7ffb\u8bd1\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\uff0c\u5176\u4e2d\u6a21\u578b\u88ab\u8981\u6c42\u5c06\u975e\u82f1\u8bed\u8bed\u97f3\u7ffb\u8bd1\u6210\u82f1\u8bed\u6587\u672c\u3002\u6211\u4eec\u5728\u8fd9\u4e9b\u8bc4\u4f30\u4e2d\u4f7f\u7528 FLEURS \u548c Covost 2 (Wang \u7b49\u4eba\uff0c2021b) \u6570\u636e\u96c6\uff0c\u5e76\u6d4b\u91cf\u7ffb\u8bd1\u82f1\u6587\u7684 BLEU \u5206\u6570\u3002\u8868 32 \u5c55\u793a\u4e86\u8fd9\u4e9b\u5b9e\u9a8c\u7684\u7ed3\u679c\u3002\u6211\u4eec\u7684\u6a21\u578b\u5728\u8bed\u97f3\u7ffb\u8bd1\u65b9\u9762\u7684\u6027\u80fd\u7a81\u663e\u4e86\u591a\u6a21\u6001\u57fa\u7840\u6a21\u578b\u5728\u8bed\u97f3\u7ffb\u8bd1\u7b49\u4efb\u52a1\u4e2d\u7684\u4f18\u52bf\u3002<\/p>\n<p><strong>\u8bed\u97f3\u95ee\u7b54\u3002<\/strong>\u00a0Llama 3 \u7684\u8bed\u97f3\u754c\u9762\u5c55\u793a\u51fa\u60ca\u4eba\u7684\u95ee\u9898\u56de\u7b54\u80fd\u529b\u3002\u8be5\u6a21\u578b\u53ef\u4ee5\u6beb\u4e0d\u8d39\u529b\u5730\u7406\u89e3\u4ee3\u7801\u5207\u6362\u7684\u8bed\u97f3\uff0c\u800c\u65e0\u9700\u4e8b\u5148\u63a5\u89e6\u6b64\u7c7b\u6570\u636e\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5c3d\u7ba1\u8be5\u6a21\u578b\u4ec5\u5728\u5355\u8f6e\u5bf9\u8bdd\u4e0a\u8fdb\u884c\u4e86\u8bad\u7ec3\uff0c\u4f46\u5b83\u80fd\u591f\u8fdb\u884c\u6269\u5c55\u4e14\u8fde\u8d2f\u7684\u591a\u8f6e\u5bf9\u8bdd\u4f1a\u8bdd\u3002\u56fe 30 \u5c55\u793a\u4e86\u4e00\u4e9b\u7a81\u663e\u8fd9\u4e9b\u591a\u8bed\u8a00\u548c\u591a\u8f6e\u80fd\u529b\u7684\u4f8b\u5b50\u3002<\/p>\n<p><strong>\u5b89\u5168\u6027\u3002<\/strong>\u00a0\u6211\u4eec\u5728 MuTox (Costa-juss\u00e0 \u7b49\u4eba\uff0c2023 \u5e74) \u4e0a\u8bc4\u4f30\u4e86\u6211\u4eec\u8bed\u97f3\u6a21\u578b\u7684\u5b89\u5168\u6027\u80fd\uff0c\u8fd9\u662f\u4e00\u4e2a\u5305\u542b 20,000 \u4e2a\u82f1\u8bed\u548c\u897f\u73ed\u7259\u8bed\u8bed\u6bb5\u4ee5\u53ca 4,000 \u4e2a\u5176\u4ed6 19 \u79cd\u8bed\u8a00\u8bed\u6bb5\u7684\u7528\u4e8e\u591a\u8bed\u8a00\u57fa\u4e8e\u97f3\u9891\u7684\u6570\u636e\u96c6\uff0c\u6bcf\u4e2a\u8bed\u6bb5\u90fd\u6709\u6bd2\u6027\u6807\u7b7e\u3002\u5c06\u97f3\u9891\u4f5c\u4e3a\u8f93\u5165\u4f20\u9012\u7ed9\u6a21\u578b\uff0c\u5e76\u5728\u6e05\u9664\u4e00\u4e9b\u7279\u6b8a\u5b57\u7b26\u540e\u8bc4\u4f30\u8f93\u51fa\u7684\u6bd2\u6027\u3002\u6211\u4eec\u5c06 MuTox \u5206\u7c7b\u5668\uff08Costa-juss\u00e0 \u7b49\u4eba\uff0c2023 \u5e74\uff09\u5e94\u7528\u4e8e Gemini 1.5 Pro \u5e76\u6bd4\u8f83\u7ed3\u679c\u3002\u6211\u4eec\u8bc4\u4f30\u4e86\u5f53\u8f93\u5165\u63d0\u793a\u5b89\u5168\u800c\u8f93\u51fa\u6709\u6bd2\u65f6\u6dfb\u52a0\u6bd2\u6027 (AT) \u7684\u767e\u5206\u6bd4\uff0c\u4ee5\u53ca\u5f53\u8f93\u5165\u63d0\u793a\u6709\u6bd2\u800c\u7b54\u6848\u5b89\u5168\u65f6\u4e22\u5931\u6bd2\u6027 (LT) \u7684\u767e\u5206\u6bd4\u3002\u8868 33 \u663e\u793a\u4e86\u82f1\u8bed\u7684\u7d50\u679c\u4ee5\u53ca\u6211\u4eec\u5728\u6240\u6709 21 \u79cd\u8bed\u8a00\u4e0a\u7684\u5e73\u5747\u7ed3\u679c\u3002\u6dfb\u52a0\u6bd2\u6027\u767e\u5206\u6bd4\u975e\u5e38\u4f4e\uff1a\u6211\u4eec\u7684\u8bed\u97f3\u6a21\u578b\u5728\u82f1\u8bed\u65b9\u9762\u5177\u6709\u6700\u4f4e\u7684\u6dfb\u52a0\u6bd2\u6027\u767e\u5206\u6bd4\uff0c\u4e0d\u5230 1%\u3002\u5b83\u53bb\u9664\u7684\u6bd2\u6027\u8fdc\u591a\u4e8e\u6dfb\u52a0\u7684\u6bd2\u6027\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3491\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/735181bd751d4fe.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"722\" height=\"295\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/735181bd751d4fe.png 1417w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/735181bd751d4fe-300x123.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/735181bd751d4fe-1024x418.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/735181bd751d4fe-768x314.png 768w\" sizes=\"auto, (max-width: 722px) 100vw, 722px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>8.5 \u8bed\u97f3\u751f\u6210\u7ed3\u679c<\/h3>\n<p>\u5728\u8bed\u97f3\u751f\u6210\u65b9\u9762\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u8bc4\u4f30\u4f7f\u7528 Llama 3 \u5411\u91cf\u8fdb\u884c\u6587\u672c\u89c4\u8303\u5316\u548c\u97f5\u5f8b\u5efa\u6a21\u4efb\u52a1\u7684\u57fa\u4e8e\u6807\u8bb0\u6d41\u5f0f\u8f93\u5165\u6a21\u578b\u7684\u8d28\u91cf\u3002\u8bc4\u4f30\u91cd\u70b9\u5728\u4e8e\u4e0e\u4e0d\u5c06 Llama 3 \u5411\u91cf\u4f5c\u4e3a\u989d\u5916\u8f93\u5165\u7684\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\u3002<\/p>\n<p><strong>\u6587\u672c\u89c4\u8303\u5316:<\/strong>\u00a0\u4e3a\u4e86\u8861\u91cf Llama 3 \u5411\u91cf\u7684\u5f71\u54cd\uff0c\u6211\u4eec\u5c1d\u8bd5\u6539\u53d8\u6a21\u578b\u4f7f\u7528\u7684\u53f3\u4fa7\u4e0a\u4e0b\u6587\u91cf\u3002\u6211\u4eec\u4f7f\u7528 3 \u4e2a\u6587\u672c\u89c4\u8303\u5316\uff08TN\uff09\u6807\u8bb0\uff08\u4ee5 Unicode \u7c7b\u522b\u5206\u9694\uff09\u7684\u53f3\u4fa7\u4e0a\u4e0b\u6587\u8bad\u7ec3\u4e86\u6a21\u578b\u3002\u8fd9\u4e2a\u6a21\u578b\u4e0e\u4e0d\u4f7f\u7528 Llama 3 \u5411\u91cf\u3001\u4f7f\u7528 3 \u4e2a\u6807\u8bb0\u53f3\u4fa7\u4e0a\u4e0b\u6587\u6216\u5b8c\u6574\u53cc\u5411\u4e0a\u4e0b\u6587\u7684\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u6b63\u5982\u9884\u671f\u7684\u90a3\u6837\uff0c\u8868 34 \u663e\u793a\u4f7f\u7528\u5b8c\u6574\u7684\u53f3\u4fa7\u4e0a\u4e0b\u6587\u53ef\u4ee5\u63d0\u9ad8\u4e0d\u4f7f\u7528 Llama 3 \u5411\u91cf\u7684\u6a21\u578b\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u5305\u542b Llama 3 \u5411\u91cf\u7684\u6a21\u578b\u4f18\u4e8e\u6240\u6709\u5176\u4ed6\u6a21\u578b\uff0c\u4ece\u800c\u5b9e\u73b0\u4e86\u65e0\u9700\u4f9d\u8d56\u8f93\u5165\u4e2d\u7684\u957f\u4e0a\u4e0b\u6587\u5373\u53ef\u8fdb\u884c\u6807\u8bb0\u7387\u8f93\u5165\/\u8f93\u51fa\u6d41\u5f0f\u4f20\u8f93\u3002\u6211\u4eec\u6bd4\u8f83\u4e86\u6709\u6216\u6ca1\u6709 Llama 3 8B \u5411\u91cf\u4ee5\u53ca\u4f7f\u7528\u4e0d\u540c\u53f3\u4fa7\u4e0a\u4e0b\u6587\u503c\u7684\u6a21\u578b\u3002<\/p>\n<p><strong>\u97f5\u5f8b\u5efa\u6a21:<\/strong>\u00a0\u4e3a\u4e86\u8bc4\u4f30\u6211\u4eec\u7684\u97f5\u5f8b\u6a21\u578b\uff08PM\uff09\u4e0e Llama 3 8B \u7684\u6027\u80fd\uff0c\u6211\u4eec\u8fdb\u884c\u4e86\u4e24\u7ec4\u4eba\u7c7b\u8bc4\u4ef7\uff0c\u6bd4\u8f83\u4e86\u6709\u548c\u6ca1\u6709 Llama 3 \u5411\u91cf\u7684\u6a21\u578b\u3002\u8bc4\u5206\u8005\u8046\u542c\u6765\u81ea\u4e0d\u540c\u6a21\u578b\u7684\u6837\u672c\uff0c\u5e76\u8868\u793a\u4ed6\u4eec\u7684\u504f\u597d\u3002<\/p>\n<p>\u8981\u751f\u6210\u6700\u7ec8\u7684\u8bed\u97f3\u6ce2\u5f62\uff0c\u6211\u4eec\u4f7f\u7528\u4e86\u4e00\u79cd\u57fa\u4e8e Transformer \u7684\u5185\u90e8\u58f0\u5b66\u6a21\u578b\uff08Wu \u7b49\u4eba\uff0c2021\uff09\uff0c\u5b83\u9884\u6d4b\u8c31\u7279\u5f81\uff0c\u5e76\u4f7f\u7528 WaveRNN \u795e\u7ecf\u58f0\u7801\u5668\uff08Kalchbrenner \u7b49\u4eba\uff0c2018\uff09\u751f\u6210\u6700\u7ec8\u7684\u8bed\u97f3\u6ce2\u5f62\u3002<\/p>\n<p>\u9996\u5148\uff0c\u6211\u4eec\u5c06\u76f4\u63a5\u4e0e\u4e0d\u4f7f\u7528 Llama 3 \u5411\u91cf\u7684\u6d41\u5f0f\u57fa\u51c6\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\u3002\u5728\u7b2c\u4e8c\u9879\u6d4b\u8bd5\u4e2d\uff0c\u5c06 Llama 3 8B PM \u4e0e\u4e0d\u4f7f\u7528 Llama 3 \u5411\u91cf\u7684\u975e\u6d41\u5f0f\u57fa\u51c6\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u5982\u8868 35 \u6240\u793a\uff0cLlama 3 8B PM \u88ab\u4f18\u9009\u4e86 60% \u7684\u65f6\u95f4\uff08\u4e0e\u6d41\u5f0f\u57fa\u51c6\u76f8\u6bd4\uff09\uff0c\u5e76\u4e14 63.6% \u7684\u65f6\u95f4\u88ab\u4f18\u9009\uff08\u4e0e\u975e\u6d41\u5f0f\u57fa\u51c6\u76f8\u6bd4\uff09\uff0c\u8fd9\u8868\u660e\u611f\u77e5\u8d28\u91cf\u6709\u663e\u8457\u63d0\u9ad8\u3002 Llama 3 8B PM \u7684\u5173\u952e\u4f18\u52bf\u5728\u4e8e\u5176\u57fa\u4e8e\u6807\u8bb0\u7684\u6d41\u5f0f\u4f20\u8f93\u80fd\u529b\uff08\u7b2c 8.2.2 \u8282\uff09\uff0c\u5b83\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u4f4e\u5ef6\u8fdf\u3002\u8fd9\u964d\u4f4e\u4e86\u6a21\u578b\u7684\u524d\u77bb\u9700\u6c42\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5b9e\u73b0\u66f4\u5177\u54cd\u5e94\u6027\u548c\u5b9e\u65f6\u6027\u7684\u8bed\u97f3\u5408\u6210\uff0c\u76f8\u6bd4\u4e8e\u975e\u6d41\u5f0f\u57fa\u51c6\u6a21\u578b\u3002\u603b\u4f53\u800c\u8a00\uff0cLlama 3 8B \u97f5\u5f8b\u6a21\u578b\u59cb\u7ec8\u4f18\u4e8e\u57fa\u51c6\u6a21\u578b\uff0c\u8bc1\u660e\u4e86\u5176\u5728\u63d0\u9ad8\u5408\u6210\u8bed\u97f3\u7684\u81ea\u7136\u5ea6\u548c\u8868\u73b0\u529b\u65b9\u9762\u7684\u6709\u6548\u6027\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3492\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/5c6dac5461f1221.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"747\" height=\"398\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/5c6dac5461f1221.png 1229w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/5c6dac5461f1221-300x160.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/5c6dac5461f1221-1024x546.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/5c6dac5461f1221-768x409.png 768w\" sizes=\"auto, (max-width: 747px) 100vw, 747px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3493\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/ee250bd81625101.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"811\" height=\"172\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/ee250bd81625101.png 1210w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/ee250bd81625101-300x64.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/ee250bd81625101-1024x217.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/ee250bd81625101-768x163.png 768w\" sizes=\"auto, (max-width: 811px) 100vw, 811px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3494\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0f9860f9385dead.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"413\" height=\"207\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0f9860f9385dead.png 562w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/0f9860f9385dead-300x151.png 300w\" sizes=\"auto, (max-width: 413px) 100vw, 413px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-3495\" title=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/75a121feea55cb2.png\" alt=\"Llama 3\uff1a\u4e00\u4e2a\u591a\u529f\u80fd\u3001\u5f00\u6e90\u7684 AI \u6a21\u578b\u7cfb\u5217-1\" width=\"804\" height=\"154\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/75a121feea55cb2.png 1214w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/75a121feea55cb2-300x57.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/75a121feea55cb2-1024x196.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/07\/75a121feea55cb2-768x147.png 768w\" sizes=\"auto, (max-width: 804px) 100vw, 804px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>9 \u76f8\u5173\u5de5\u4f5c<\/h2>\n<p>&nbsp;<\/p>\n<p>Llama 3 \u7684\u5f00\u53d1\u5efa\u7acb\u5728\u5bf9\u8bed\u8a00\u3001\u56fe\u50cf\u3001\u89c6\u9891\u548c\u8bed\u97f3\u57fa\u7840\u6a21\u578b\u7684\u5927\u91cf\u5148\u524d\u7814\u7a76\u4e4b\u4e0a\u3002\u672c\u6587\u8303\u56f4\u4e0d\u5305\u62ec\u5bf9\u8fd9\u4e9b\u5de5\u4f5c\u7684\u5168\u9762\u6982\u8ff0\uff1b\u6211\u4eec\u5efa\u8bae\u8bfb\u8005\u53c2\u8003 Bordes \u7b49\u4eba (2024)\uff1bMadan \u7b49\u4eba (2024)\uff1bZhao \u7b49\u4eba (2023a) \u4ee5\u83b7\u53d6\u8fd9\u6837\u7684\u6982\u8ff0\u3002\u4e0b\u9762\uff0c\u6211\u4eec\u5c06\u7b80\u8981\u6982\u8ff0\u76f4\u63a5\u5f71\u54cd Llama 3 \u5f00\u53d1\u7684\u5f00\u521b\u6027\u8457\u4f5c\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>9.1 \u8bed\u8a00<\/h3>\n<p><strong>\u89c4\u6a21:<\/strong>\u00a0Llama 3 \u9075\u5faa\u7740\u5c06\u7b80\u5355\u65b9\u6cd5\u5e94\u7528\u4e8e\u4e0d\u65ad\u589e\u5927\u7684\u89c4\u6a21\u7684\u6301\u4e45\u8d8b\u52bf\uff0c\u8fd9\u662f\u57fa\u7840\u6a21\u578b\u7684\u7279\u5f81\u3002\u6539\u8fdb\u662f\u7531\u8ba1\u7b97\u80fd\u529b\u7684\u63d0\u9ad8\u548c\u6570\u636e\u8d28\u91cf\u7684\u63d0\u5347\u9a71\u52a8\u7684\uff0c\u5176\u4e2d405B \u6a21\u578b\u4f7f\u7528\u7684\u9884\u8bad\u7ec3\u8ba1\u7b97\u9884\u7b97\u51e0\u4e4e\u662fLlama 2 70B \u7684\u4e94\u5341\u500d\u3002\u5c3d\u7ba1\u6211\u4eec\u7684\u6700\u5927 Llama 3 \u5305\u542b405B \u4e2a\u53c2\u6570\uff0c\u4f46\u5b9e\u9645\u4e0a\u5176\u53c2\u6570\u6570\u91cf\u5c11\u4e8e\u65e9\u671f\u4e14\u6027\u80fd\u66f4\u5dee\u7684\u6a21\u578b\uff0c\u4f8b\u5982 PALM (Chowdhery \u7b49\u4eba\uff0c2023)\uff0c\u8fd9\u662f\u7531\u4e8e\u5bf9\u7f29\u653e\u5b9a\u5f8b\uff08Kaplan \u7b49\u4eba\uff0c2020\uff1bHoffmann \u7b49\u4eba\uff0c2022\uff09\u6709\u4e86\u66f4\u597d\u7684\u7406\u89e3\u3002\u5176\u4ed6\u524d\u6cbf\u6a21\u578b\u7684\u5927\u5c0f\uff0c\u5982 Claude 3 \u6216 GPT 4\uff08OpenAI\uff0c2023a\uff09\uff0c\u516c\u5f00\u4fe1\u606f\u6709\u9650\uff0c\u4f46\u603b\u4f53\u6027\u80fd\u76f8\u5f53\u3002<\/p>\n<p><strong>\u5c0f\u89c4\u6a21\u6a21\u578b:<\/strong>\u00a0\u5c0f\u89c4\u6a21\u6a21\u578b\u7684\u53d1\u5c55\u4e0e\u5927\u89c4\u6a21\u6a21\u578b\u7684\u53d1\u5c55\u540c\u6b65\u8fdb\u884c\u4e86\u3002\u53c2\u6570\u8f83\u5c11\u7684\u6a21\u578b\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u63a8\u7406\u6210\u672c\u5e76\u7b80\u5316\u90e8\u7f72\uff08Mehta \u7b49\u4eba\uff0c2024\uff1bTeam \u7b49\u4eba\uff0c2024\uff09\u3002\u8f83\u5c0f\u7684 Llama 3 \u6a21\u578b\u901a\u8fc7\u8fdc\u8fdc\u8d85\u8fc7\u8ba1\u7b97\u6700\u4f18\u8bad\u7ec3\u70b9\u6765\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u6709\u6548\u5730\u5c06\u8bad\u7ec3\u8ba1\u7b97\u6743\u8861\u4e3a\u63a8\u7406\u6548\u7387\u3002\u53e6\u4e00\u79cd\u9014\u5f84\u662f\u5c06\u8f83\u5927\u6a21\u578b\u84b8\u998f\u6210\u8f83\u5c0f\u6a21\u578b\uff0c\u5982 Phi\uff08Abdin \u7b49\u4eba\uff0c2024\uff09\u3002<\/p>\n<p><strong>\u67b6\u6784:<\/strong>\u00a0\u4e0e Llama 2 \u76f8\u6bd4\uff0cLlama 3 \u5728\u67b6\u6784\u4e0a\u505a\u4e86\u6700\u5c0f\u7684\u4fee\u6539\uff0c\u4f46\u5176\u4ed6\u6700\u8fd1\u7684\u57fa\u7840\u6a21\u578b\u63a2\u7d22\u4e86\u5176\u4ed6\u8bbe\u8ba1\u3002\u6700\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4e13\u5bb6\u6df7\u5408\u67b6\u6784\uff08Shazeer \u7b49\u4eba\uff0c2017\uff1bLewis \u7b49\u4eba\uff0c2021\uff1bFedus \u7b49\u4eba\uff0c2022\uff1bZhou \u7b49\u4eba\uff0c2022\uff09\u53ef\u4ee5\u4f5c\u4e3a\u6709\u6548\u5730\u63d0\u9ad8\u6a21\u578b\u5bb9\u91cf\u7684\u4e00\u79cd\u65b9\u6cd5\uff0c\u4f8b\u5982\u5728 Mixtral\uff08Jiang \u7b49\u4eba\uff0c2024\uff09\u548c Arctic\uff08Snowflake\uff0c2024\uff09\u4e2d\u3002Llama 3 \u7684\u6027\u80fd\u4f18\u4e8e\u8fd9\u4e9b\u6a21\u578b\uff0c\u8fd9\u8868\u660e\u5bc6\u96c6\u67b6\u6784\u5e76\u4e0d\u662f\u9650\u5236\u56e0\u7d20\uff0c\u4f46\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u6548\u7387\u3001\u4ee5\u53ca\u5927\u89c4\u6a21\u4e0b\u7684\u6a21\u578b\u7a33\u5b9a\u6027\u65b9\u9762\u4ecd\u7136\u5b58\u5728\u8bb8\u591a\u6743\u8861\u3002<\/p>\n<p><strong>\u5f00\u6e90:<\/strong>\u00a0\u5f00\u6e90\u57fa\u7840\u6a21\u578b\u5728\u8fc7\u53bb\u4e00\u5e74\u4e2d\u8fc5\u901f\u53d1\u5c55\uff0c\u5176\u4e2d Llama3-405B \u73b0\u5728\u4e0e\u5f53\u524d\u95ed\u6e90\u7684\u6700\u5148\u8fdb\u6c34\u5e73\u76f8\u5f53\u3002\u6700\u8fd1\u5f00\u53d1\u4e86\u8bb8\u591a\u6a21\u578b\u7cfb\u5217\uff0c\u5305\u62ec Mistral\uff08Jiang \u7b49\u4eba\uff0c2023\uff09\u3001Falcon\uff08Almazrouei \u7b49\u4eba\uff0c2023\uff09\u3001MPT\uff08Databricks\uff0c2024\uff09\u3001Pythia\uff08Biderman \u7b49\u4eba\uff0c2023\uff09\u3001Arctic\uff08Snowflake\uff0c2024\uff09\u3001OpenELM\uff08Mehta \u7b49\u4eba\uff0c2024\uff09\u3001OLMo\uff08Groeneveld \u7b49\u4eba\uff0c2024\uff09\u3001StableLM\uff08Bellagente \u7b49\u4eba\uff0c2024\uff09\u3001OpenLLaMA\uff08Geng \u548c Liu\uff0c2023\uff09\u3001Qwen\uff08Bai \u7b49\u4eba\uff0c2023\uff09\u3001Gemma\uff08Team \u7b49\u4eba\uff0c2024\uff09\u3001Grok\uff08XAI\uff0c2024\uff09\u548c Phi\uff08Abdin \u7b49\u4eba\uff0c2024\uff09\u3002<\/p>\n<p><strong>\u540e\u8bad\u7ec3:<\/strong>\u00a0Llama 3 \u7684\u540e\u8bad\u7ec3\u9075\u5faa\u4e86\u65e2\u5b9a\u7684\u6307\u4ee4\u8c03\u6574\u7b56\u7565\uff08Chung \u7b49\u4eba\uff0c2022\uff1bOuyang \u7b49\u4eba\uff0c2022\uff09\uff0c\u63a5\u7740\u662f\u4e0e\u4eba\u7c7b\u53cd\u9988\u8fdb\u884c\u5bf9\u9f50\uff08Kaufmann \u7b49\u4eba\uff0c2023\uff09\u3002\u5c3d\u7ba1\u4e00\u4e9b\u7814\u7a76\u8868\u660e\u8f7b\u91cf\u7ea7\u5bf9\u9f50\u7a0b\u5e8f\u7684\u6548\u679c\u51fa\u4e4e\u610f\u6599\uff08Zhou \u7b49\u4eba\uff0c2024\uff09\uff0c\u4f46 Llama 3 \u4f7f\u7528\u6570\u767e\u4e07\u4e2a\u4eba\u7c7b\u6307\u4ee4\u548c\u504f\u597d\u5224\u65ad\u6765\u6539\u8fdb\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u5305\u62ec\u62d2\u7edd\u91c7\u6837\uff08Bai \u7b49\u4eba\uff0c2022\uff09\u3001\u76d1\u7763\u5fae\u8c03\uff08Sanh \u7b49\u4eba\uff0c2022\uff09\u548c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08Rafailov \u7b49\u4eba\uff0c2023\uff09\u3002\u4e3a\u4e86\u7b56\u5212\u8fd9\u4e9b\u6307\u4ee4\u548c\u504f\u597d\u793a\u4f8b\uff0c\u6211\u4eec\u90e8\u7f72\u4e86\u66f4\u65e9\u7248\u672c\u7684 Llama 3 \u6765\u8fc7\u6ee4\uff08Liu \u7b49\u4eba\uff0c2024c\uff09\u3001\u91cd\u5199\uff08Pan \u7b49\u4eba\uff0c2024\uff09\u6216\u751f\u6210\u63d0\u793a\u548c\u54cd\u5e94\uff08Liu \u7b49\u4eba\uff0c2024b\uff09\uff0c\u5e76\u5c06\u8fd9\u4e9b\u6280\u672f\u901a\u8fc7\u591a\u8f6e\u540e\u8bad\u7ec3\u5e94\u7528\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>9.2 \u591a\u6a21\u6001\u6027<\/h3>\n<p>\u6211\u4eec\u7684 Llama 3 \u591a\u6a21\u6001\u80fd\u529b\u5b9e\u9a8c\u662f\u5173\u4e8e\u8054\u5408\u5efa\u6a21\u591a\u4e2a\u6a21\u6001\u7684\u57fa\u7840\u6a21\u578b\u957f\u671f\u7814\u7a76\u7684\u4e00\u90e8\u5206\u3002\u6211\u4eec\u7684 Llama 3 \u65b9\u6cd5\u7ed3\u5408\u4e86\u8bb8\u591a\u8bba\u6587\u4e2d\u7684\u60f3\u6cd5\uff0c\u53d6\u5f97\u4e86\u4e0e Gemini 1.0 Ultra\uff08\u8c37\u6b4c\uff0c2023 \u5e74\uff09\u548c GPT-4 Vision\uff08OpenAI\uff0c2023b\uff09\u76f8\u5f53\u7684\u7ed3\u679c\uff1b\u8bf7\u53c2\u89c1\u7b2c 7.6 \u8282\u3002<\/p>\n<p><strong>\u89c6\u9891<\/strong>\uff1a\u5c3d\u7ba1\u8d8a\u6765\u8d8a\u591a\u7684\u57fa\u7840\u6a21\u578b\u652f\u6301\u89c6\u9891\u8f93\u5165\uff08\u8c37\u6b4c\uff0c2023 \u5e74\uff1b OpenAI\uff0c2023b\uff09\uff0c\u4f46\u5173\u4e8e\u89c6\u9891\u548c\u8bed\u8a00\u8054\u5408\u5efa\u6a21\u7684\u7814\u7a76\u5e76\u4e0d\u591a\u3002\u7c7b\u4f3c\u4e8e Llama 3\uff0c\u76ee\u524d\u5927\u591a\u6570\u7814\u7a76\u90fd\u91c7\u7528\u9002\u914d\u5668\u65b9\u6cd5\u6765\u5bf9\u9f50\u89c6\u9891\u548c\u8bed\u8a00\u8868\u793a\uff0c\u5e76\u89e3\u5f00\u6709\u5173\u89c6\u9891\u7684\u95ee\u7b54\u548c\u63a8\u7406\uff08Lin \u7b49\u4eba\uff0c2023 \u5e74\uff1bLi \u7b49\u4eba\uff0c2023a\uff1bMaaz \u7b49\u4eba\uff0c2024 \u5e74\uff1bZhang \u7b49\u4eba\uff0c2023 \u5e74\uff1bZhao \u7b49\u4eba\uff0c2022 \u5e74\uff09\u3002\u6211\u4eec\u53d1\u73b0\u6b64\u7c7b\u65b9\u6cd5\u4ea7\u751f\u7684\u7ed3\u679c\u4e0e\u6700\u5148\u8fdb\u7684\u7ed3\u679c\u5177\u6709\u7ade\u4e89\u529b\uff1b\u8bf7\u53c2\u89c1\u7b2c 7.7 \u8282\u3002<\/p>\n<p><strong>\u8bed\u97f3<\/strong>\uff1a\u6211\u4eec\u7684\u5de5\u4f5c\u4e5f\u878d\u5165\u4e86\u4e00\u9879\u66f4\u5927\u7684\u5de5\u4f5c\uff0c\u8be5\u5de5\u4f5c\u7ed3\u5408\u4e86\u8bed\u8a00\u548c\u8bed\u97f3\u5efa\u6a21\u3002\u65e9\u671f\u7684\u6587\u672c\u548c\u8bed\u97f3\u8054\u5408\u6a21\u578b\u5305\u62ec AudioPaLM\uff08Rubenstein \u7b49\u4eba\uff0c2023 \u5e74\uff09\u3001VioLA\uff08Wang \u7b49\u4eba\uff0c2023b\uff09\u3001VoxtLM Maiti \u7b49\u4eba\uff082023 \u5e74\uff09, SUTLM\uff08Chou \u7b49\u4eba\uff0c2023 \u5e74\uff09\u548c Spirit-LM\uff08Nguyen \u7b49\u4eba\uff0c2024 \u5e74\uff09\u3002\u6211\u4eec\u7684\u5de5\u4f5c\u57fa\u4e8e Fathullah \u7b49\u4eba\uff082024 \u5e74\uff09\u7b49\u5148\u524d\u7ec4\u5408\u8bed\u97f3\u548c\u8bed\u8a00\u7684\u7ec4\u6210\u65b9\u6cd5\u3002\u4e0e\u5927\u591a\u6570\u5148\u524d\u7684\u5de5\u4f5c\u4e0d\u540c\uff0c\u6211\u4eec\u9009\u62e9\u4e0d\u9488\u5bf9\u8bed\u97f3\u4efb\u52a1\u5bf9\u8bed\u8a00\u6a21\u578b\u672c\u8eab\u8fdb\u884c\u5fae\u8c03\uff0c\u56e0\u4e3a\u8fd9\u6837\u505a\u53ef\u80fd\u4f1a\u5bfc\u81f4\u975e\u8bed\u97f3\u4efb\u52a1\u7684\u7ade\u4e89\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u5373\u4f7f\u6ca1\u6709\u8fd9\u6837\u7684\u5fae\u8c03\uff0c\u5728\u66f4\u5927\u7684\u6a21\u578b\u89c4\u6a21\u4e0b\u4e5f\u80fd\u53d6\u5f97\u826f\u597d\u7684\u6027\u80fd\uff1b\u8bf7\u53c2\u89c1\u7b2c 8.4 \u8282\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>10 \u7ed3\u8bba<\/h2>\n<p>&nbsp;<\/p>\n<p>\u9ad8\u54c1\u8d28\u57fa\u7840\u6a21\u578b\u7684\u5f00\u53d1\u4ecd\u5904\u4e8e\u521d\u671f\u9636\u6bb5\u3002\u6211\u4eec\u5f00\u53d1Llama 3\u7684\u7ecf\u9a8c\u8868\u660e\uff0c\u8fd9\u4e9b\u6a21\u578b\u672a\u6765\u8fd8\u6709\u5f88\u5927\u7684\u6539\u8fdb\u7a7a\u95f4\u3002\u5728\u5f00\u53d1Llama 3\u6a21\u578b\u5bb6\u65cf\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0\u5f3a\u70c8\u7684\u5173\u6ce8\u9ad8\u8d28\u91cf\u6570\u636e\u3001\u89c4\u6a21\u548c\u7b80\u5355\u6027\u59cb\u7ec8\u80fd\u5e26\u6765\u6700\u4f73\u7ed3\u679c\u3002\u5728\u521d\u6b65\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u63a2\u7d22\u4e86\u66f4\u590d\u6742\u6a21\u578b\u67b6\u6784\u548c\u8bad\u7ec3\u65b9\u6848\uff0c\u4f46\u5e76\u6ca1\u6709\u53d1\u73b0\u8fd9\u4e9b\u65b9\u6cd5\u7684\u4f18\u52bf\u80fd\u591f\u8d85\u8fc7\u5b83\u4eec\u5728\u6a21\u578b\u5f00\u53d1\u4e2d\u5f15\u5165\u7684\u989d\u5916\u590d\u6742\u6027\u3002<\/p>\n<p>\u5f00\u53d1\u50cfLlama 3\u8fd9\u6837\u7684\u65d7\u8230\u57fa\u7840\u6a21\u578b\u4e0d\u4ec5\u9700\u8981\u514b\u670d\u8bb8\u591a\u6df1\u5c42\u6b21\u7684\u6280\u672f\u95ee\u9898\uff0c\u8fd8\u9700\u8981\u505a\u51fa\u660e\u667a\u7684\u7ec4\u7ec7\u51b3\u7b56\u3002\u4f8b\u5982\uff0c\u4e3a\u4e86\u786e\u4fddLlama 3\u4e0d\u4f1a\u610f\u5916\u8fc7\u5ea6\u62df\u5408\u4e8e\u5e38\u7528\u57fa\u51c6\u6d4b\u8bd5\uff0c\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u6570\u636e\u7531\u4e00\u4e2a\u72ec\u7acb\u7684\u56e2\u961f\u91c7\u8d2d\u548c\u5904\u7406\uff0c\u8be5\u56e2\u961f\u88ab\u5f3a\u70c8\u6fc0\u52b1\u9632\u6b62\u5c06\u5916\u90e8\u57fa\u51c6\u6d4b\u8bd5\u4e0e\u9884\u8bad\u7ec3\u6570\u636e\u6c61\u67d3\u3002\u53e6\u4e00\u4e2a\u4f8b\u5b50\u662f\uff0c\u6211\u4eec\u786e\u4fdd\u4eba\u7c7b\u8bc4\u4f30\u7684\u53ef\u4fe1\u5ea6\uff0c\u53ea\u5141\u8bb8\u4e00\u5c0f\u90e8\u5206\u4e0d\u53c2\u4e0e\u6a21\u578b\u5f00\u53d1\u7684\u7814\u7a76\u4eba\u5458\u6267\u884c\u548c\u8bbf\u95ee\u8fd9\u4e9b\u8bc4\u4f30\u3002\u867d\u7136\u8fd9\u4e9b\u7ec4\u7ec7\u51b3\u7b56\u5728\u6280\u672f\u8bba\u6587\u4e2d\u5f88\u5c11\u88ab\u8ba8\u8bba\uff0c\u4f46\u6211\u4eec\u53d1\u73b0\u5b83\u4eec\u5bf9\u4e8eLlama 3\u6a21\u578b\u5bb6\u65cf\u7684\u6210\u529f\u5f00\u53d1\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>\u6211\u4eec\u5206\u4eab\u4e86\u6211\u4eec\u7684\u5f00\u53d1\u8fc7\u7a0b\u7ec6\u8282\uff0c\u56e0\u4e3a\u6211\u4eec\u76f8\u4fe1\u8fd9\u5c06\u6709\u52a9\u4e8e\u66f4\u5e7f\u6cdb\u7684\u7814\u7a76\u793e\u533a\u4e86\u89e3\u57fa\u7840\u6a21\u578b\u5f00\u53d1\u7684\u5173\u952e\u56e0\u7d20\uff0c\u5e76\u4e3a\u516c\u4f17\u5173\u4e8e\u57fa\u7840\u6a21\u578b\u672a\u6765\u53d1\u5c55\u8fdb\u884c\u66f4\u6709\u89c1\u5730\u7684\u8ba8\u8bba\u505a\u51fa\u8d21\u732e\u3002\u6211\u4eec\u8fd8\u5206\u4eab\u4e86\u5c06\u591a\u6a21\u6001\u529f\u80fd\u6574\u5408\u5230Llama 3\u4e2d\u7684\u521d\u6b65\u5b9e\u9a8c\u7ed3\u679c\u3002\u867d\u7136\u8fd9\u4e9b\u6a21\u578b\u4ecd\u5728\u79ef\u6781\u5f00\u53d1\u4e2d\uff0c\u5c1a\u672a\u51c6\u5907\u597d\u53d1\u5e03\uff0c\u4f46\u6211\u4eec\u5e0c\u671b\u5c3d\u65e9\u5206\u4eab\u6211\u4eec\u7684\u7ed3\u679c\u53ef\u4ee5\u52a0\u901f\u8be5\u65b9\u5411\u7684\u7814\u7a76\u3002<\/p>\n<p>\u9274\u4e8e\u672c\u6587\u4e2d\u8be6\u7ec6\u7684\u5b89\u5168\u5206\u6790\u53d6\u5f97\u7684\u79ef\u6781\u6210\u679c\uff0c\u6211\u4eec\u5c06\u516c\u5f00\u53d1\u5e03\u6211\u4eec\u7684Llama 3\u8bed\u8a00\u6a21\u578b\uff0c\u4ee5\u52a0\u901f\u4e3a\u4f17\u591a\u793e\u4f1a\u76f8\u5173\u7528\u4f8b\u5f00\u53d1\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u8fdb\u7a0b\uff0c\u5e76\u4f7f\u7814\u7a76\u754c\u80fd\u591f\u5ba1\u67e5\u6211\u4eec\u7684\u6a21\u578b\u5e76\u627e\u5230\u4f7f\u8fd9\u4e9b\u6a21\u578b\u66f4\u597d\u3001\u66f4\u5b89\u5168\u7684\u9014\u5f84\u3002\u6211\u4eec\u76f8\u4fe1\u57fa\u7840\u6a21\u578b\u7684\u516c\u5f00\u53d1\u5e03\u5bf9\u4e8e\u6b64\u7c7b\u6a21\u578b\u7684\u8d1f\u8d23\u4efb\u53d1\u5c55\u81f3\u5173\u91cd\u8981\uff0c\u5e76\u4e14\u6211\u4eec\u5e0c\u671bLlama 3\u7684\u53d1\u5e03\u53ef\u4ee5\u9f13\u52b1\u6574\u4e2a\u884c\u4e1a\u62e5\u62b1\u5f00\u653e\u3001\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u5f00\u53d1\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6458\u8981: \u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u7cfb\u5217\u65b0\u7684\u57fa\u7840\u6a21\u578b\uff0c\u79f0\u4e3a Llama 3\u3002Llama 3 \u662f\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b\u7fa4\u4f53\uff0c\u5929\u751f\u652f\u6301\u591a\u8bed\u8a00\u3001\u4ee3\u7801\u7f16\u5199\u3001\u63a8\u7406\u548c\u5de5\u5177\u4f7f\u7528\u3002\u6211\u4eec\u6700\u5927\u7684\u6a21\u578b\u662f\u4e00\u4e2a\u5177\u6709 4050 \u4ebf\u4e2a\u53c2\u6570\u548c\u9ad8\u8fbe 128,000 \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u5bc6\u96c6\u578b Tra&#8230;<\/p>\n","protected":false},"author":1,"featured_media":3496,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-3429","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/3429","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=3429"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/3429\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/3496"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=3429"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=3429"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=3429"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}