{"id":28399,"date":"2025-03-13T00:41:18","date_gmt":"2025-03-12T16:41:18","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=28399"},"modified":"2025-03-13T00:41:18","modified_gmt":"2025-03-12T16:41:18","slug":"gemma-3-jishubaogao","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/gemma-3-jishubaogao\/","title":{"rendered":"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248"},"content":{"rendered":"<h2>Gemma 3 \u5173\u952e\u4fe1\u606f\u603b\u7ed3<\/h2>\n<h3>\u4e00\u3001\u5173\u952e\u6307\u6807<\/h3>\n<table>\n<thead>\n<tr>\n<th><strong>\u53c2\u6570<\/strong><\/th>\n<th><strong>\u8be6\u60c5<\/strong><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>\u6a21\u578b\u89c4\u6a21<\/strong><\/td>\n<td>1 \u4ebf\u5230 270 \u4ebf\u53c2\u6570\uff0c\u5171\u56db\u4e2a\u7248\u672c\uff1a1B\u30014B\u300112B\u300127B<\/td>\n<\/tr>\n<tr>\n<td><strong>\u67b6\u6784<\/strong><\/td>\n<td>\u57fa\u4e8e <a href=\"https:\/\/www.kdjingpai.com\/transformer\/\">Transformer<\/a> \u7684\u89e3\u7801\u5668\u4e13\u7528\u67b6\u6784\uff0c\u7ee7\u627f\u81ea Gemma 2\uff0c\u5e76\u8fdb\u884c\u591a\u9879\u6539\u8fdb<\/td>\n<\/tr>\n<tr>\n<td><strong>\u591a\u6a21\u6001\u80fd\u529b<\/strong><\/td>\n<td>\u652f\u6301\u6587\u672c\u548c\u56fe\u50cf\u8f93\u5165\uff0c\u4f7f\u7528\u5b9a\u5236\u7684 SigLIP \u89c6\u89c9\u7f16\u7801\u5668\uff0c\u5c06\u56fe\u50cf\u7f16\u7801\u4e3a 256 \u4e2a\u8f6f\u6807\u8bb0<\/td>\n<\/tr>\n<tr>\n<td><strong>\u4e0a\u4e0b\u6587\u957f\u5ea6<\/strong><\/td>\n<td>1B \u6a21\u578b\u652f\u6301 32K \u6807\u8bb0\uff0c\u5176\u4f59\u6a21\u578b\u652f\u6301 128K \u6807\u8bb0<\/td>\n<\/tr>\n<tr>\n<td><strong>\u6ce8\u610f\u529b\u673a\u5236<\/strong><\/td>\n<td>\u91c7\u7528 5:1 \u7684\u5c40\u90e8\/\u5168\u5c40\u6ce8\u610f\u529b\u5c42\u4ea4\u9519\u6a21\u5f0f\uff0c\u5c40\u90e8\u5c42\u8de8\u5ea6\u4e3a 1024 \u4e2a\u6807\u8bb0<\/td>\n<\/tr>\n<tr>\n<td><strong>\u8bad\u7ec3\u65b9\u6cd5<\/strong><\/td>\n<td>\u4f7f\u7528\u77e5\u8bc6\u84b8\u998f\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u5e76\u91c7\u7528\u6539\u8fdb\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03<\/td>\n<\/tr>\n<tr>\n<td><strong>\u9884\u8bad\u7ec3\u6570\u636e<\/strong><\/td>\n<td>\u9884\u8bad\u7ec3\u6570\u636e\u91cf\uff1a1B \u6a21\u578b 2T \u6807\u8bb0\uff0c4B \u6a21\u578b 4T \u6807\u8bb0\uff0c12B \u6a21\u578b 12T \u6807\u8bb0\uff0c27B \u6a21\u578b 14T \u6807\u8bb0<\/td>\n<\/tr>\n<tr>\n<td><strong>\u591a\u8bed\u8a00\u80fd\u529b<\/strong><\/td>\n<td>\u4f7f\u7528\u4e0e <a href=\"https:\/\/www.kdjingpai.com\/geminibardgubu\/\">Gemini<\/a> 2.0 \u76f8\u540c\u7684 SentencePiece \u5206\u8bcd\u5668\uff0c\u8bcd\u6c47\u8868\u5305\u542b 262k \u4e2a\u6761\u76ee\uff0c\u652f\u6301\u591a\u79cd\u8bed\u8a00<\/td>\n<\/tr>\n<tr>\n<td><strong>\u91cf\u5316\u652f\u6301<\/strong><\/td>\n<td>\u63d0\u4f9b\u4e0d\u540c\u6807\u51c6\u683c\u5f0f\u7684\u91cf\u5316\u7248\u672c\uff0c\u5305\u62ec\u6bcf\u901a\u9053 int4\u3001\u6bcf\u5757 int4 \u548c\u5207\u6362 fp8<\/td>\n<\/tr>\n<tr>\n<td><strong>\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd<\/strong><\/td>\n<td>\u4f7f\u7528 TPUv4\u3001TPUv5e \u548c TPUv5p \u8fdb\u884c\u8bad\u7ec3\uff0c\u91c7\u7528 ZeRO-3 \u4f18\u5316\u5668\u72b6\u6001\u5206\u7247\u548c\u6570\u636e\u590d\u5236\u7b56\u7565<\/td>\n<\/tr>\n<tr>\n<td><strong>\u78b3\u8db3\u8ff9<\/strong><\/td>\n<td>\u9884\u8bad\u7ec3\u78b3\u6392\u653e\u91cf\u4e3a 1497.13 \u5428 CO2e\uff0c\u8c37\u6b4c\u6570\u636e\u4e2d\u5fc3\u5df2\u5b9e\u73b0\u78b3\u4e2d\u548c<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u4e8c\u3001\u4f18\u70b9<\/h3>\n<ol>\n<li><strong>\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\u51fa\u8272<\/strong>\n<ul>\n<li>Gemma 3 \u96c6\u6210\u4e86\u89c6\u89c9\u7f16\u7801\u5668\uff0c\u80fd\u591f\u5904\u7406\u6587\u672c\u548c\u56fe\u50cf\u8f93\u5165\uff0c\u5728\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\u3002\n<ul>\n<li>\u5728 DocVQA\u3001InfoVQA\u3001TextVQA \u7b49\u89c6\u89c9\u95ee\u7b54\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u5c24\u5176\u662f\u5728\u6d89\u53ca\u6587\u6863\u7406\u89e3\u548c\u56fe\u50cf\u4e2d\u6587\u672c\u9605\u8bfb\u7684\u4efb\u52a1\u4e0a\u3002<\/li>\n<li>\u4e0e PaliGemma 2 \u76f8\u6bd4\uff0cGemma 3 \u5728\u6587\u6863\u7406\u89e3\u4efb\u52a1\u4e0a\u8868\u73b0\u66f4\u4f73\uff0c\u4f8b\u5982 DocVQA \u548c InfoVQA\u3002<\/li>\n<li>Pan &amp; Scan (P&amp;S) \u6280\u672f\u7684\u5e94\u7528\uff0c\u4f7f\u5f97 Gemma 3 \u80fd\u591f\u4ee5\u63a5\u8fd1\u539f\u751f\u5206\u8fa8\u7387\u5904\u7406\u56fe\u50cf\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u5176\u5728\u89c6\u89c9\u8bed\u8a00\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u80fd\u529b\u5f3a\u5927<\/strong>\n<ul>\n<li>\u652f\u6301 128K \u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff081B \u6a21\u578b\u4e3a 32K\uff09\uff0c\u5728\u5904\u7406\u957f\u6587\u672c\u4efb\u52a1\u65f6\u5177\u6709\u4f18\u52bf\u3002\n<ul>\n<li>\u5728 RULER \u548c MRCR \u7b49\u957f\u4e0a\u4e0b\u6587\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u826f\u597d\u3002<\/li>\n<li>\u5c40\u90e8\/\u5168\u5c40\u6ce8\u610f\u529b\u5c42\u4ea4\u9519\u6a21\u5f0f\u548c\u5c40\u90e8\u5c42\u77ed\u8de8\u5ea6\u7684\u8bbe\u8ba1\uff0c\u5728\u4fdd\u6301\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u80fd\u529b\u7684\u540c\u65f6\uff0c\u6709\u6548\u63a7\u5236\u4e86 KV \u7f13\u5b58\u7684\u5185\u5b58\u6d88\u8017\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u591a\u8bed\u8a00\u80fd\u529b\u63d0\u5347<\/strong>\n<ul>\n<li>\u4f7f\u7528\u66f4\u5e73\u8861\u7684\u975e\u82f1\u8bed\u8bed\u8a00\u5206\u8bcd\u5668\uff0c\u5e76\u589e\u52a0\u4e86\u591a\u8bed\u8a00\u6570\u636e\u7684\u8bad\u7ec3\u91cf\u3002\n<ul>\n<li>\u5728 MGSM\u3001Global-MMLU-Lite\u3001WMT24++ \u7b49\u591a\u8bed\u8a00\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u826f\u597d\u3002<\/li>\n<li>\u5728 XQuAD \u548c IndicGenBench \u7b49\u8de8\u8bed\u8a00\u95ee\u7b54\u548c\u5370\u5ea6\u8bed\u8a00\u751f\u6210\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6570\u5b66\u3001\u63a8\u7406\u548c\u4ee3\u7801\u80fd\u529b\u589e\u5f3a<\/strong>\n<ul>\n<li>\u540e\u8bad\u7ec3\u65b9\u6cd5\u9488\u5bf9\u6570\u5b66\u3001\u63a8\u7406\u548c\u4ee3\u7801\u80fd\u529b\u8fdb\u884c\u4e86\u4f18\u5316\u3002\n<ul>\n<li>\u5728 MATH\u3001GSM8K\u3001HumanEval \u7b49\u6570\u5b66\u548c\u4ee3\u7801\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u4f18\u5f02\u3002<\/li>\n<li>\u5728 MMLU\u3001MMLU-Pro\u3001AGIEval \u7b49\u591a\u4efb\u52a1\u8bed\u8a00\u7406\u89e3\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u826f\u597d\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6a21\u578b\u6548\u7387\u9ad8<\/strong>\n<ul>\n<li>1B \u548c 4B \u6a21\u578b\u53c2\u6570\u89c4\u6a21\u8f83\u5c0f\uff0c\u9002\u5408\u5728\u8d44\u6e90\u53d7\u9650\u7684\u8bbe\u5907\u4e0a\u90e8\u7f72\u3002<\/li>\n<li>\u901a\u8fc7\u5e73\u5747\u6c60\u5316\u6280\u672f\uff0cGemma 3 4B \u548c 12B \u6a21\u578b\u5728\u89c6\u89c9\u7f16\u7801\u5668\u8f93\u5165\u5206\u8fa8\u7387\u76f8\u540c\u7684\u60c5\u51b5\u4e0b\uff0c\u8fc1\u79fb\u6210\u672c\u6bd4 PaliGemma 2 9B \u548c 27B \u6a21\u578b\u4f4e\u7ea6 10 \u500d\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h3>\u4e09\u3001\u7f3a\u70b9<\/h3>\n<ol>\n<li><strong>\u5bf9\u5316\u5b66\u3001\u751f\u7269\u3001\u653e\u5c04\u6027\u548c\u6838\uff08CBRN\uff09\u77e5\u8bc6\u7684\u638c\u63e1\u6709\u9650<\/strong>\n<ul>\n<li>\u7531\u4e8e\u9884\u8bad\u7ec3\u6570\u636e\u4e2d CBRN \u76f8\u5173\u5185\u5bb9\u7684\u7f3a\u4e4f\uff0cGemma 3 \u5728 CBRN \u77e5\u8bc6\u8bc4\u4f30\u4e2d\u8868\u73b0\u4e0d\u4f73\u3002<\/li>\n<li>\u8fd9\u610f\u5473\u7740\u5728\u6d89\u53ca CBRN \u9886\u57df\u7684\u5e94\u7528\u573a\u666f\u4e2d\uff0cGemma 3 \u53ef\u80fd\u65e0\u6cd5\u63d0\u4f9b\u51c6\u786e\u53ef\u9760\u7684\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u53ef\u80fd\u5b58\u5728\u6027\u80fd\u74f6\u9888<\/strong>\n<ul>\n<li>\u867d\u7136 P&amp;S \u6280\u672f\u53ef\u4ee5\u7f13\u89e3\u975e\u6b63\u65b9\u5f62\u957f\u5bbd\u6bd4\u548c\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5e26\u6765\u7684\u95ee\u9898\uff0c\u4f46\u9891\u7e41\u7684\u56fe\u50cf\u88c1\u526a\u548c\u7f29\u653e\u53ef\u80fd\u4f1a\u5f71\u54cd\u63a8\u7406\u901f\u5ea6\u3002<\/li>\n<li>\u5728\u9700\u8981\u5904\u7406\u8d85\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u6216\u5b9e\u65f6\u89c6\u89c9\u4efb\u52a1\u7684\u5e94\u7528\u4e2d\uff0cGemma 3 \u53ef\u80fd\u9762\u4e34\u6027\u80fd\u6311\u6218\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5bf9\u67d0\u4e9b\u7279\u5b9a\u9886\u57df\u7684\u7406\u89e3\u53ef\u80fd\u4e0d\u8db3<\/strong>\n<ul>\n<li>\u5c3d\u7ba1 Gemma 3 \u5728\u591a\u4e2a\u9886\u57df\u8868\u73b0\u51fa\u8272\uff0c\u4f46\u5bf9\u4e8e\u67d0\u4e9b\u9ad8\u5ea6\u4e13\u4e1a\u5316\u7684\u9886\u57df\uff08\u4f8b\u5982\u6cd5\u5f8b\u3001\u533b\u5b66\u7b49\uff09\uff0c\u5176\u7406\u89e3\u80fd\u529b\u53ef\u80fd\u6709\u9650\u3002<\/li>\n<li>\u8fd9\u9700\u8981\u6839\u636e\u5177\u4f53\u5e94\u7528\u573a\u666f\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u5fae\u8c03\u6216\u9886\u57df\u9002\u914d\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5b58\u5728\u4e00\u5b9a\u7684\u8bb0\u5fc6\u548c\u9690\u79c1\u98ce\u9669<\/strong>\n<ul>\n<li>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u90fd\u5b58\u5728\u751f\u6210\u8bad\u7ec3\u6570\u636e\u4e2d\u51fa\u73b0\u7684\u6587\u672c\u7684\u98ce\u9669\uff0cGemma 3 \u4e5f\u4e0d\u4f8b\u5916\u3002<\/li>\n<li>\u867d\u7136 Gemma 3 \u7684\u8bb0\u5fc6\u7387\u4f4e\u4e8e\u4e4b\u524d\u7684\u6a21\u578b\uff0c\u4f46\u4ecd\u9700\u8c28\u614e\u5904\u7406\u654f\u611f\u4fe1\u606f\uff0c\u907f\u514d\u6cc4\u9732\u7528\u6237\u9690\u79c1\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5bf9\u590d\u6742\u56e0\u679c\u5173\u7cfb\u7684\u63a8\u7406\u80fd\u529b\u6709\u5f85\u63d0\u9ad8<\/strong>\n<ul>\n<li>\u5728\u6d89\u53ca\u590d\u6742\u56e0\u679c\u5173\u7cfb\u63a8\u7406\u7684\u4efb\u52a1\u4e2d\uff0cGemma 3 \u7684\u8868\u73b0\u53ef\u80fd\u4e0d\u5982\u9884\u671f\u3002<\/li>\n<li>\u8fd9\u9700\u8981\u8fdb\u4e00\u6b65\u7684\u7814\u7a76\u548c\u6a21\u578b\u6539\u8fdb\uff0c\u4ee5\u63d0\u9ad8\u5176\u5728\u56e0\u679c\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h3>Gemma 3 \u8bc4\u5206<\/h3>\n<p>Gemma 3\u7cfb\u5217\u6a21\u578b\u5305\u542b4\u4e2a\u7248\u672c\uff0c\u6bcf\u4e2a\u7248\u672c\u5747\u5f00\u6e90\u4e86\u9884\u8bad\u7ec3\u57fa\u5ea7\u7248\u672c\uff08pt\u540e\u7f00\u7248\u672c\uff0c\u8868\u793apre-training\uff09\u548c\u6307\u4ee4\u5fae\u8c03\u540e\u7684\u7248\u672c\uff08it\u540e\u7f00\u7248\u672c\uff0c\u8868\u793ainstruction fine-tuned\uff09\uff0c\u4e5f\u5c31\u662f\u8bf4\u5171\u5f00\u6e90\u4e868\u4e2a\u7248\u672c\u7684\u5927\u6a21\u578b\u3002<\/p>\n<p>\u800c\u6700\u5927\u53c2\u6570\u89c4\u6a21\u7684Gemma 3-27B IT\u7684fp16\u7cbe\u5ea6\u5927\u5c0f\u4e3a54.8GB\uff0cint8\u91cf\u5316\u540e27GB\uff0c\u4e24\u5f204090\u53ef\u7528\uff0cINT4\u91cf\u5316\u540e\u9700\u898114GB\u663e\u5b58\uff0c\u5355\u5f204090\u5b8c\u5168\u6ca1\u95ee\u9898\u3002<\/p>\n<p>\u800c\u8fd9\u4e2a\u7248\u672c\u7684\u6a21\u578b\u8bc4\u6d4b\u7ed3\u679c\u975e\u5e38\u597d\uff0c\u5728\u5927\u6a21\u578b\u533f\u540d\u7ade\u6280\u573a\uff08Chatbot Arena\uff09\u4e0a\u5f97\u52061338\u5206\uff08\u622a\u6b622025\u5e743\u67088\u65e5\uff09\uff0c\u6392\u540d\u5168\u7403\u7b2c9\uff0c\u4ec5\u6b21\u4e8eo1-2024-12-17\u6a21\u578b\uff0c\u8d85\u8fc7\u4e86Qwen2.5-Max\u4ee5\u53caDeepSeek V3\u7b49\u3002<\/p>\n<p>\u6309\u7167Google\u5b98\u65b9\u7684\u8bf4\u6cd5\uff0c\u672c\u6b21Gemma 3\u7cfb\u5217\u63d0\u5347\u5f88\u660e\u663e\uff0c\u5176\u4e2dGemma 3-4B\u7248\u672c\u6a21\u578b\u6c34\u5e73\u63a5\u8fd1Gemma 2-27B\uff0c\u800cGemma 3-27B\u5219\u63a5\u8fd1Gemini 1.5-Pro\uff01<\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/e8c7fa32725b603.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-1\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>Gemma 3 \u4f53\u9a8c\u5730\u5740<\/h3>\n<p>https:\/\/aistudio.google.com\/prompts\/new_chat?model=gemma-3-27b-it<\/p>\n<p>https:\/\/ollama.com\/library\/gemma3<\/p>\n<p>&nbsp;<\/p>\n<h3>\u603b\u7ed3<\/h3>\n<p>Gemma 3 \u662f\u4e00\u6b3e\u529f\u80fd\u5f3a\u5927\u7684\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u5728\u591a\u6a21\u6001\u7406\u89e3\u3001\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u3001\u591a\u8bed\u8a00\u80fd\u529b\u3001\u6570\u5b66\u3001\u63a8\u7406\u548c\u4ee3\u7801\u7b49\u65b9\u9762\u8868\u73b0\u51fa\u8272\u3002\u7136\u800c\uff0c\u5b83\u5728 CBRN \u77e5\u8bc6\u3001\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5904\u7406\u3001\u7279\u5b9a\u9886\u57df\u7406\u89e3\u3001\u8bb0\u5fc6\u548c\u9690\u79c1\u98ce\u9669\u4ee5\u53ca\u590d\u6742\u56e0\u679c\u63a8\u7406\u7b49\u65b9\u9762\u4ecd\u6709\u6539\u8fdb\u7a7a\u95f4\u3002<\/p>\n<hr \/>\n<p>&nbsp;<\/p>\n<p><strong>\u6280\u672f\u62a5\u544a\u5168\u6587\uff1a<\/strong>https:\/\/storage.googleapis.com\/deepmind-media\/gemma\/Gemma3Report.pdf<\/p>\n<p>\u6211\u4eec\u4ecb\u7ecd\u4e86 Gemma 3\uff0c\u8fd9\u662f Gemma \u7cfb\u5217\u8f7b\u91cf\u7ea7\u5f00\u6e90\u6a21\u578b\u7684\u591a\u6a21\u6001\u6269\u5c55\uff0c\u53c2\u6570\u89c4\u6a21\u4ece 1 \u4ebf\u5230 270 \u4ebf\u4e0d\u7b49\u3002\u8fd9\u4e2a\u7248\u672c\u5f15\u5165\u4e86\u89c6\u89c9\u7406\u89e3\u80fd\u529b\uff0c\u652f\u6301\u66f4\u591a\u8bed\u8a00\uff0c\u5e76\u4e14\u4e0a\u4e0b\u6587\u957f\u5ea6\u66f4\u957f\u2014\u2014\u81f3\u5c11\u652f\u6301 128K \u4e2a\u6807\u8bb0\u3002\u6211\u4eec\u8fd8\u5bf9\u6a21\u578b\u67b6\u6784\u8fdb\u884c\u4e86\u6539\u8fdb\uff0c\u4ee5\u51cf\u5c11\u968f\u7740\u4e0a\u4e0b\u6587\u957f\u5ea6\u589e\u52a0\u800c\u6025\u5267\u589e\u957f\u7684 KV \u7f13\u5b58\u5185\u5b58\u3002\u8fd9\u662f\u901a\u8fc7\u589e\u52a0\u5c40\u90e8\u5230\u5168\u5c40\u6ce8\u610f\u529b\u5c42\u7684\u6bd4\u4f8b\uff0c\u5e76\u5c06\u5c40\u90e8\u6ce8\u610f\u529b\u7684\u8de8\u5ea6\u4fdd\u6301\u5728\u8f83\u77ed\u8303\u56f4\u5185\u6765\u5b9e\u73b0\u7684\u3002Gemma 3 \u6a21\u578b\u901a\u8fc7\u77e5\u8bc6\u84b8\u998f\u8fdb\u884c\u8bad\u7ec3\uff0c\u5728\u9884\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u7248\u672c\u4e0a\u5747\u4f18\u4e8e Gemma 2\u3002\u7279\u522b\u662f\uff0c\u6211\u4eec\u65b0\u9896\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u663e\u8457\u63d0\u9ad8\u4e86\u6570\u5b66\u3001\u804a\u5929\u3001\u6307\u4ee4\u9075\u5faa\u548c\u591a\u8bed\u8a00\u80fd\u529b\uff0c\u4f7f\u5f97 Gemma3-4B-IT \u4e0e Gemma2-27B-IT \u6027\u80fd\u76f8\u5f53\uff0c\u800c Gemma3-27B-IT \u5728\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u53ef\u4e0e Gemini-1.5-Pro \u76f8\u5ab2\u7f8e\u3002\u6211\u4eec\u5c06\u6240\u6709\u6a21\u578b\u53d1\u5e03\u7ed9\u793e\u533a\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>1. \u5f15\u8a00<\/h2>\n<p>\u6211\u4eec\u4ecb\u7ecd\u4e86 Gemma \u5f00\u6e90\u8bed\u8a00\u6a21\u578b\uff08Gemma \u56e2\u961f\uff0c2024a\uff09\u7684\u6700\u65b0\u7248\u672c\uff0c\u8be5\u7248\u672c\u4e0e Gemini \u524d\u6cbf\u6a21\u578b\u5bb6\u65cf\uff08Gemini \u56e2\u961f\uff0c2023\uff09\u5171\u540c\u8bbe\u8ba1\u3002\u8fd9\u4e2a\u65b0\u7248\u672c\u7684\u5927\u5c0f\u4e0e Gemma 2\uff08Gemma \u56e2\u961f\uff0c2024b\uff09\u76f8\u5f53\uff0c\u5e76\u589e\u52a0\u4e86\u4e00\u4e2a 10 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u3002\u8fd9\u4e9b\u6a21\u578b\u65e8\u5728\u5728\u624b\u673a\u3001\u7b14\u8bb0\u672c\u7535\u8111\u548c\u9ad8\u7aef GPU \u7b49\u6807\u51c6\u6d88\u8d39\u7ea7\u786c\u4ef6\u4e0a\u8fd0\u884c\u3002\u8fd9\u4e2a\u7248\u672c\u4e3a Gemma \u7cfb\u5217\u5e26\u6765\u4e86\u51e0\u9879\u65b0\u529f\u80fd\uff0c\u5373\u591a\u6a21\u6001\u3001\u957f\u4e0a\u4e0b\u6587\u548c\u591a\u8bed\u8a00\u6027\uff0c\u540c\u65f6\u4fdd\u6301\u6216\u8d85\u8d8a\u4e86\u4e4b\u524d\u7248\u672c\u7684\u6027\u80fd\u3002<\/p>\n<p>\u5728\u591a\u6a21\u6001\u65b9\u9762\uff0c\u5927\u591a\u6570 Gemma 3 \u6a21\u578b\u90fd\u4e0e\u5b9a\u5236\u7684 SigLIP \u89c6\u89c9\u7f16\u7801\u5668\uff08Zhai \u7b49\u4eba\uff0c2023\uff09\u517c\u5bb9\u3002\u8bed\u8a00\u6a21\u578b\u5c06\u56fe\u50cf\u89c6\u4e3a\u7531 SigLIP \u7f16\u7801\u7684\u8f6f\u6807\u8bb0\u5e8f\u5217\u3002\u6211\u4eec\u901a\u8fc7\u5c06\u89c6\u89c9\u5d4c\u5165\u538b\u7f29\u4e3a\u56fa\u5b9a\u5927\u5c0f\u7684 256 \u4e2a\u5411\u91cf\u6765\u964d\u4f4e\u56fe\u50cf\u5904\u7406\u7684\u63a8\u7406\u6210\u672c\u3002\u7f16\u7801\u5668\u4ee5\u56fa\u5b9a\u5206\u8fa8\u7387\u5de5\u4f5c\uff0c\u6211\u4eec\u4ece LLaVA\uff08Liu \u7b49\u4eba\uff0c2024\uff09\u4e2d\u6c72\u53d6\u7075\u611f\uff0c\u901a\u8fc7\u5e73\u79fb\u548c\u626b\u63cf\uff08P&amp;S\uff09\u65b9\u6cd5\u5b9e\u73b0\u7075\u6d3b\u7684\u5206\u8fa8\u7387\u3002<\/p>\n<p>\u7b2c\u4e8c\u4e2a\u4e3b\u8981\u7684\u67b6\u6784\u6539\u8fdb\u662f\u5c06\u4e0a\u4e0b\u6587\u5927\u5c0f\u589e\u52a0\u5230 128K \u4e2a\u6807\u8bb0\uff0c\u540c\u65f6\u4e0d\u5f71\u54cd\u6027\u80fd\u3002\u957f\u4e0a\u4e0b\u6587\u7684\u4e00\u4e2a\u6311\u6218\u662f\u63a8\u7406\u8fc7\u7a0b\u4e2d KV \u7f13\u5b58\u7684\u5185\u5b58\u7206\u70b8\u3002\u4e3a\u4e86\u51cf\u5c11\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5728\u6bcf\u5c42\u5168\u5c40\u5c42\u4e4b\u95f4\u4ea4\u9519\u591a\u4e2a\u5c40\u90e8\u5c42\uff0c\u5e76\u5c06\u5c40\u90e8\u5c42\u7684\u8de8\u5ea6\u8bbe\u7f6e\u4e3a\u4ec5 1024 \u4e2a\u6807\u8bb0\u3002\u56e0\u6b64\uff0c\u53ea\u6709\u5168\u5c40\u5c42\u5173\u6ce8\u957f\u4e0a\u4e0b\u6587\uff0c\u6211\u4eec\u6bcf 5 \u5c42\u5c40\u90e8\u5c42\u5c31\u6709 1 \u5c42\u5168\u5c40\u5c42\u3002<\/p>\n<p>\u9884\u8bad\u7ec3\u4f18\u5316\u65b9\u6cd5\u4e0e Gemma 2 \u7c7b\u4f3c\uff0c\u4f46\u5728\u67b6\u6784\u8bbe\u8ba1\u4e0a\u8fdb\u884c\u4e86\u4e00\u4e9b\u4fee\u6539\u3002\u6211\u4eec\u4f7f\u7528\u4e0e Gemini 2.0 \u76f8\u540c\u7684\u5206\u8bcd\u5668\uff0c\u5e76\u4e14\u6211\u4eec\u8fd8\u91cd\u65b0\u5ba1\u89c6\u4e86\u6211\u4eec\u7684\u6570\u636e\u6df7\u5408\u7b56\u7565\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u591a\u8bed\u8a00\u80fd\u529b\uff0c\u540c\u65f6\u5f15\u5165\u56fe\u50cf\u7406\u89e3\u3002\u6240\u6709 Gemma 3 \u6a21\u578b\u90fd\u901a\u8fc7\u77e5\u8bc6\u84b8\u998f\uff08Hinton \u7b49\u4eba\uff0c2015\uff09\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n<p>\u5728\u540e\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u63d0\u9ad8\u6570\u5b66\u3001\u63a8\u7406\u548c\u804a\u5929\u80fd\u529b\uff0c\u4ee5\u53ca\u6574\u5408 Gemma 3 \u7684\u65b0\u529f\u80fd\u3001\u957f\u4e0a\u4e0b\u6587\u548c\u56fe\u50cf\u8f93\u5165\u3002\u6211\u4eec\u4f7f\u7528\u4e00\u79cd\u65b0\u9896\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5728\u6570\u5b66\u3001\u7f16\u7801\u3001\u804a\u5929\u3001\u6307\u4ee4\u9075\u5faa\u548c\u591a\u8bed\u8a00\u7b49\u6240\u6709\u80fd\u529b\u4e0a\u90fd\u6709\u6240\u63d0\u5347\u3002\u7531\u6b64\u4ea7\u751f\u7684 Gemma 3 \u6307\u4ee4\u5fae\u8c03\u6a21\u578b\u65e2\u5f3a\u5927\u53c8\u591a\u529f\u80fd\uff0c\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5927\u5e45\u8d85\u8d8a\u5176\u524d\u8eab\u3002<\/p>\n<p>\u5728\u4ee5\u4e0b\u7ae0\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u7b80\u8981\u6982\u8ff0\u6211\u4eec\u7684\u6a21\u578b\uff0c\u5305\u62ec\u67b6\u6784\u4ee5\u53ca\u9884\u8bad\u7ec3\u548c\u540e\u8bad\u7ec3\u65b9\u6cd5\u3002\u6211\u4eec\u8fd8\u63d0\u4f9b\u4e86\u5728\u5404\u79cd\u5b9a\u91cf\u548c\u5b9a\u6027\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u8be6\u7ec6\u8bc4\u4f30\u3002\u6211\u4eec\u8ba8\u8bba\u4e86\u5b89\u5168\u8d1f\u8d23\u90e8\u7f72\u7684\u65b9\u6cd5\uff0c\u5e76\u6982\u8ff0\u4e86 Gemma 3 \u7684\u66f4\u5e7f\u6cdb\u5f71\u54cd\u3001\u5176\u5c40\u9650\u6027\u4ee5\u53ca\u4f18\u52bf\u3002<\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/bbb1fe238e110d5.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-1\" \/><\/p>\n<p>\u8fd9\u662f\u57fa\u4e8e\u6536\u636e\u7684\u6210\u672c\u548c\u5c0f\u8d39\u8ba1\u7b97\u660e\u7ec6\uff1a<\/p>\n<ol>\n<li><\/li>\n<\/ol>\n<p>1.<strong>\u8bc6\u522b\u5207\u7247\u8089\u7684\u6210\u672c\uff1a<\/strong>\u6536\u636e\u4e0a\u5217\u51fa\u4e86\u201c1x\u00a0Zurcher\u00a0Geschnetzeltes\u00a0+\u00a0Rosti\u201d\uff0c\u4ef7\u683c\u4e3a\u00a036.50\u00a0\u745e\u58eb\u6cd5\u90ce\u3002\u8fd9\u662f\u5e26\u6709\u70b8\u571f\u8c46\uff08Rosti\uff09\u914d\u83dc\u7684\u5207\u7247\u8089\u83dc\u30022.<strong>\u8ba1\u7b97 18% \u7684\u5c0f\u8d39\uff1a<\/strong>\u5c0f\u8d39\u91d1\u989d\u00a0=\u00a036.5\u00a0\u745e\u58eb\u6cd5\u90ce\u00a0*\u00a00.18\u00a0=\u00a06.57\u00a0\u745e\u58eb\u6cd5\u90ce3.<strong>\u8ba1\u7b97\u603b\u91d1\u989d\uff1a<\/strong>\u603b\u91d1\u989d\u00a0=\u00a036.50\u00a0\u745e\u58eb\u6cd5\u90ce\u00a0+\u00a06.57\u00a0\u745e\u58eb\u6cd5\u90ce\u00a0=\u00a043.07\u00a0\u745e\u58eb\u6cd5\u90ce<\/p>\n<p>&nbsp;<\/p>\n<h2>2. \u6a21\u578b\u67b6\u6784<\/h2>\n<p>Gemma 3 \u6a21\u578b\u9075\u5faa\u4e0e\u524d\u51e0\u4ee3\u6a21\u578b\uff08Vaswani \u7b49\u4eba\uff0c2017\uff09\u76f8\u540c\u7684\u901a\u7528\u89e3\u7801\u5668\u4e13\u7528 Transformer \u67b6\u6784\uff0c\u5927\u591a\u6570\u67b6\u6784\u5143\u7d20\u4e0e\u524d\u4e24\u4e2a Gemma \u7248\u672c\u76f8\u4f3c\u3002\u6211\u4eec\u4f7f\u7528\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff08GQA\uff09\uff08Ainslie \u7b49\u4eba\uff0c2023\uff09\u4ee5\u53ca\u540e\u5f52\u4e00\u5316\u548c\u9884\u5f52\u4e00\u5316\u4e0e RMSNorm\uff08Zhang \u548c Sennrich\uff0c2019\uff09\u3002\u53d7 Dehghani \u7b49\u4eba\uff082023\uff09\u3001Wortsman \u7b49\u4eba\uff082023\uff09\u548c Chameleon \u56e2\u961f\uff082024\uff09\u7684\u542f\u53d1\uff0c\u6211\u4eec\u7528 QK \u5f52\u4e00\u5316\u66ff\u6362\u4e86 Gemma 2 \u7684\u8f6f\u5e3d\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u91cd\u70b9\u4ecb\u7ecd\u4e0e\u524d\u51e0\u4ee3\u6a21\u578b\u7684\u4e00\u4e9b\u5173\u952e\u533a\u522b\u3002<\/p>\n<p><strong>5:1 \u5c40\u90e8\/\u5168\u5c40\u5c42\u4ea4\u9519\u3002<\/strong>\u00a0\u6211\u4eec\u5728\u5c40\u90e8\u6ed1\u52a8\u7a97\u53e3\u81ea\u6ce8\u610f\u529b\uff08Beltagy \u7b49\u4eba\uff0c2020\uff09\u548c\u5168\u5c40\u81ea\u6ce8\u610f\u529b\uff08Luong \u7b49\u4eba\uff0c2015\uff09\u4e4b\u95f4\u4ea4\u66ff\uff0c\u4ea4\u66ff\u6a21\u5f0f\u4e3a\u6bcf 5 \u5c42\u5c40\u90e8\u5c42\u5bf9\u5e94 1 \u5c42\u5168\u5c40\u5c42\uff0c\u4ece\u5c40\u90e8\u5c42\u4f5c\u4e3a\u6a21\u578b\u7684\u7b2c\u4e00\u5c42\u5f00\u59cb\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>\u89c6\u89c9\u7f16\u7801\u5668<\/th>\n<th>\u5d4c\u5165\u53c2\u6570<\/th>\n<th>\u975e\u5d4c\u5165\u53c2\u6570<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>1B<\/td>\n<td>0<\/td>\n<td>302M<\/td>\n<td>698M<\/td>\n<\/tr>\n<tr>\n<td>4B<\/td>\n<td>417M<\/td>\n<td>675M<\/td>\n<td>3,209M<\/td>\n<\/tr>\n<tr>\n<td>12B<\/td>\n<td>417M<\/td>\n<td>1,012M<\/td>\n<td>10,759M<\/td>\n<\/tr>\n<tr>\n<td>27B<\/td>\n<td>417M<\/td>\n<td>1,416M<\/td>\n<td>25,600M<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u957f\u4e0a\u4e0b\u6587\u3002<\/strong>\u00a0Gemma 3 \u6a21\u578b\u652f\u6301 128K \u4e2a\u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c1B \u6a21\u578b\u9664\u5916\uff0c\u5176\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e3a 32K\u3002\u6211\u4eec\u5c06\u5168\u5c40\u81ea\u6ce8\u610f\u529b\u5c42\u7684 RoPE \u57fa\u9891\u4ece 10k \u589e\u52a0\u5230 1M\uff0c\u5e76\u5c06\u5c40\u90e8\u5c42\u7684\u9891\u7387\u4fdd\u6301\u5728 10k\u3002\u6211\u4eec\u9075\u5faa\u4e0e Chen \u7b49\u4eba\uff082023\uff09\u7684\u4f4d\u7f6e\u63d2\u503c\u7c7b\u4f3c\u7684\u8fc7\u7a0b\uff0c\u4ee5\u6269\u5c55\u5168\u5c40\u81ea\u6ce8\u610f\u529b\u5c42\u7684\u8de8\u5ea6\u3002<\/p>\n<h3>2.1 \u89c6\u89c9\u6a21\u6001<\/h3>\n<p><strong>\u89c6\u89c9\u7f16\u7801\u5668\u3002<\/strong>\u00a0\u6211\u4eec\u4f7f\u7528 SigLIP \u7f16\u7801\u5668\uff08Zhai \u7b49\u4eba\uff0c2023\uff09\u7684 400M \u53d8\u4f53\uff0c\u8fd9\u662f\u4e00\u4e2a\u5728 CLIP \u635f\u5931\uff08Radford \u7b49\u4eba\uff0c2021\uff09\u7684\u53d8\u4f53\u4e0a\u8bad\u7ec3\u7684\u89c6\u89c9 Transformer\uff08Dosovitskiy\uff0c2020\uff09\u3002Gemma \u89c6\u89c9\u7f16\u7801\u5668\u4ee5\u8c03\u6574\u4e3a 896&#215;896 \u7684\u6b63\u65b9\u5f62\u56fe\u50cf\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u5728\u89c6\u89c9\u52a9\u624b\u4efb\u52a1\u6570\u636e\u4e0a\u8fdb\u884c\u5fae\u8c03\u3002\u4e3a\u4e86\u7b80\u5355\u8d77\u89c1\uff0c\u6211\u4eec\u5728 4B\u300112B \u548c 27B \u6a21\u578b\u4e4b\u95f4\u5171\u4eab\u89c6\u89c9\u7f16\u7801\u5668\uff0c\u5e76\u5728\u8bad\u7ec3\u671f\u95f4\u4fdd\u6301\u5176\u51bb\u7ed3\u72b6\u6001\u3002<\/p>\n<p><strong>\u5e73\u79fb\u548c\u626b\u63cf\uff08P&amp;S\uff09\u3002<\/strong>\u00a0Gemma \u89c6\u89c9\u7f16\u7801\u5668\u4ee5 896&#215;896 \u7684\u56fa\u5b9a\u5206\u8fa8\u7387\u8fd0\u884c\u3002\u5f53\u5904\u7406\u975e\u6b63\u65b9\u5f62\u957f\u5bbd\u6bd4\u548c\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\uff0c\u8fd9\u4f1a\u5bfc\u81f4\u56fe\u50cf\u5931\u771f\uff0c\u5bfc\u81f4\u6587\u672c\u4e0d\u53ef\u8bfb\u6216\u5c0f\u7269\u4f53\u6d88\u5931\u3002\u6211\u4eec\u901a\u8fc7\u5728\u63a8\u7406\u65f6\u4f7f\u7528\u81ea\u9002\u5e94\u7a97\u53e3\u7b97\u6cd5\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002\u8be5\u7b97\u6cd5\u5c06\u56fe\u50cf\u5206\u5272\u6210\u5927\u5c0f\u76f8\u7b49\u7684\u975e\u91cd\u53e0\u88c1\u526a\u5757\uff0c\u8986\u76d6\u6574\u4e2a\u56fe\u50cf\uff0c\u5e76\u5c06\u5176\u8c03\u6574\u4e3a 896&#215;896 \u50cf\u7d20\u4f20\u9012\u7ed9\u7f16\u7801\u5668\u3002\u8fd9\u79cd\u7a97\u53e3\u5316\u4ec5\u5728\u5fc5\u8981\u65f6\u5e94\u7528\uff0c\u5e76\u4e14\u63a7\u5236\u6700\u5927\u88c1\u526a\u6570\u91cf\u3002\u8fd9\u662f\u4e00\u79cd\u4ec5\u5728\u63a8\u7406\u65f6\u8fdb\u884c\u7684\u4f18\u5316\uff0c\u53ef\u4ee5\u7981\u7528\u4ee5\u52a0\u5feb\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>Shards<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6a21\u578b<\/td>\n<td>\u7c7b\u578b<\/td>\n<td>\u82af\u7247\u6570\u91cf<\/td>\n<td>\u6570\u636e\u5e8f\u5217<\/td>\n<td>\u526f\u672c<\/td>\n<\/tr>\n<tr>\n<td>1B<\/td>\n<td>TPUv5e<\/td>\n<td>512<\/td>\n<td>16 16<\/td>\n<td>2<\/td>\n<\/tr>\n<tr>\n<td>4B<\/td>\n<td>TPUv5e<\/td>\n<td>2048<\/td>\n<td>16 16<\/td>\n<td>8<\/td>\n<\/tr>\n<tr>\n<td>12B<\/td>\n<td>TPUv4<\/td>\n<td>6144<\/td>\n<td>16 16<\/td>\n<td>24<\/td>\n<\/tr>\n<tr>\n<td>27B<\/td>\n<td>TPUv5p<\/td>\n<td>6144<\/td>\n<td>24 8<\/td>\n<td>32<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>2.2 \u9884\u8bad\u7ec3<\/h3>\n<p>\u6211\u4eec\u9075\u5faa\u4e0e Gemma 2 \u7c7b\u4f3c\u7684\u9884\u8bad\u7ec3\u914d\u65b9\uff0c\u5e76\u8fdb\u884c\u77e5\u8bc6\u84b8\u998f\u3002<\/p>\n<p><strong>\u8bad\u7ec3\u6570\u636e\u3002<\/strong>\u00a0\u6211\u4eec\u5728\u6bd4 Gemma 2 \u7a0d\u5927\u7684\u6807\u8bb0\u9884\u7b97\u4e0a\u9884\u8bad\u7ec3\u6211\u4eec\u7684\u6a21\u578b\uff0c\u5373\u6211\u4eec\u4f7f\u7528 14T \u6807\u8bb0\u8bad\u7ec3 Gemma 3 27B\uff0c12T \u6807\u8bb0\u8bad\u7ec3 12B \u7248\u672c\uff0c4T \u6807\u8bb0\u8bad\u7ec3 4B\uff0c2T \u6807\u8bb0\u8bad\u7ec3 1B\u3002\u6807\u8bb0\u7684\u589e\u52a0\u662f\u4e3a\u4e86\u8003\u8651\u9884\u8bad\u7ec3\u671f\u95f4\u4f7f\u7528\u7684\u56fe\u50cf\u548c\u6587\u672c\u7684\u6df7\u5408\u3002\u6211\u4eec\u8fd8\u589e\u52a0\u4e86\u591a\u8bed\u8a00\u6570\u636e\u7684\u6570\u91cf\uff0c\u4ee5\u63d0\u9ad8\u8bed\u8a00\u8986\u76d6\u8303\u56f4\u3002\u6211\u4eec\u6dfb\u52a0\u4e86\u5355\u8bed\u548c\u5e76\u884c\u6570\u636e\uff0c\u5e76\u4f7f\u7528\u53d7 Chung \u7b49\u4eba\uff082023\uff09\u542f\u53d1\u7684\u7b56\u7565\u6765\u5904\u7406\u8bed\u8a00\u8868\u793a\u7684\u4e0d\u5e73\u8861\u3002<\/p>\n<p><strong>\u5206\u8bcd\u5668\u3002<\/strong>\u00a0\u6211\u4eec\u4f7f\u7528\u4e0e Gemini 2.0 \u76f8\u540c\u7684\u5206\u8bcd\u5668\uff1a\u4e00\u4e2a\u5e26\u6709\u5206\u5272\u6570\u5b57\u3001\u4fdd\u7559\u7a7a\u767d\u548c\u5b57\u8282\u7ea7\u7f16\u7801\uff08Kudo \u548c Richardson\uff0c2018\uff09\u7684 SentencePiece \u5206\u8bcd\u5668\u3002\u7531\u6b64\u4ea7\u751f\u7684\u8bcd\u6c47\u8868\u6709 262k \u4e2a\u6761\u76ee\u3002\u8fd9\u4e2a\u5206\u8bcd\u5668\u5bf9\u4e8e\u975e\u82f1\u8bed\u8bed\u8a00\u66f4\u52a0\u5e73\u8861\u3002<\/p>\n<p><strong>\u8fc7\u6ee4\u3002<\/strong>\u00a0\u6211\u4eec\u4f7f\u7528\u8fc7\u6ee4\u6280\u672f\u6765\u51cf\u5c11\u4ea7\u751f\u4e0d\u9700\u8981\u6216\u4e0d\u5b89\u5168\u7684\u8bdd\u8bed\u7684\u98ce\u9669\uff0c\u5e76\u5220\u9664\u67d0\u4e9b\u4e2a\u4eba\u4fe1\u606f\u548c\u5176\u4ed6\u654f\u611f\u6570\u636e\u3002\u6211\u4eec\u5bf9\u9884\u8bad\u7ec3\u6570\u636e\u6df7\u5408\u4e2d\u7684\u8bc4\u4f30\u96c6\u8fdb\u884c\u53bb\u6c61\u67d3\uff0c\u5e76\u901a\u8fc7\u6700\u5c0f\u5316\u654f\u611f\u8f93\u51fa\u7684\u6269\u6563\u6765\u964d\u4f4e\u80cc\u8bf5\u7684\u98ce\u9669\u3002\u6211\u4eec\u8fd8\u5e94\u7528\u4e86\u53d7 Sachdeva \u7b49\u4eba\uff082024\uff09\u542f\u53d1\u7684\u8d28\u91cf\u91cd\u65b0\u52a0\u6743\u6b65\u9aa4\uff0c\u4ee5\u51cf\u5c11\u4f4e\u8d28\u91cf\u6570\u636e\u7684\u51fa\u73b0\u3002<\/p>\n<p><strong>\u77e5\u8bc6\u84b8\u998f\u3002<\/strong>\u00a0\u6211\u4eec\u5bf9\u6bcf\u4e2a\u6807\u8bb0\u91c7\u6837 256 \u4e2a logits\uff0c\u5e76\u6839\u636e\u6559\u5e08\u6982\u7387\u8fdb\u884c\u52a0\u6743\u3002\u5b66\u751f\u901a\u8fc7\u4ea4\u53c9\u71b5\u635f\u5931\u5b66\u4e60\u6559\u5e08\u5728\u8fd9\u4e9b\u6837\u672c\u4e2d\u7684\u5206\u5e03\u3002\u6559\u5e08\u7684\u76ee\u6807\u5206\u5e03\u5bf9\u4e8e\u672a\u91c7\u6837\u7684 logits \u8bbe\u7f6e\u4e3a\u96f6\u6982\u7387\uff0c\u5e76\u91cd\u65b0\u5f52\u4e00\u5316\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th><\/th>\n<th>\u539f\u59cb\uff08GB\uff09<\/th>\n<th>\u91cf\u5316\uff08GB\uff09<\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6a21\u578b 1B<\/td>\n<td>bf16 2.0<\/td>\n<td>0.5<\/td>\n<td>0.7 1.0<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>+KV<\/td>\n<td>2.9<\/td>\n<td>1.4<\/td>\n<td>1.6 1.9<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>4B<\/td>\n<td>8.0<\/td>\n<td>2.6<\/td>\n<td>2.9 4.4<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>+KV<\/td>\n<td>12.7<\/td>\n<td>7.3<\/td>\n<td>7.6 9.1<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>12B<\/td>\n<td>24.0<\/td>\n<td>6.6<\/td>\n<td>7.1 12.4<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>+KV<\/td>\n<td>38.9<\/td>\n<td>21.5<\/td>\n<td>22.0 27.3<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>27B<\/td>\n<td>54.0<\/td>\n<td>14.1<\/td>\n<td>15.3 27.4<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>+KV<\/td>\n<td>72.7<\/td>\n<td>32.8<\/td>\n<td>34.0 46.1<\/td>\n<td><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>2.3 \u91cf\u5316\u611f\u77e5\u8bad\u7ec3<\/h3>\n<p>\u9664\u4e86\u539f\u59cb\u68c0\u67e5\u70b9\u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u63d0\u4f9b\u4e0d\u540c\u6807\u51c6\u683c\u5f0f\u7684\u6a21\u578b\u91cf\u5316\u7248\u672c\u3002\u8fd9\u4e9b\u7248\u672c\u662f\u901a\u8fc7\u5bf9\u6bcf\u4e2a\u6a21\u578b\u8fdb\u884c\u5c11\u91cf\u6b65\u9aa4\uff08\u901a\u5e38\u4e3a 5000 \u6b65\uff09\u7684\u5fae\u8c03\uff0c\u4f7f\u7528\u91cf\u5316\u611f\u77e5\u8bad\u7ec3\uff08QAT\uff09\uff08Jacob \u7b49\u4eba\uff0c2018\uff09\u83b7\u5f97\u7684\u3002\u6211\u4eec\u4f7f\u7528\u672a\u91cf\u5316\u68c0\u67e5\u70b9\u7684\u6982\u7387\u4f5c\u4e3a\u76ee\u6807\uff0c\u5e76\u8c03\u6574\u6570\u636e\u4ee5\u5339\u914d\u9884\u8bad\u7ec3\u548c\u540e\u8bad\u7ec3\u5206\u5e03\u3002\u57fa\u4e8e\u6700\u6d41\u884c\u7684\u5f00\u6e90\u91cf\u5316\u63a8\u7406\u5f15\u64ce\uff08\u4f8b\u5982 llama.cpp\uff09\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u4e09\u79cd\u6743\u91cd\u8868\u793a\uff1a\u6bcf\u901a\u9053 int4\u3001\u6bcf\u5757 int4 \u548c\u5207\u6362 fp8\u3002\u5728\u8868 3 \u4e2d\uff0c\u6211\u4eec\u62a5\u544a\u4e86\u539f\u59cb\u548c\u91cf\u5316\u6a21\u578b\u5728\u4e0d\u540c\u6743\u91cd\u8868\u793a\u4e0b\u4ee5\u53ca\u5728 32k \u6807\u8bb0\u5e8f\u5217\u957f\u5ea6\u4e0b\u6709\u548c\u6ca1\u6709 KV \u7f13\u5b58\u7684\u5185\u5b58\u5360\u7528\u3002<\/p>\n<h3>2.4 \u8ba1\u7b97\u57fa\u7840\u8bbe\u65bd<\/h3>\n<p>\u6211\u4eec\u4f7f\u7528 TPUv4\u3001TPUv5e \u548c TPUv5p \u8bad\u7ec3\u6211\u4eec\u7684\u6a21\u578b\uff0c\u5982\u8868 2 \u6240\u793a\u3002\u6bcf\u4e2a\u6a21\u578b\u914d\u7f6e\u90fd\u7ecf\u8fc7\u4f18\u5316\uff0c\u4ee5\u6700\u5c0f\u5316\u8bad\u7ec3\u6b65\u9aa4\u65f6\u95f4\u3002\u5bf9\u4e8e\u89c6\u89c9\u7f16\u7801\u5668\uff0c\u6211\u4eec\u9884\u5148\u8ba1\u7b97\u6bcf\u4e2a\u56fe\u50cf\u7684\u5d4c\u5165\uff0c\u5e76\u76f4\u63a5\u4f7f\u7528\u5d4c\u5165\u8fdb\u884c\u8bad\u7ec3\uff0c\u4e0d\u4f1a\u589e\u52a0\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u7684\u6210\u672c\u3002<\/p>\n<p>\u4f18\u5316\u5668\u72b6\u6001\u4f7f\u7528 ZeRO-3\uff08Ren \u7b49\u4eba\uff0c2021\uff09\u7684\u5b9e\u73b0\u8fdb\u884c\u5206\u7247\u3002\u5bf9\u4e8e\u591a pod \u8bad\u7ec3\uff0c\u6211\u4eec\u6267\u884c\u6570\u636e\u4e2d\u5fc3\u7f51\u7edc\u4e0a\u7684\u6570\u636e\u526f\u672c\u91cd<\/p>\n<table>\n<thead>\n<tr>\n<th>\u4e0a\u4e0b\u6587<\/th>\n<th>\u683c\u5f0f\u5316<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u7528\u6237\u8f6e\u6b21<\/td>\n<td>user<\/td>\n<\/tr>\n<tr>\n<td>\u6a21\u578b\u8f6e\u6b21<\/td>\n<td>model<\/td>\n<\/tr>\n<tr>\n<td>\u8f6e\u6b21\u7ed3\u675f<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>\u8ba8\u8bba\u793a\u4f8b\uff1a<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>\u7528\u6237\uff1a\u4f60\u662f\u8c01\uff1f \u6a21\u578b\uff1a\u6211\u7684\u540d\u5b57\u662f Gemma\uff01 \u7528\u6237\uff1a2+2 \u662f\u591a\u5c11\uff1f \u6a21\u578b\uff1a2+2=4\u3002<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>\u6a21\u578b\u8f93\u5165\uff1a<\/td>\n<\/tr>\n<tr>\n<td>[BOS]user Whoare you? model My name is Gemma! user<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>Whatis2+2? model<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>2+2=4.<\/td>\n<td>\u6a21\u578b\u8f93\u51fa\uff1a<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>2.5 \u78b3\u8db3\u8ff9<\/h3>\n<p>\u9884\u8bad\u7ec3 Gemma 3 \u6a21\u578b\u7684\u78b3\u6392\u653e\u91cf\u4e3a 1497.13 \u5428 CO2e\u3002\u8fd9\u662f\u6839\u636e\u6211\u4eec TPU \u6570\u636e\u4e2d\u5fc3\u62a5\u544a\u7684\u5c0f\u65f6\u80fd\u8017\u4f30\u7b97\u7684\uff0c\u5e76\u8fdb\u884c\u4e86\u6269\u5c55\uff0c\u4ee5\u8003\u8651\u521b\u5efa\u548c\u7ef4\u62a4\u6570\u636e\u4e2d\u5fc3\u6240\u6d88\u8017\u7684\u989d\u5916\u80fd\u6e90\u3002\u8c37\u6b4c\u6570\u636e\u4e2d\u5fc3\u662f\u78b3\u4e2d\u548c\u7684\uff0c\u901a\u8fc7\u80fd\u6e90\u6548\u7387\u3001\u53ef\u518d\u751f\u80fd\u6e90\u8d2d\u4e70\u548c\u78b3\u62b5\u6d88\u7684\u7ed3\u5408\u5b9e\u73b0\u3002\u8fd9\u79cd\u78b3\u4e2d\u548c\u9002\u7528\u4e8e\u6211\u4eec\u7684\u8bad\u7ec3\u548c\u8fd0\u884c\u5b83\u4eec\u7684\u673a\u5668\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>3. \u6307\u4ee4\u5fae\u8c03<\/h2>\n<p>\u9884\u8bad\u7ec3\u6a21\u578b\u901a\u8fc7\u6539\u8fdb\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u8f6c\u53d8\u4e3a\u6307\u4ee4\u5fae\u8c03\u6a21\u578b\uff08\u89c1\u8868 6\uff09\u3002<\/p>\n<h3>\u6280\u672f<\/h3>\n<p>\u6211\u4eec\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u4f9d\u8d56\u4e8e\u6539\u8fdb\u7248\u672c\u7684\u77e5\u8bc6\u84b8\u998f\uff08Agarwal \u7b49\u4eba\uff0c2024\uff1bAnil \u7b49\u4eba\uff0c2018\uff1bHinton \u7b49\u4eba\uff0c2015\uff09\uff0c\u6765\u81ea\u4e00\u4e2a\u5927\u578b IT \u6559\u5e08\uff0c\u4ee5\u53ca\u57fa\u4e8e\u6539\u8fdb\u7248\u672c\u7684 BOND\uff08Sessa \u7b49\u4eba\uff0c2024\uff09\u3001WARM\uff08Ram \u7b49\u4eba\uff0c2024b\uff09\u548c WARP\uff08Ram \u7b49\u4eba\uff0c2024a\uff09\u7684 RL \u5fae\u8c03\u9636\u6bb5\u3002<\/p>\n<h3>\u5f3a\u5316\u5b66\u4e60\u76ee\u6807<\/h3>\n<p>\u6211\u4eec\u4f7f\u7528\u5404\u79cd\u5956\u52b1\u51fd\u6570\u6765\u63d0\u9ad8\u6709\u7528\u6027\u3001\u6570\u5b66\u3001\u7f16\u7801\u3001\u63a8\u7406\u3001\u6307\u4ee4\u9075\u5faa\u548c\u591a\u8bed\u8a00\u80fd\u529b\uff0c\u540c\u65f6\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u6a21\u578b\u5371\u5bb3\u6027\u3002\u8fd9\u5305\u62ec\u4ece\u4f7f\u7528\u4eba\u7c7b\u53cd\u9988\u6570\u636e\u8bad\u7ec3\u7684\u6743\u91cd\u805a\u5408\u5956\u52b1\u6a21\u578b\uff08Ram \u7b49\u4eba\uff0c2024b\uff09\u3001\u4ee3\u7801\u6267\u884c\u53cd\u9988\uff08Gehring \u7b49\u4eba\uff0c2024\uff09\u4ee5\u53ca\u89e3\u51b3\u6570\u5b66\u95ee\u9898\u7684\u771f\u5b9e\u5956\u52b1\uff08DeepSeek-AI\uff0c2025\uff1bLambert \u7b49\u4eba\uff0c2024\uff09\u4e2d\u8fdb\u884c\u5b66\u4e60\u3002<\/p>\n<h3>\u6570\u636e\u8fc7\u6ee4<\/h3>\n<p>\u6211\u4eec\u4ed4\u7ec6\u4f18\u5316\u540e\u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684\u6570\u636e\uff0c\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u6a21\u578b\u6027\u80fd\u3002\u6211\u4eec\u8fc7\u6ee4\u6389\u663e\u793a\u67d0\u4e9b\u4e2a\u4eba\u4fe1\u606f\u3001\u4e0d\u5b89\u5168\u6216\u6709\u6bd2\u7684\u6a21\u578b\u8f93\u51fa\u3001\u9519\u8bef\u7684\u81ea\u6211\u8bc6\u522b\u6570\u636e\u548c\u91cd\u590d\u793a\u4f8b\u3002\u5305\u62ec\u9f13\u52b1\u66f4\u597d\u7684\u4e0a\u4e0b\u6587\u5f52\u56e0\u3001\u7f13\u548c\u548c\u62d2\u7edd\u7684\u6570\u636e\u5b50\u96c6\uff0c\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u5e7b\u89c9\uff0c\u4e5f\u63d0\u9ad8\u4e86\u4e8b\u5b9e\u6027\u6307\u6807\u4e0a\u7684\u6027\u80fd\uff0c\u800c\u4e0d\u4f1a\u964d\u4f4e\u6a21\u578b\u5728\u5176\u4ed6\u6307\u6807\u4e0a\u7684\u6027\u80fd\u3002<\/p>\n<h3>[BOS] \u6807\u8bb0<\/h3>\n<p>\u5bf9\u4e8e PT \u548c IT \u6a21\u578b\uff0c\u6587\u672c\u4ee5 [BOS] \u6807\u8bb0\u5f00\u5934\uff0c\u9700\u8981\u663e\u5f0f\u6dfb\u52a0\uff0c\u56e0\u4e3a\u6587\u672c\u201c[BOS]\u201d\u4e0d\u4f1a\u6620\u5c04\u5230 [BOS] \u6807\u8bb0\u3002\u4f8b\u5982\uff0cFlax \u6709\u4e00\u4e2a\u9009\u9879\uff0cadd_bos = True\uff0c\u53ef\u4ee5\u5728\u5206\u8bcd\u65f6\u81ea\u52a8\u6dfb\u52a0\u6b64\u6807\u8bb0\u3002\u8868 4 \u663e\u793a\u4e86 IT \u6a21\u578b\u7684\u683c\u5f0f\u5316\u793a\u4f8b\uff0c<\/p>\n<h3>PT \u4e0e IT \u683c\u5f0f\u5316<\/h3>\n<p>\u6240\u6709\u6a21\u578b\u5171\u4eab\u76f8\u540c\u7684\u5206\u8bcd\u5668\uff0c\u4e00\u4e9b\u63a7\u5236\u6807\u8bb0\u4e13\u95e8\u7528\u4e8e IT \u683c\u5f0f\u5316\u3002\u4e00\u4e2a\u5173\u952e\u7684\u533a\u522b\u662f PT \u6a21\u578b\u5728\u751f\u6210\u7ed3\u675f\u65f6\u8f93\u51fa\u4e00\u4e2a\u00a0\u6807\u8bb0\uff0c\u800c IT \u6a21\u578b\u5728\u751f\u6210\u7ed3\u675f\u65f6\u8f93\u51fa\u4e00\u4e2a &lt;end_of_turn&gt; \u6807\u8bb0\uff0c\u5982\u8868 4 \u4e2d IT \u6240\u793a\u6240\u793a\u3002\u56e0\u6b64\uff0c\u5fae\u8c03\u4efb\u4e00\u6a21\u578b\u7c7b\u578b\u4e5f\u9700\u8981\u6dfb\u52a0\u5b83\u4eec\u5404\u81ea\u7684\u7ed3\u675f\u6807\u8bb0\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6392\u540d<\/th>\n<th>\u6a21\u578b<\/th>\n<th>Elo<\/th>\n<th>95% CI<\/th>\n<th>\u5f00\u653e<\/th>\n<th>\u7c7b\u578b<\/th>\n<th>#\u53c2\u6570\/#\u6fc0\u6d3b<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>1<\/td>\n<td>Grok-3-Preview-02-24<\/td>\n<td>1412<\/td>\n<td>+8\/-10<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>1<\/td>\n<td>GPT-4.5-Preview<\/td>\n<td>1411<\/td>\n<td>+11\/-11<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td>Gemini-2.0-Flash-Thinking-Exp-01-21<\/td>\n<td>1384<\/td>\n<td>+6\/-5<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td>Gemini-2.0-Pro-Exp-02-05<\/td>\n<td>1380<\/td>\n<td>+ 5\/-6<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td>ChatGPT-4o-latest (2025-01-29)<\/td>\n<td>1377<\/td>\n<td>+ 5\/-4<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>6<\/td>\n<td><a href=\"https:\/\/www.kdjingpai.com\/deepseek-r1nenglixiang\/\">DeepSeek-R1<\/a><\/td>\n<td>1363<\/td>\n<td>+8\/-6<\/td>\n<td>\u662f<\/td>\n<td>MoE<\/td>\n<td>671B\/37B<\/td>\n<\/tr>\n<tr>\n<td>6<\/td>\n<td>Gemini-2.0-Flash-001<\/td>\n<td>1357<\/td>\n<td>+6\/-5<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>8<\/td>\n<td>01-2024-12-17<\/td>\n<td>1352<\/td>\n<td>+ 4\/-6<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>9<\/td>\n<td>Gemma-3-27B-IT<\/td>\n<td>1338<\/td>\n<td>+8\/-9<\/td>\n<td>\u662f<\/td>\n<td>\u5bc6\u96c6<\/td>\n<td>27B<\/td>\n<\/tr>\n<tr>\n<td>9<\/td>\n<td>Qwen2.5-Max<\/td>\n<td>1336<\/td>\n<td>+ 7\/-5<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>9<\/td>\n<td>01-preview<\/td>\n<td>1335<\/td>\n<td>+4\/-3<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>9<\/td>\n<td>03-mini-high<\/td>\n<td>1329<\/td>\n<td>+8\/-6<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>13<\/td>\n<td>DeepSeek-V3<\/td>\n<td>1318<\/td>\n<td>+8\/-6<\/td>\n<td>\u662f<\/td>\n<td>MoE<\/td>\n<td>671B\/37B<\/td>\n<\/tr>\n<tr>\n<td>14<\/td>\n<td>GLM-4-Plus-0111<\/td>\n<td>1311<\/td>\n<td>+8\/-8<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>14<\/td>\n<td>Qwen-Plus-0125<\/td>\n<td>1310<\/td>\n<td>+7\/-5<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>14<\/td>\n<td><a href=\"https:\/\/www.kdjingpai.com\/claudeanquanfubai\/\">Claude<\/a> 3.7 Sonnet<\/td>\n<td>1309<\/td>\n<td>+ 9\/-11<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>14<\/td>\n<td>Gemini-2.0-Flash-Lite<\/td>\n<td>1308<\/td>\n<td>+5\/-5<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>18<\/td>\n<td>Step-2-16K-Exp<\/td>\n<td>1305<\/td>\n<td>+ 7\/-6<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>18<\/td>\n<td>03-mini<\/td>\n<td>1304<\/td>\n<td>+ 5\/-4<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>18<\/td>\n<td>0l-mini<\/td>\n<td>1304<\/td>\n<td>+4\/-3<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>18<\/td>\n<td>Gemini-1.5-Pro-002<\/td>\n<td>1302<\/td>\n<td>+3\/-3<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>28<\/td>\n<td>Meta-Llama-3.1-405B-Instruct-bf16<\/td>\n<td>1269<\/td>\n<td>+4\/-3<\/td>\n<td>\u662f<\/td>\n<td>\u5bc6\u96c6<\/td>\n<td>405B<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>38<\/td>\n<td>Llama-3.3-70B-Instruct<\/td>\n<td>1257<\/td>\n<td>+5\/-3<\/td>\n<td>\u662f<\/td>\n<td>\u5bc6\u96c6<\/td>\n<td>70B<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>39<\/td>\n<td>Qwen2.5-72B-Instruct<\/td>\n<td>1257<\/td>\n<td>+3\/-3<\/td>\n<td>\u662f<\/td>\n<td>\u5bc6\u96c6<\/td>\n<td>72B<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>59<\/td>\n<td>Gemma-2-27B-it<\/td>\n<td>1220<\/td>\n<td>+3\/-2<\/td>\n<td>\u662f<\/td>\n<td>\u5bc6\u96c6<\/td>\n<td>27B<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8868 5 | Gemma 3 27B IT \u6a21\u578b\u5728 Chatbot Arena\uff08Chiang \u7b49\u4eba\uff0c2024\uff09\u4e2d\u7684\u8bc4\u4f30\u3002\u6240\u6709\u6a21\u578b\u90fd\u901a\u8fc7\u4eba\u7c7b\u8bc4\u4f30\u8005\u8fdb\u884c\u76f2\u76ee\u7684\u5e76\u6392\u8bc4\u4f30\u3002\u6bcf\u4e2a\u6a21\u578b\u6839\u636e Elo \u8bc4\u5206\u7cfb\u7edf\u88ab\u8d4b\u4e88\u4e00\u4e2a\u5206\u6570\u3002Gemma-3-27B-IT \u7684\u6570\u5b57\u662f 2025 \u5e74 3 \u6708 8 \u65e5\u6536\u5230\u7684\u521d\u6b65\u7ed3\u679c\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>4. \u6700\u7ec8\u6a21\u578b\u8bc4\u4f30<\/h2>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u901a\u8fc7\u4e00\u7cfb\u5217\u81ea\u52a8\u5316\u57fa\u51c6\u6d4b\u8bd5\u548c\u4eba\u7c7b\u8bc4\u4f30\u4ee5\u53ca\u9759\u6001\u57fa\u51c6\u6d4b\u8bd5\uff08\u5982 MMLU\uff09\u6765\u8bc4\u4f30 IT \u6a21\u578b\u3002<\/p>\n<h3>4.1 LMSYS Chatbot Arena<\/h3>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u62a5\u544a\u4e86\u6211\u4eec\u7684 IT 27B \u6a21\u578b\u5728 <a href=\"https:\/\/www.kdjingpai.com\/lmsys-org\/\">LMSys<\/a> Chatbot Arena\uff08Chiang \u7b49\u4eba\uff0c2024\uff09\u4e2d\u5bf9\u5176\u4ed6\u6700\u5148\u8fdb\u6a21\u578b\u7684\u76f2\u6d4b\u5e76\u6392\u8bc4\u4f30\u4e2d\u7684\u8868\u73b0\u3002\u6211\u4eec\u5728\u8868 5 \u4e2d\u62a5\u544a\u4e86 Elo \u5206\u6570\u3002Gemma 3 27B IT\uff081338\uff09\u662f\u6392\u540d\u524d 10 \u7684\u6700\u4f73\u6a21\u578b\uff0c\u5176\u5f97\u5206\u9ad8\u4e8e\u5176\u4ed6\u975e\u601d\u7ef4\u5f00\u653e\u6a21\u578b\uff0c\u4f8b\u5982 DeepSeek-V3\uff081318\uff09\u3001LLaMA 3 405B\uff081257\uff09\u548c Qwen2.5-70B\uff081257\uff09\uff0c\u8fd9\u4e9b\u662f\u66f4\u5927\u7684\u6a21\u578b\u3002\u6700\u540e\uff0cGemma 3 \u7684 Elo \u660e\u663e\u9ad8\u4e8e Gemma 2\uff0c\u4e3a 1220\u3002\u8bf7\u6ce8\u610f\uff0cElo \u5206\u6570\u6ca1\u6709\u8003\u8651\u89c6\u89c9\u80fd\u529b\uff0c\u800c\u4e0a\u8ff0\u6a21\u578b\u90fd\u6ca1\u6709\u3002<\/p>\n<h3>4.2 \u6807\u51c6\u57fa\u51c6\u6d4b\u8bd5<\/h3>\n<p>\u5728\u8868 6 \u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u6211\u4eec\u7684\u6700\u7ec8\u6a21\u578b\u5728\u4e0d\u540c\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u8868\u73b0\uff0c\u4e0e\u6211\u4eec\u4e4b\u524d\u7684\u6a21\u578b\u8fed\u4ee3\u548c Gemini 1.5 \u76f8\u6bd4\u3002\u6211\u4eec\u6ca1\u6709\u76f4\u63a5\u4e0e\u7ecf\u5e38\u62a5\u544a\u81ea\u5df1\u8bc4\u4f30\u8bbe\u7f6e\u7684\u5916\u90e8\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\uff0c\u56e0\u4e3a\u5728\u6211\u4eec\u7684\u8bbe\u7f6e\u4e2d\u8fd0\u884c\u5b83\u4eec\u5e76\u4e0d\u80fd\u4fdd\u8bc1\u516c\u5e73\u6bd4\u8f83\u3002\u6211\u4eec\u9f13\u52b1\u8bfb\u8005\u5173\u6ce8\u7b2c\u4e09\u65b9\u9759\u6001\u6392\u884c\u699c\uff0c\u4ee5\u4fbf\u66f4\u516c\u5e73\u5730\u6bd4\u8f83\u4e0d\u540c\u6a21\u578b\u3002\u6211\u4eec\u5728\u9644\u5f55\u4e2d\u5305\u62ec\u4e86\u6211\u4eec\u7684\u6a21\u578b\u5728\u5176\u4ed6\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u989d\u5916\u8bc4\u4f30\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th><\/th>\n<th><\/th>\n<th>Gemini 1.5<\/th>\n<th><\/th>\n<th>Gemini 2.0<\/th>\n<th><\/th>\n<th><\/th>\n<th>Gemma 2<\/th>\n<th><\/th>\n<th><\/th>\n<th>Gemma 3<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td><\/td>\n<td>Flash<\/td>\n<td>Pro<\/td>\n<td>Flash<\/td>\n<td>Pro<\/td>\n<td>2B<\/td>\n<td>9B<\/td>\n<td>27B<\/td>\n<td>1B<\/td>\n<td>4B<\/td>\n<td>12B<\/td>\n<td>27B<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>MMLU-Pro<\/td>\n<td><\/td>\n<td>67.3<\/td>\n<td>75.8<\/td>\n<td>77.6<\/td>\n<td>79.1<\/td>\n<td>15.6<\/td>\n<td>46.8<\/td>\n<td>56.9<\/td>\n<td>14.7<\/td>\n<td>43.6<\/td>\n<td>60.6<\/td>\n<td>67.5<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>LiveCodeBench<\/td>\n<td><\/td>\n<td>30.7<\/td>\n<td>34.2<\/td>\n<td>34.5<\/td>\n<td>36.0<\/td>\n<td>1.2<\/td>\n<td>10.8<\/td>\n<td>20.4<\/td>\n<td>1.9<\/td>\n<td>12.6<\/td>\n<td>24.6<\/td>\n<td>29.7<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>Bird-SQL (dev)<\/td>\n<td><\/td>\n<td>45.6<\/td>\n<td>54.4<\/td>\n<td>58.7<\/td>\n<td>59.3<\/td>\n<td>12.2<\/td>\n<td>33.8<\/td>\n<td>46.7<\/td>\n<td>6.4<\/td>\n<td>36.3<\/td>\n<td>47.9<\/td>\n<td>54.4<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>GPQA Diamond<\/td>\n<td><\/td>\n<td>51.0<\/td>\n<td>59.1<\/td>\n<td>60.1<\/td>\n<td>64.7<\/td>\n<td>24.7<\/td>\n<td>28.8<\/td>\n<td>34.3<\/td>\n<td>19.2<\/td>\n<td>30.8<\/td>\n<td>40.9<\/td>\n<td>42.4<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>SimpleQA<\/td>\n<td><\/td>\n<td>8.6<\/td>\n<td>24.9<\/td>\n<td>29.9<\/td>\n<td>44.3<\/td>\n<td>2.8<\/td>\n<td>5.3<\/td>\n<td>9.2<\/td>\n<td>2.2<\/td>\n<td>4.0<\/td>\n<td>6.3<\/td>\n<td>10.0<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>FACTS Grounding<\/td>\n<td><\/td>\n<td>82.9<\/td>\n<td>80.0<\/td>\n<td>84.6<\/td>\n<td>82.8<\/td>\n<td>43.8<\/td>\n<td>62.0<\/td>\n<td>62.4<\/td>\n<td>36.4<\/td>\n<td>70.1<\/td>\n<td>75.8<\/td>\n<td>74.9<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>MATH<\/td>\n<td><\/td>\n<td>77.9<\/td>\n<td>86.5<\/td>\n<td>90.9<\/td>\n<td>91.8<\/td>\n<td>27.2<\/td>\n<td>49.4<\/td>\n<td>55.6<\/td>\n<td>48.0<\/td>\n<td>75.6<\/td>\n<td>83.8<\/td>\n<td>89.0<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>HiddenMath<\/td>\n<td><\/td>\n<td>47.2<\/td>\n<td>52.0<\/td>\n<td>63.5<\/td>\n<td>65.2<\/td>\n<td>1.8<\/td>\n<td>10.4<\/td>\n<td>14.8<\/td>\n<td>15.8<\/td>\n<td>43.0<\/td>\n<td>54.5<\/td>\n<td>60.3<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>MMMU (val)<\/td>\n<td><\/td>\n<td>62.3<\/td>\n<td>65.9<\/td>\n<td>71.7<\/td>\n<td>72.7<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td>48.8<\/td>\n<td>59.6<\/td>\n<td>64.9<\/td>\n<td><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8868 6 | \u4e0e Gemini 1.5\u3001Gemini 2.0 \u548c Gemma 2 \u7684\u6307\u4ee4\u5fae\u8c03\uff08IT\uff09\u6a21\u578b\u5728\u96f6\u6837\u672c\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6027\u80fd\u5bf9\u6bd4\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>5. \u6d88\u878d\u7814\u7a76<\/h2>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u91cd\u70b9\u5173\u6ce8\u67b6\u6784\u53d8\u5316\u7684\u5f71\u54cd\uff0c\u4ee5\u53ca\u4e00\u4e9b\u65b0\u6a21\u578b\u7684\u65b0\u89c6\u89c9\u80fd\u529b\u3002<\/p>\n<h3>5.1 \u9884\u8bad\u7ec3\u80fd\u529b\u63a2\u6d4b<\/h3>\n<p>\u6211\u4eec\u5728\u9884\u8bad\u7ec3\u671f\u95f4\u4f7f\u7528\u51e0\u4e2a\u6807\u51c6\u57fa\u51c6\u4f5c\u4e3a\u63a2\u6d4b\u5de5\u5177\uff0c\u4ee5\u786e\u4fdd\u6211\u4eec\u7684\u6a21\u578b\u80fd\u591f\u6355\u83b7\u4e00\u822c\u80fd\u529b\u3002\u5728\u56fe 2 \u4e2d\uff0c\u6211\u4eec\u6bd4\u8f83\u4e86 Gemma 2 \u548c 3 \u7684\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u8fd9\u4e9b\u4e00\u822c\u80fd\u529b\u4e0a\u7684\u8d28\u91cf\uff0c\u5373\u79d1\u5b66\u3001\u4ee3\u7801\u3001\u4e8b\u5b9e\u6027\u3001\u591a\u8bed\u8a00\u6027\u3001\u63a8\u7406\u548c\u89c6\u89c9\u3002\u9644\u5f55\u4e2d\u603b\u7ed3\u4e86\u8fd9\u4e9b\u56fe\u8868\u4e2d\u4f7f\u7528\u7684\u4e0d\u540c\u516c\u5171\u57fa\u51c6\u7684\u6027\u80fd\u7ec6\u8282\u3002\u603b\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u770b\u5230\u5c3d\u7ba1\u589e\u52a0\u4e86\u89c6\u89c9\uff0c\u65b0\u7248\u672c\u5728\u5927\u591a\u6570\u7c7b\u522b\u4e2d\u90fd\u6709\u6240\u6539\u8fdb\u3002\u6211\u4eec\u7279\u522b\u5173\u6ce8\u8fd9\u4e00\u7248\u672c\u7684\u591a\u8bed\u8a00\u6027\uff0c\u8fd9\u76f4\u63a5\u5f71\u54cd\u6211\u4eec\u6a21\u578b\u7684\u8d28\u91cf\u3002\u7136\u800c\uff0c\u5c3d\u7ba1\u4f7f\u7528\u4e86\u53bb\u6c61\u67d3\u6280\u672f\uff0c\u8fd9\u4e9b\u63a2\u6d4b\u603b\u662f\u5b58\u5728\u6c61\u67d3\u7684\u98ce\u9669\uff08Mirzadeh \u7b49\u4eba\uff0c2024\uff09\uff0c\u8fd9\u4f7f\u5f97\u66f4\u96be\u5f97\u51fa\u66f4\u660e\u786e\u7684\u7ed3\u8bba\u3002<\/p>\n<h3>5.2 \u5c40\u90e8\uff1a\u5168\u5c40\u6ce8\u610f\u529b\u5c42<\/h3>\n<p>\u6211\u4eec\u6d4b\u91cf\u4e86\u5c40\u90e8\u548c\u5168\u5c40\u81ea\u6ce8\u610f\u529b\u5c42\u7684\u53d8\u5316\u5bf9\u63a8\u7406\u8fc7\u7a0b\u4e2d\u6027\u80fd\u548c\u5185\u5b58\u6d88\u8017\u7684\u5f71\u54cd\u3002<\/p>\n<p><strong>\u5c40\u90e8\uff1a\u5168\u5c40\u6bd4\u7387\u3002<\/strong>\u00a0\u5728\u56fe 3 \u4e2d\uff0c\u6211\u4eec\u6bd4\u8f83\u4e86\u5c40\u90e8\u5230\u5168\u5c40\u6ce8\u610f\u529b\u5c42\u7684\u4e0d\u540c\u6bd4\u7387\u3002Gemma 2 \u6a21\u578b\u4f7f\u7528 1:1\uff0cGemma 3 \u4f7f\u7528 5:1\u3002\u6211\u4eec\u89c2\u5bdf\u5230\u6539\u53d8\u8fd9\u4e2a\u6bd4\u7387\u5bf9\u56f0\u60d1\u5ea6\u7684\u5f71\u54cd\u5f88\u5c0f\u3002<\/p>\n<p><strong>\u6ed1\u52a8\u7a97\u53e3\u5927\u5c0f\u3002<\/strong>\u00a0\u5728\u56fe 4 \u4e2d\uff0c\u6211\u4eec\u6bd4\u8f83\u4e86\u4e0d\u540c\u6ed1\u52a8\u7a97\u53e3\u5927\u5c0f\u5bf9\u4e8e\u4e0d\u540c\u5168\u5c40\uff1a\u5c40\u90e8\u6bd4\u7387\u914d\u7f6e\u7684\u5c40\u90e8\u6ce8\u610f\u529b\u5c42\u7684\u56f0\u60d1\u5ea6\u3002\u6ed1\u52a8\u7a97\u53e3\u53ef\u4ee5\u663e\u8457\u51cf\u5c11\uff0c\u800c\u4e0d\u4f1a\u5f71\u54cd\u56f0\u60d1\u5ea6\u3002<\/p>\n<p><strong>\u5bf9 KV \u7f13\u5b58\u5185\u5b58\u7684\u5f71\u54cd\u3002<\/strong>\u00a0\u5728\u56fe 5 \u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u6a21\u578b\u4e0e KV \u7f13\u5b58\u5185\u5b58\u4e4b\u95f4\u7684\u5e73\u8861\uff0c\u5728 32k \u6807\u8bb0\u7684\u4e0a\u4e0b\u6587\u4e2d\u3002\u6211\u4eec\u8003\u8651\u4e86\u4e00\u4e2a 2B \u6a21\u578b\uff0c\u5177\u6709\u4e0d\u540c\u7684\u5c40\u90e8\u5230\u5168\u5c40\u6bd4\u7387\u548c\u6ed1\u52a8\u7a97\u53e3\u5927\u5c0f\uff08sw\uff09\u3002\u6211\u4eec\u5c06\u5176\u4e0e\u5168\u5c40\u552f\u4e00\u914d\u7f6e\u8fdb\u884c\u6bd4\u8f83\uff0c\u8fd9\u662f Gemma 1 \u548c Llama \u4e2d\u4f7f\u7528\u7684\u6807\u51c6\u914d\u7f6e\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u201c\u5168\u5c40\u552f\u4e00\u201d\u914d\u7f6e\u5bfc\u81f4 60% \u7684\u5185\u5b58\u5f00\u9500\uff0c\u800c\u4f7f\u7528 1:3 \u548c 1024 \u7684\u6ed1\u52a8\u7a97\u53e3\uff08<strong>sw=1024<\/strong>\uff09\u65f6\u51cf\u5c11\u5230\u4e0d\u5230 15%\u3002\u5728\u56fe 6 \u4e2d\uff0c\u6211\u4eec\u8ba1\u7b97\u4e86 KV \u7f13\u5b58\u7684\u5185\u5b58\u4f7f\u7528\u91cf\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u51fd\u6570\uff0c\u4f7f\u7528\u6211\u4eec\u7684 2B \u67b6\u6784\uff08**L:G=5:1, sw=1024**\uff09\u4e0e\u4e00\u4e2a\u201c\u5168\u5c40\u552f\u4e00\u201d\u7684 2B \u6a21\u578b\u3002<\/p>\n<h3>5.3 \u542f\u7528\u957f\u4e0a\u4e0b\u6587<\/h3>\n<p>\u6211\u4eec\u4e0d\u662f\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3 128k \u5e8f\u5217\uff0c\u800c\u662f\u9884\u8bad\u7ec3\u6211\u4eec\u7684\u6a21\u578b\u4f7f\u7528 32k \u5e8f\u5217\uff0c\u7136\u540e\u5728\u9884\u8bad\u7ec3\u7ed3\u675f\u65f6\u5c06 4B\u300112B \u548c 27B \u6a21\u578b\u6269\u5c55\u5230 128k \u6807\u8bb0\uff0c\u540c\u65f6\u91cd\u65b0\u7f29\u653e RoPE\uff08Chen \u7b49\u4eba\uff0c2023\uff09\u3002\u6211\u4eec\u53d1\u73b0 8 \u7684\u7f29\u653e\u56e0\u5b50\u5728\u5b9e\u8df5\u4e2d\u6548\u679c\u826f\u597d\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4e0e Gemma 2 \u76f8\u6bd4\uff0c\u6211\u4eec\u8fd8\u5c06\u5168\u5c40\u81ea\u6ce8\u610f\u529b\u5c42\u7684 RoPE \u57fa\u9891\u4ece 10k \u589e\u52a0\u5230 1M\uff0c\u800c\u4fdd\u6301\u5c40\u90e8\u81ea\u6ce8\u610f\u529b\u5c42\u7684\u9891\u7387\u4e3a 10k\u3002\u5728\u56fe 7 \u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u4e0d\u540c\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u56f0\u60d1\u5ea6\u5f71\u54cd\u3002\u6211\u4eec\u7684\u6a21\u578b\u63a8\u5e7f\u5230 128k\uff0c\u4f46\u968f\u7740\u6211\u4eec\u7ee7\u7eed\u6269\u5c55\uff0c\u56f0\u60d1\u5ea6\u8fc5\u901f\u4e0b\u964d\u3002<\/p>\n<h3>5.4 \u5c0f\u578b\u4e0e\u5927\u578b\u6559\u5e08<\/h3>\n<p>\u4e00\u4e2a\u5e38\u89c1\u7684\u53d1\u73b0\u662f\uff0c\u8981\u8bad\u7ec3\u4e00\u4e2a\u5c0f\u6a21\u578b\uff0c\u6700\u597d\u4ece\u4e00\u4e2a\u5c0f\u6559\u5e08\u90a3\u91cc\u8fdb\u884c\u84b8\u998f\u3002<\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-2\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/f6854cc6d3b0304.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-2\" \/><\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-3\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d699452d4a68f51.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-3\" \/><\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-3\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d699452d4a68f51.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-3\" \/><\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-4\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/394a13f65673032.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-4\" \/><\/p>\n<p>\u6211\u4eec\u6000\u7591\u8fd9\u662f\u56e0\u4e3a\u8fd9\u4e9b\u7814\u7a76\u901a\u5e38\u5728\u8bbe\u7f6e\u4e2d\u4f7f\u7528\u66f4\u5dee\u7684\u6559\u5e08\u4f1a\u8d85\u8fc7\u4f7f\u7528\u66f4\u597d\u6559\u5e08\u7684\u76ca\u5904\u3002\u6211\u4eec\u4f7f\u7528\u4e24\u4e2a\u4e0d\u540c\u5927\u5c0f\u7684\u6559\u5e08\u8bad\u7ec3\u5b66\u751f\uff0c\u4e00\u4e2a\u5927\u4e00\u4e2a\u5c0f\uff0c\u7528\u4e8e\u4e0d\u540c\u7684\u8bad\u7ec3\u671f\u9650\u3002\u5728\u56fe 8 \u4e2d\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u5bf9\u4e8e\u77ed\u8bad\u7ec3\u671f\u9650\uff0c\u5c0f\u6559\u5e08\u66f4\u597d\uff0c\u4f46\u8d8b\u52bf\u5728\u66f4\u957f\u8bad\u7ec3\u4e2d\u9006\u8f6c\u3002<\/p>\n<h3>5.5 \u89c6\u89c9\u7f16\u7801\u5668<\/h3>\n<p><strong>\u56fe\u50cf\u5206\u8fa8\u7387\u7684\u5f71\u54cd\u3002<\/strong>\u00a0\u6211\u4eec\u4f7f\u7528\u57fa\u4e8e SigLIP\uff08Zhai \u7b49\u4eba\uff0c2023\uff09\u7684\u89c6\u89c9\u7f16\u7801\u5668\u3002\u89c6\u89c9\u7f16\u7801\u5668\u88ab\u51bb\u7ed3\uff0c\u53ea\u6709\u8bed\u8a00\u6a21\u578b\u63a5\u53d7\u8bad\u7ec3\u3002\u8fd9\u4e2a\u591a\u6a21\u6001\u6570\u636e\u4e2d\u7684\u6bcf\u4e2a\u56fe\u50cf\u90fd\u7531\u6765\u81ea\u76f8\u5e94\u89c6\u89c9\u7f16\u7801\u5668\u7684 256 \u4e2a\u56fe\u50cf\u6807\u8bb0\u8868\u793a\u3002\u56e0\u6b64\uff0c\u66f4\u9ad8\u5206\u8fa8\u7387\u7684\u7f16\u7801\u5668\u4f7f\u7528\u5e73\u5747\u6c60\u5316\u6765\u5c06\u5176\u8f93\u51fa\u51cf\u5c11\u5230 256 \u4e2a\u6807\u8bb0\u3002\u4f8b\u5982\uff0c896 \u5206\u8fa8\u7387\u7f16\u7801\u5668\u5728\u5176\u8f93\u51fa\u4e0a\u6709 4&#215;4 \u7684\u5e73\u5747\u6c60\u5316\u3002\u5982\u8868 7 \u6240\u793a\uff0c\u66f4\u9ad8\u5206\u8fa8\u7387\u7684\u7f16\u7801\u5668\u8868\u73b0\u4f18\u4e8e\u8f83\u5c0f\u7684\u7f16\u7801\u5668\u3002<\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-5\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/74b9445453fd4d0.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-5\" \/><\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-6\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/a825cbfe4be0010.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-6\" \/><\/p>\n<p><strong>\u5e73\u79fb\u548c\u626b\u63cf\uff08P&amp;S)\u3002<\/strong>\u00a0P&amp;S \u4f7f\u5f97\u4ee5\u63a5\u8fd1\u5176\u539f\u751f\u957f\u5bbd\u6bd4\u548c\u56fe\u50cf\u5206\u8fa8\u7387\u6355\u83b7\u56fe\u50cf\u6210\u4e3a\u53ef\u80fd\u3002\u5728\u8868 8 \u4e2d\uff0c\u6211\u4eec\u6bd4\u8f83\u4e86\u6211\u4eec\u7684 27B IT \u6a21\u578b\uff0c\u6709\u548c\u6ca1\u6709 P&amp;S\u3002\u5982\u9884\u671f\u7684\u90a3\u6837\uff0c\u80fd\u591f\u4ee5\u63a5\u8fd1\u539f\u751f\u5206\u8fa8\u7387\u5904\u7406\u56fe\u50cf\u7684\u80fd\u529b\u5bf9\u4e8e\u9700\u8981\u67d0\u79cd\u5f62\u5f0f\u8bfb\u53d6\u56fe\u50cf\u4e0a\u7684\u6587\u672c\u7684\u4efb\u52a1\u6709\u5f88\u5927\u5e2e\u52a9\uff0c\u8fd9\u5bf9\u4e8e\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u6765\u8bf4\u5c24\u4e3a\u91cd\u8981\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>6. \u8bb0\u5fc6\u548c\u9690\u79c1<\/h2>\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u53ef\u80fd\u4f1a\u4ea7\u751f\u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684\u67d0\u4e9b\u6587\u672c\u7684\u8fd1\u4e4e\u590d\u5236\uff08Biderman \u7b49\u4eba\uff0c2023\uff1bCarlini \u7b49\u4eba\uff0c2021\uff0c2022\uff1bIppolito \u7b49\u4eba\uff0c2022\uff1bNasr \u7b49\u4eba\uff0c2023\uff09\u3002\u4e00\u4e9b\u4ee5\u524d\u7684\u62a5\u544a\u53d1\u5e03\u4e86\u5ba1\u8ba1\uff0c\u901a\u8fc7\u6d4b\u91cf\u8bb0\u5fc6\u7387\uff08Anil \u7b49\u4eba\uff0c2023\uff1bChowdhery \u7b49\u4eba\uff0c2022\uff1bGemini \u56e2\u961f\uff0c2023\uff0c2024\uff1bGemma \u56e2\u961f\uff0c2024a\uff0cb\uff1bLLaMa \u56e2\u961f\uff0c2024\uff09\u6765\u91cf\u5316\u8fd9\u79cd\u98ce\u9669\u3002\u8fd9\u4e2a\u201c\u8bb0\u5fc6\u7387\u201d1 \u5b9a\u4e49\u4e3a\u6a21\u578b\u751f\u6210\u4e0e\u8bad\u7ec3\u6570\u636e\u5339\u914d\u7684\u751f\u6210\u4e0e\u6240\u6709\u6a21\u578b\u751f\u6210\u4e4b\u95f4\u7684\u6bd4\u7387\uff0c\u4f7f\u7528\u4ee5\u4e0b\u8bbe\u7f6e\u3002\u6211\u4eec\u9075\u5faa Gemma \u56e2\u961f\uff082024b\uff09\u4e2d\u63cf\u8ff0\u7684\u65b9\u6cd5\u6765\u6d4b\u91cf\u5b83\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u4ece\u8bad\u7ec3\u6570\u636e\u4e2d\u7edf\u4e00\u5206\u5e03\u7684\u8f83\u5927\u90e8\u5206\u8fdb\u884c\u5b50\u91c7\u6837\uff0c\u5e76\u4f7f\u7528\u957f\u5ea6\u4e3a 50 \u7684\u524d\u7f00\u548c\u957f\u5ea6\u4e3a 50 \u7684\u540e\u7f00\u6d4b\u8bd5\u5bf9\u6b64\u5185\u5bb9\u7684\u53ef\u53d1\u73b0\u63d0\u53d6\uff08Nasr \u7b49\u4eba\uff0c2023\uff09\u3002\u5982\u679c\u5ef6\u7eed\u4e2d\u7684\u6240\u6709\u6807\u8bb0\u4e0e\u6e90\u540e\u7f00\u5339\u914d\uff0c\u5219\u5c06\u6587\u672c\u8868\u793a\u4e3a\u201c\u5b8c\u5168\u8bb0\u5fc6\u201d\uff0c\u5982\u679c\u5b83\u4eec\u5339\u914d\u5230\u7f16\u8f91\u8ddd\u79bb\u4e3a 10% \u7684\u8303\u56f4\u5185\uff0c\u5219\u8868\u793a\u4e3a\u201c\u8fd1\u4f3c\u8bb0\u5fc6\u201d\u3002<\/p>\n<p>\u56fe 9 \u6bd4\u8f83\u4e86 Gemma \u548c Gemini \u6a21\u578b\u7684\u8bb0\u5fc6\u7387\uff1b\u8fd9\u4e9b\u6a21\u578b\u6309\u9006\u65f6\u95f4\u987a\u5e8f\u6392\u5217\uff0c\u6700\u65b0\u7684 Gemma 3 \u6a21\u578b\u5728\u5de6\u8fb9\u3002\u6211\u4eec\u53d1\u73b0\uff0cGemma 3 \u6a21\u578b\u4ee5\u6bd4\u4e4b\u524d\u6a21\u578b\u4f4e\u5f97\u591a\u7684\u901f\u7387\u8bb0\u5fc6\u957f\u7bc7\u6587\u672c\uff08\u6ce8\u610f\u5bf9\u6570 y \u8f74\uff09\u3002\u6211\u4eec\u89c2\u5bdf\u5230 4B\u300112B \u548c 27B \u6a21\u578b\u4e4b\u95f4\u7684\u8bb0\u5fc6\u7387\u53ea\u6709\u5fae\u5c0f\u7684\u5dee\u5f02\uff0c1B \u8bb0\u5fc6\u8f83\u5c11\u6bd4\u5176\u4ed6\u66f4\u5927\u7684\u6a21\u578b\u3002\u6b64\u5916\uff0c\u6211\u4eec\u53d1\u73b0\u66f4\u5927\u6bd4\u4f8b\u7684\u6587\u672c\u88ab\u8868\u5f81\u4e3a\u8fd1\u4f3c\u8bb0\u5fc6\uff0c\u4e0e\u5b8c\u5168\u8bb0\u5fc6\u76f8\u6bd4\uff0c\u8fd1\u4f3c\u8bb0\u5fc6\u7684\u76f8\u5bf9\u589e\u52a0\u5927\u7ea6\u662f 24 \u500d\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u7814\u7a76\u4e86\u751f\u6210\u4e2d\u53ef\u80fd\u5305\u542b\u4e2a\u4eba\u4fe1\u606f\u7684\u901f\u7387\u3002\u4e3a\u4e86\u8bc6\u522b\u6f5c\u5728\u7684\u4e2a\u4eba\u4fe1\u606f\uff0c\u6211\u4eec\u4f7f\u7528\u8c37\u6b4c\u4e91\u654f\u611f\u6570\u636e\u4fdd\u62a4\uff08SDP\uff09\u670d\u52a1\u30022 SDP \u4f7f\u7528\u5e7f\u6cdb\u7684\u68c0\u6d4b\u89c4\u5219\u6765\u8bc6\u522b\u53ef\u80fd\u5305\u542b\u4e2a\u4eba\u4fe1\u606f\u7684\u6587\u672c\u3002SDP \u65e8\u5728\u5177\u6709\u9ad8\u53ec\u56de\u7387\uff0c\u4e0d\u8003\u8651\u4fe1\u606f\u53ef\u80fd\u51fa\u73b0\u7684\u73af\u5883\uff0c\u8fd9\u5bfc\u81f4\u8bb8\u591a\u8bef\u62a5\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u53ef\u80fd\u9ad8\u4f30\u4e86\u8f93\u51fa\u4e2d\u5305\u542b\u7684\u4e2a\u4eba\u4fe1\u606f\u7684\u771f\u5b9e\u6570\u91cf\u3002SDP \u8fd8\u63d0\u4f9b\u5e7f\u6cdb\u7684\u4e25\u91cd\u7a0b\u5ea6\u7ea7\u522b\uff1a\u4f4e\u3001\u4e2d\u548c\u9ad8\u3002\u5982\u679c SDP \u5c06\u5176\u5206\u7c7b\u4e3a\u4efb\u4f55\u4e25\u91cd\u7a0b\u5ea6\u7ea7\u522b\u7684\u4e2a\u4eba\u4fe1\u606f\uff0c\u6211\u4eec\u5219\u5c06\u6587\u672c\u5206\u7c7b\u4e3a\u4e2a\u4eba\u4fe1\u606f\u3002\u6211\u4eec\u89c2\u5bdf\u5230\u6240\u6709 Gemma 3 \u6a21\u578b\u7684\u8f93\u51fa\u4e2d\u90fd\u6ca1\u6709\u4e2a\u4eba\u4fe1\u606f\u3002\u8fd9\u8868\u660e\u5728\u5206\u7c7b\u4e3a\u8bb0\u5fc6\u7684\u8f93\u51fa\u4e2d\uff0c\u4e2a\u4eba\u6570\u636e\u7684\u901f\u7387\u4f4e\u4e8e\u6211\u4eec\u7684\u68c0\u6d4b\u9608\u503c\u3002<\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-7\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/7ac2e62fb32276f.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-7\" \/><\/p>\n<p><img decoding=\"async\" title=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-8\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/0922010a8cd74d1.jpg\" alt=\"Gemma 3 \u6280\u672f\u62a5\u544a\u4e2d\u6587\u7248-8\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>7. \u8d23\u4efb\u3001\u5b89\u5168\u3001\u4fdd\u969c<\/h2>\n<p>\u8d23\u4efb\u3001\u5b89\u5168\u548c\u4fdd\u969c\u5728 Gemma \u6a21\u578b\u5f00\u53d1\u4e2d\u81f3\u5173\u91cd\u8981\u3002\u4e3a\u4e86\u51cf\u5c11\u5bf9 Gemma 3 \u7528\u6237\u7684\u5f71\u54cd\uff0c\u6211\u4eec\u7ee7\u7eed\u6574\u5408\u589e\u5f3a\u7684\u5185\u90e8\u5b89\u5168\u6d41\u7a0b\uff0c\u8fd9\u4e9b\u6d41\u7a0b\u8d2f\u7a7f\u4e8e\u5f00\u53d1\u5de5\u4f5c\u6d41\u7a0b\uff0c\u4e0e\u6700\u8fd1\u7684\u8c37\u6b4c AI \u6a21\u578b\uff08Gemini \u56e2\u961f\uff0c2024\uff09\u4fdd\u6301\u4e00\u81f4\u3002\u8fd9\u4fa7\u91cd\u4e8e\u8bad\u7ec3\u65f6\u7684\u5b89\u5168\u7f13\u89e3\u63aa\u65bd\uff0c\u4ee5\u53ca\u5bf9\u65b0\u7684\u56fe\u50cf\u5230\u6587\u672c\u529f\u80fd\u7684\u5f3a\u5927\u800c\u900f\u660e\u7684\u6a21\u578b\u8bc4\u4f30\u3002<\/p>\n<h3>7.1 \u6cbb\u7406\u4e0e\u8bc4\u4f30<\/h3>\n<p>\u6211\u4eec\u5bf9\u8bc4\u4f30 Gemma \u7684\u76ca\u5904\u548c\u98ce\u9669\u7684\u65b9\u6cd5\u53cd\u6620\u4e86 Gemma 1\uff08Gemma \u56e2\u961f\uff0c2024a\uff09\u4e2d\u6240\u6982\u8ff0\u7684\u65b9\u6cd5\uff0c\u5e76\u8003\u8651\u4e86\u652f\u6301\u6a21\u5f0f\u7684\u53d8\u5316\u3002\u6211\u4eec\u7ee7\u7eed\u76f8\u4fe1\uff0cAI \u7684\u5f00\u653e\u6027\u53ef\u4ee5\u5c06\u8fd9\u4e9b\u6280\u672f\u7684\u76ca\u5904\u4f20\u64ad\u5230\u6574\u4e2a\u793e\u4f1a\uff0c\u4f46\u5fc5\u987b\u6839\u636e\u53ef\u80fd\u9020\u6210\u7684\u5371\u5bb3\u98ce\u9669\u8fdb\u884c\u8bc4\u4f30\uff0c\u65e0\u8bba\u662f\u5728\u4e2a\u4eba\u8fd8\u662f\u673a\u6784\u5c42\u9762\uff08Weidinger \u7b49\u4eba\uff0c2021\uff09\u3002\u81ea Gemma \u9996\u6b21\u53d1\u5e03\u4ee5\u6765\uff0c\u6211\u4eec\u5df2\u7ecf\u770b\u5230\u8fd9\u4e9b\u6a21\u578b\u63a8\u52a8\u4e86\u591a\u9879\u793e\u4f1a\u6709\u76ca\u7684\u5e94\u7528\uff0c\u4f8b\u5982\u6211\u4eec\u81ea\u5df1\u7684 ShieldGemma 2\uff0c\u4e00\u4e2a\u4f7f\u7528 Gemma 3 \u6784\u5efa\u7684 4B \u56fe\u50cf\u5b89\u5168\u5206\u7c7b\u5668\uff0c\u5b83\u63d0\u4f9b\u4e86\u4e00\u4e2a\u73b0\u6210\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u7528\u4e8e\u56fe\u50cf\u5b89\u5168\uff0c\u8f93\u51fa\u5371\u9669\u5185\u5bb9\u3001\u6027\u5265\u524a\u548c\u66b4\u529b\u7c7b\u522b\u7684\u5b89\u5168\u6807\u7b7e\u3002<\/p>\n<p>\u53d1\u5e03 Gemma 3 \u6a21\u578b\u9700\u8981\u5bf9\u6a21\u578b\u80fd\u529b\u7684\u53d8\u5316\u7ed9\u4e88\u7279\u522b\u5173\u6ce8\uff0c\u5e76\u5bc6\u5207\u76d1\u63a7\u73b0\u6709\u591a\u6a21\u6001 LLM\uff08Lin \u7b49\u4eba\uff0c2024\uff09\u7684\u4e0d\u65ad\u6f14\u53d8\u7684\u98ce\u9669\uff0c\u4ee5\u53ca\u4e86\u89e3\u6a21\u578b\u5728\u91ce\u5916\u7684\u4f7f\u7528\u65b9\u5f0f\u3002\u5c3d\u7ba1\u6211\u4eec\u5c1a\u672a\u6536\u5230\u4efb\u4f55\u5173\u4e8e Gemma \u7684\u6076\u610f\u4f7f\u7528\u62a5\u544a\uff0c\u4f46\u6211\u4eec\u4ecd\u7136\u81f4\u529b\u4e8e\u8c03\u67e5\u4efb\u4f55\u6b64\u7c7b\u62a5\u544a\uff0c\u5e76\u4e0e\u5b66\u672f\u754c\u548c\u5f00\u53d1\u8005\u793e\u533a\u5408\u4f5c\uff0c\u4ee5\u53ca\u8fdb\u884c\u6211\u4eec\u81ea\u5df1\u7684\u76d1\u63a7\uff0c\u4ee5\u6807\u8bb0\u6b64\u7c7b\u6848\u4f8b\u3002<\/p>\n<p>\u5c3d\u7ba1\u80fd\u529b\u6709\u6240\u63d0\u5347\uff0c\u6211\u4eec\u8ba4\u4e3a\uff0c\u9274\u4e8e\u6709\u66f4\u591a\u5f3a\u5927\u7684\u5f00\u653e\u6a21\u578b\u53ef\u7528\uff0c\u8fd9\u4e2a\u7248\u672c\u5bf9\u6574\u4f53\u98ce\u9669\u683c\u5c40\u7684\u5f71\u54cd\u53ef\u4ee5\u5ffd\u7565\u4e0d\u8ba1\u3002<\/p>\n<h3>7.2 \u5b89\u5168\u653f\u7b56\u548c\u8bad\u7ec3\u65f6\u7f13\u89e3\u63aa\u65bd<\/h3>\n<p>Gemma \u5b89\u5168\u65b9\u6cd5\u7684\u4e00\u4e2a\u5173\u952e\u652f\u67f1\u662f\u4e0e\u8c37\u6b4c\u5b89\u5168\u653f\u7b56\u4fdd\u6301\u4e00\u81f4\uff0c\u4e0e Gemini \u6a21\u578b\uff08Gemini \u56e2\u961f\uff0c2023\uff09\u4fdd\u6301\u4e00\u81f4\u3002\u5b83\u4eec\u65e8\u5728\u5e2e\u52a9\u9632\u6b62\u6211\u4eec\u7684\u6a21\u578b\u4ea7\u751f\u6709\u5bb3\u5185\u5bb9\uff0c\u5373\uff1a<\/p>\n<p>\u2022 \u513f\u7ae5\u6027\u8650\u5f85\u548c\u5265\u524a<br \/>\n\u2022 \u6cc4\u9732\u53ef\u80fd\u5bfc\u81f4\u5371\u5bb3\u7684\u4e2a\u4eba\u8eab\u4efd\u4fe1\u606f\uff08\u4f8b\u5982\uff0c\u793e\u4f1a\u5b89\u5168\u53f7\u7801\uff09<br \/>\n\u2022 \u4ec7\u6068\u8a00\u8bba\u548c\u9a9a\u6270<br \/>\n\u2022 \u5371\u9669\u6216\u6076\u610f\u5185\u5bb9\uff08\u5305\u62ec\u63d0\u5021\u81ea\u6211\u4f24\u5bb3\u6216\u6307\u5bfc\u6709\u5bb3\u6d3b\u52a8\uff09<br \/>\n\u2022 \u8272\u60c5\u5185\u5bb9<br \/>\n\u2022 \u4e0e\u79d1\u5b66\u6216\u533b\u5b66\u5171\u8bc6\u76f8\u53cd\u7684\u533b\u7597\u5efa\u8bae<\/p>\n<p>\u6211\u4eec\u8fdb\u884c\u4e86\u5927\u91cf\u7684\u5b89\u5168\u8fc7\u6ee4\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u6570\u636e\uff0c\u4ee5\u51cf\u5c11\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u68c0\u67e5\u70b9\u4ea7\u751f\u6709\u5bb3\u5185\u5bb9\u7684\u53ef\u80fd\u6027\u3002\u5bf9\u4e8e\u5fae\u8c03\u6a21\u578b\uff0c\u6211\u4eec\u8fd8\u4f7f\u7528 SFT \u548c RLHF \u6765\u5f15\u5bfc\u6a21\u578b\u8fdc\u79bb\u4e0d\u826f\u884c\u4e3a\u3002<\/p>\n<h3>7.3 \u4fdd\u969c\u8bc4\u4f30<\/h3>\n<p>\u6211\u4eec\u8fd8\u5c06\u6211\u4eec\u7684 IT \u6a21\u578b\u901a\u8fc7\u4e00\u7cfb\u5217\u57fa\u7ebf\u4fdd\u969c\u8bc4\u4f30\uff0c\u4ee5\u4e86\u89e3\u6211\u4eec\u7684\u6a21\u578b\u53ef\u80fd\u9020\u6210\u7684\u6f5c\u5728\u5371\u5bb3\u3002\u7531\u4e8e\u6211\u4eec\u63d0\u5021\u5f00\u653e\u6a21\u578b\uff0c\u6211\u4eec\u4e5f\u8ba4\u8bc6\u5230\uff0c\u6743\u91cd\u53d1\u5e03\u7684\u4e0d\u53ef\u9006\u6027\u9700\u8981\u4e25\u683c\u7684\u8bc4\u4f30\u98ce\u9669\u3002\u6211\u4eec\u7684\u5185\u90e8\u5b89\u5168\u6d41\u7a0b\u76f8\u5e94\u5730\u8bbe\u8ba1\uff0c\u5bf9\u4e8e\u4e4b\u524d\u7684 Gemma \u6a21\u578b\uff0c\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u4e0e\u6781\u7aef\u98ce\u9669\u76f8\u5173\u7684\u80fd\u529b\u7684\u8bc4\u4f30\uff08Phuong \u7b49\u4eba\uff0c2024\uff1bShevlane \u7b49\u4eba\uff0c2023\uff09\u3002\u968f\u7740\u6211\u4eec\u7ee7\u7eed\u5f00\u53d1\u548c\u5171\u4eab\u5f00\u653e\u6a21\u578b\uff0c\u6211\u4eec\u5c06\u9075\u5faa\u8fd9\u6837\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\uff0c\u5f7b\u5e95\u8bc4\u4f30\u4e00\u4e2a\u66f4\u5f3a\u5927\u7684\u6a21\u578b\u901a\u5e38\u4e3a\u80fd\u529b\u8f83\u4f4e\u7684\u6a21\u578b\u63d0\u4f9b\u8db3\u591f\u7684\u4fdd\u969c\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4f18\u5148\u8003\u8651\u4e3a Gemma 3 \u7b80\u5316\u8bc4\u4f30\u96c6\uff0c\u4e3a\u7279\u5b9a\u6a21\u578b\u53ef\u80fd\u5b58\u5728\u7684\u6f5c\u5728\u66f4\u9ad8\u98ce\u9669\u7684\u60c5\u51b5\u4fdd\u7559\u6df1\u5165\u7684\u5371\u9669\u80fd\u529b\u8bc4\u4f30\uff08\u5982\u4e0b\u9762\u5173\u4e8e CBRN \u8bc4\u4f30\u7684\u63cf\u8ff0\uff09\u3002\u6211\u4eec\u5728\u5e73\u8861\u5f00\u53d1\u901f\u5ea6\u4e0e\u76ee\u6807\u5b89\u5168\u6d4b\u8bd5\u7684\u540c\u65f6\uff0c\u786e\u4fdd\u6211\u4eec\u7684\u8bc4\u4f30\u662f\u4e13\u6ce8\u4e14\u9ad8\u6548\u7684\uff0c\u540c\u65f6\u575a\u6301\u6211\u4eec\u5728 Frontier Safety Framework \u4e2d\u6982\u8ff0\u7684\u627f\u8bfa\u3002<\/p>\n<h3>\u57fa\u7ebf\u8bc4\u4f30<\/h3>\n<p>\u57fa\u7ebf\u4fdd\u969c\u6355\u83b7\u4e86\u6a21\u578b\u8fdd\u53cd\u5b89\u5168\u653f\u7b56\u7684\u6bd4\u7387\uff0c\u4f7f\u7528\u5927\u91cf\u5408\u6210\u7684\u5bf9\u6297\u7528\u6237\u67e5\u8be2\uff0c\u5e76\u4f7f\u7528\u4eba\u7c7b\u8bc4\u4f30\u8005\u5c06\u7b54\u6848\u6807\u8bb0\u4e3a\u662f\u5426\u8fdd\u53cd\u653f\u7b56\u3002\u603b\u4f53\u800c\u8a00\uff0cGemma 3 \u5bf9\u8fd9\u4e9b\u5b89\u5168\u653f\u7b56\u7684\u8fdd\u89c4\u7387\u975e\u5e38\u4f4e\u3002<\/p>\n<h3>\u5316\u5b66\u3001\u751f\u7269\u3001\u653e\u5c04\u6027\u548c\u6838\uff08CBRN\uff09\u77e5\u8bc6<\/h3>\n<p>\u7531\u4e8e\u5728 STEM \u76f8\u5173\u4efb\u52a1\u4e0a\u7684\u589e\u5f3a\u6027\u80fd\uff0c\u6211\u4eec\u4f7f\u7528\u5185\u90e8\u6570\u636e\u96c6\u7684\u5c01\u95ed\u5f0f\u77e5\u8bc6\u578b\u591a\u9009\u9898\u8bc4\u4f30\u4e86\u4e0e\u751f\u7269\u3001\u653e\u5c04\u6027\u548c\u6838\u98ce\u9669\u76f8\u5173\u7684\u77e5\u8bc6\u3002\u5bf9\u4e8e\u5316\u5b66\u77e5\u8bc6\u7684\u8bc4\u4f30\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u4e00\u79cd\u5c01\u95ed\u578b\u77e5\u8bc6\u578b\u65b9\u6cd5\u5bf9\u5316\u5b66\u5371\u5bb3\u8fdb\u884c\u8bc4\u4f30\uff0c\u8be5\u65b9\u6cd5\u7531 Macknight \u7b49\u4eba\u5f00\u53d1\u3002\u6211\u4eec\u7684\u8bc4\u4f30\u8868\u660e\uff0cGemma 3 \u6a21\u578b\u5728\u8fd9\u4e9b\u9886\u57df\u4e2d\u7684\u77e5\u8bc6\u8f83\u4f4e\u3002<\/p>\n<h3>7.4 \u6211\u4eec\u8d1f\u8d23\u4efb\u7684\u5f00\u653e\u6a21\u578b\u65b9\u6cd5<\/h3>\n<p>\u8bbe\u8ba1\u5b89\u5168\u3001\u53ef\u9760\u548c\u8d1f\u8d23\u4efb\u7684\u5e94\u7528\u9700\u8981\u7cfb\u7edf\u7ea7\u7684\u65b9\u6cd5\uff0c\u52aa\u529b\u7f13\u89e3\u4e0e\u6bcf\u4e2a\u7279\u5b9a\u7528\u4f8b\u548c\u73af\u5883\u76f8\u5173\u7684\u98ce\u9669\u3002\u6211\u4eec\u5c06\u7ee7\u7eed\u91c7\u7528\u4e0e\u6211\u4eec\u7684\u6a21\u578b\u6f5c\u5728\u98ce\u9669\u6210\u6bd4\u4f8b\u7684\u8bc4\u4f30\u548c\u5b89\u5168\u7f13\u89e3\u63aa\u65bd\uff0c\u5e76\u4e14\u53ea\u6709\u5f53\u6211\u4eec\u786e\u4fe1\u6536\u76ca\u660e\u663e\u8d85\u8fc7\u53ef\u9884\u89c1\u7684\u98ce\u9669\u65f6\uff0c\u624d\u4f1a\u4e0e\u793e\u533a\u5171\u4eab\u8fd9\u4e9b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>8. \u8ba8\u8bba\u4e0e\u7ed3\u8bba<\/h2>\n<p>\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86 Gemma 3\uff0c\u8fd9\u662f Gemma \u7cfb\u5217\u5f00\u6e90\u8bed\u8a00\u6a21\u578b\u7684\u6700\u65b0\u6210\u5458\uff0c\u7528\u4e8e\u6587\u672c\u3001\u56fe\u50cf\u548c\u4ee3\u7801\u3002\u5728\u8fd9\u4e2a\u7248\u672c\u4e2d\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u6dfb\u52a0\u56fe\u50cf\u7406\u89e3\u548c\u957f\u4e0a\u4e0b\u6587\uff0c\u540c\u65f6\u63d0\u9ad8\u591a\u8bed\u8a00\u80fd\u529b\u548c STEM \u76f8\u5173\u80fd\u529b\u3002\u6211\u4eec\u7684\u6a21\u578b\u5927\u5c0f\u548c\u67b6\u6784\u65e8\u5728\u4e0e\u6807\u51c6\u786c\u4ef6\u517c\u5bb9\uff0c\u6211\u4eec\u7684\u5927\u90e8\u5206\u67b6\u6784\u6539\u8fdb\u90fd\u662f\u4e3a\u4e86\u9002\u5e94\u8fd9\u79cd\u786c\u4ef6\uff0c\u540c\u65f6\u4fdd\u6301\u6027\u80fd\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Gemma 3 \u5173\u952e\u4fe1\u606f\u603b\u7ed3 \u4e00\u3001\u5173\u952e\u6307\u6807 \u53c2\u6570 \u8be6\u60c5 \u6a21\u578b\u89c4\u6a21 1 \u4ebf\u5230 270 \u4ebf\u53c2\u6570\uff0c\u5171\u56db\u4e2a\u7248\u672c\uff1a1B\u30014B\u300112B\u300127B \u67b6\u6784 \u57fa\u4e8e Transformer \u7684\u89e3\u7801\u5668\u4e13\u7528\u67b6\u6784\uff0c\u7ee7\u627f\u81ea Gemma 2\uff0c\u5e76\u8fdb\u884c\u591a\u9879\u6539\u8fdb \u591a\u6a21\u6001\u80fd\u529b \u652f&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-28399","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/28399","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=28399"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/28399\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=28399"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=28399"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=28399"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}