{"id":27775,"date":"2025-03-06T09:09:36","date_gmt":"2025-03-06T01:09:36","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=27775"},"modified":"2025-03-07T11:22:26","modified_gmt":"2025-03-07T03:22:26","slug":"xiaomoxingdanengba","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/xiaomoxingdanengba\/","title":{"rendered":"\u5c0f\u6a21\u578b\uff0c\u5927\u80fd\u91cf\uff1aQwQ-32B \u4ee5 1\/20 \u53c2\u6570\u786c\u521a\u6ee1\u8840 DeepSeek-R1"},"content":{"rendered":"<p>\u8fd1\u671f\uff0cAI \u9886\u57df\u6d8c\u73b0\u51fa\u4ee4\u4eba\u77a9\u76ee\u7684\u8fdb\u5c55\uff0c\u5c24\u5176\u662f\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u63a8\u7406\u80fd\u529b\u63d0\u5347\u65b9\u9762\u3002\u5176\u4e2d\uff0c\u5f3a\u5316\u5b66\u4e60 (Reinforcement Learning, RL) \u6b63\u9010\u6e10\u6210\u4e3a\u7a81\u7834\u4f20\u7edf\u6a21\u578b\u6027\u80fd\u74f6\u9888\u7684\u5173\u952e\u6280\u672f\u3002\u4e0d\u5c11\u7814\u7a76\u5df2\u8bc1\u5b9e\uff0cRL \u80fd\u591f\u663e\u8457\u589e\u5f3a\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002\u4f8b\u5982\uff0cDeepSeek R1 \u6a21\u578b\u901a\u8fc7\u6574\u5408\u51b7\u542f\u52a8\u6570\u636e\u548c\u591a\u9636\u6bb5\u8bad\u7ec3\uff0c\u5b9e\u73b0\u4e86\u6df1\u5ea6\u601d\u8003\u548c\u590d\u6742\u63a8\u7406\uff0c\u8fbe\u5230\u4e86\u5f53\u65f6\u7684\u9886\u5148\u6c34\u5e73\u3002<\/p>\n<p>\u5728\u8fd9\u4e00\u80cc\u666f\u4e0b\uff0c\u963f\u91cc\u4e91\u63a8\u51fa\u4e86 QwQ-32B \u6a21\u578b\uff0c\u518d\u6b21\u5f15\u53d1\u4e1a\u754c\u5173\u6ce8\u3002\u8fd9\u6b3e\u62e5\u6709 320 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\uff0c\u5728\u6027\u80fd\u4e0a\u5ab2\u7f8e <a href=\"https:\/\/www.kdjingpai.com\/deepseek-r1nenglixiang\/\">DeepSeek-R1<\/a> \u6a21\u578b\uff0c\u800c\u540e\u8005\u53c2\u6570\u91cf\u9ad8\u8fbe 6710 \u4ebf\uff08\u6fc0\u6d3b\u53c2\u6570\u91cf\u4e3a 370 \u4ebf\uff09\u3002 QwQ-32B \u7684\u5353\u8d8a\u8868\u73b0\uff0c\u6709\u529b\u5730\u8bc1\u660e\u4e86\u5f3a\u5316\u5b66\u4e60\u5728\u63d0\u5347\u57fa\u4e8e\u6d77\u91cf\u4e16\u754c\u77e5\u8bc6\u9884\u8bad\u7ec3\u7684\u5f3a\u5927\u57fa\u7840\u6a21\u578b\u667a\u80fd\u6c34\u5e73\u65b9\u9762\u7684\u6709\u6548\u6027\u3002<\/p>\n<p>\u66f4\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u963f\u91cc\u4e91\u8fd8\u5c06 Agent (\u667a\u80fd\u4ee3\u7406) \u76f8\u5173\u80fd\u529b\u878d\u5165\u5230 QwQ-32B \u7684\u63a8\u7406\u6a21\u578b\u4e2d\uff0c\u4f7f\u5176\u4e0d\u4ec5\u80fd\u591f\u8fdb\u884c\u6279\u5224\u6027\u601d\u8003\uff0c\u8fd8\u80fd\u5229\u7528\u5de5\u5177\uff0c\u5e76\u6839\u636e\u73af\u5883\u53cd\u9988\u8c03\u6574\u63a8\u7406\u8fc7\u7a0b\u3002\u8fd9\u4e9b\u6280\u672f\u8fdb\u6b65\uff0c\u5145\u5206\u5c55\u793a\u4e86 RL \u6280\u672f\u7684\u53d8\u9769\u6f5c\u529b\uff0c\u4e5f\u4e3a\u901a\u5411\u901a\u7528\u4eba\u5de5\u667a\u80fd (AGI) \u7684\u9053\u8def\uff0c \u94fa\u5e73\u4e86\u9053\u8def\u3002<\/p>\n<p>\u76ee\u524d\uff0cQwQ-32B \u5df2\u5728 Hugging Face \u548c ModelScope \u5e73\u53f0\u4ee5 Apache 2.0 \u5f00\u6e90\u534f\u8bae\u53d1\u5e03\uff0c\u7528\u6237\u53ef\u4ee5\u901a\u8fc7 <a href=\"https:\/\/www.kdjingpai.com\/qwen-chat\/\">Qwen Chat<\/a> \u4f53\u9a8c\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u5c0f\u6a21\u578b\uff0c\u5927\u80fd\u91cf\uff1aQwQ-32B \u4ee5 320 \u4ebf\u53c2\u6570\u6bd4\u80a9 6710 \u4ebf\u53c2\u6570 DeepSeek-R1-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/17a16c00baf290f.png\" alt=\"\u5c0f\u6a21\u578b\uff0c\u5927\u80fd\u91cf\uff1aQwQ-32B \u4ee5 320 \u4ebf\u53c2\u6570\u6bd4\u80a9 6710 \u4ebf\u53c2\u6570 DeepSeek-R1-1\" \/><\/p>\n<h2>\u6a21\u578b\u4ecb\u7ecd (Introduction)<\/h2>\n<p>QwQ \u662f \u201c\u5343\u95ee\u201d (Qwen) \u7cfb\u5217\u6a21\u578b\u7684\u63a8\u7406\u6a21\u578b\uff0c\u4e0e\u4f20\u7edf\u7684\u6307\u4ee4\u5fae\u8c03\u6a21\u578b\u76f8\u6bd4\uff0cQwQ \u6a21\u578b\u5177\u5907\u66f4\u5f3a\u7684\u601d\u8003\u548c\u63a8\u7406\u80fd\u529b\uff0c\u5728\u4e0b\u6e38\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff0c\u5c24\u5176\u662f\u5728\u89e3\u51b3\u590d\u6742\u96be\u9898\u65f6\u3002QwQ-32B \u4f5c\u4e3a\u4e2d\u7b49\u89c4\u6a21\u7684\u63a8\u7406\u6a21\u578b\uff0c\u5176\u6027\u80fd\u8db3\u4ee5\u4e0e DeepSeek-R1\u3001o1-mini \u7b49\u5148\u8fdb\u7684\u63a8\u7406\u6a21\u578b\u76f8\u5ab2\u7f8e\u3002<\/p>\n<p><strong>\u6a21\u578b\u7279\u70b9:<\/strong><\/p>\n<ul>\n<li><strong>\u7c7b\u578b<\/strong>: \u81ea\u56de\u5f52\u8bed\u8a00\u6a21\u578b (Causal Language Models)<\/li>\n<li><strong>\u8bad\u7ec3\u9636\u6bb5<\/strong>: \u9884\u8bad\u7ec3 (Pretraining) &amp; \u540e\u8bad\u7ec3 (Post-training\uff0c\u5305\u62ec\u76d1\u7763\u5fae\u8c03 Supervised Finetuning \u548c\u5f3a\u5316\u5b66\u4e60 Reinforcement Learning)<\/li>\n<li><strong>\u67b6\u6784<\/strong>: Transformers \u7ed3\u6784\uff0c\u91c7\u7528 RoPE \u4f4d\u7f6e\u7f16\u7801\uff0cSwiGLU \u6fc0\u6d3b\u51fd\u6570\uff0cRMSNorm \u5f52\u4e00\u5316\uff0c\u4ee5\u53ca Attention QKV bias \u6ce8\u610f\u529b\u673a\u5236\u504f\u7f6e<\/li>\n<li><strong>\u53c2\u6570\u89c4\u6a21<\/strong>: 325 \u4ebf\uff0832.5B\uff09<\/li>\n<li><strong>\u975e\u5d4c\u5165\u5c42\u53c2\u6570\u89c4\u6a21<\/strong>: 310 \u4ebf\uff0831.0B\uff09<\/li>\n<li><strong>\u5c42\u6570<\/strong>: 64<\/li>\n<li><strong>\u6ce8\u610f\u529b\u5934\u6570 (GQA)<\/strong>: Query \u7aef 40 \u4e2a\uff0cKey\/Value \u7aef 8 \u4e2a<\/li>\n<li><strong>\u4e0a\u4e0b\u6587\u957f\u5ea6<\/strong>: \u5b8c\u6574 131,072 <a href=\"https:\/\/www.kdjingpai.com\/tokenization\/\">tokens<\/a><\/li>\n<\/ul>\n<p><strong>\u6ce8\u610f<\/strong>: \u4e3a\u4e86\u83b7\u5f97\u6700\u4f73\u4f7f\u7528\u4f53\u9a8c\uff0c\u8bf7\u52a1\u5fc5\u53c2\u8003\u00a0<a href=\"https:\/\/huggingface.co\/Qwen\/QwQ-32B#usage-guidelines\">\u4f7f\u7528\u6307\u5357<\/a>\u00a0\u540e\u518d\u90e8\u7f72 QwQ \u6a21\u578b\u3002<\/p>\n<p>\u7528\u6237\u53ef\u4ee5\u901a\u8fc7\u00a0<a href=\"https:\/\/huggingface.co\/spaces\/Qwen\/QwQ-32B-Demo\">Demo<\/a>\u00a0\u8fdb\u884c\u4f53\u9a8c\uff0c\u6216\u901a\u8fc7\u00a0<a href=\"https:\/\/www.kdjingpai.com\/qwen-chat\/\">QwenChat<\/a>\u00a0\u8bbf\u95ee QwQ \u6a21\u578b\uff0c\u8bb0\u5f97\u6253\u5f00Thinking (QwQ)\u3002<\/p>\n<p>\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u8003 <a href=\"https:\/\/github.com\/QwenLM\/Qwen2.5\">GitHub \u4ed3\u5e93<\/a>\u00a0\u4ee5\u53ca\u00a0<a href=\"https:\/\/qwen.readthedocs.io\/en\/latest\/\">\u5b98\u65b9\u6587\u6863<\/a>\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u6027\u80fd\u8868\u73b0 (Performance)<\/h2>\n<p>QwQ-32B \u6a21\u578b\u5728\u4e00\u7cfb\u5217\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u65e8\u5728\u5168\u9762\u8003\u5bdf\u5176\u5728\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u7f16\u5199\u4ee5\u53ca\u901a\u7528\u95ee\u9898\u89e3\u51b3\u7b49\u65b9\u9762\u7684\u80fd\u529b\u3002\u4e0b\u56fe\u5c55\u793a\u4e86 QwQ-32B \u4e0e\u5305\u62ec DeepSeek-R1-Distilled-Qwen-32B\u3001DeepSeek-R1-Distilled-Llama-70B\u3001o1-mini \u4ee5\u53ca\u539f\u59cb DeepSeek-R1 \u5728\u5185\u7684\u5176\u4ed6\u9886\u5148\u6a21\u578b\u7684\u6027\u80fd\u5bf9\u6bd4\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u5c0f\u6a21\u578b\uff0c\u5927\u80fd\u91cf\uff1aQwQ-32B \u4ee5 320 \u4ebf\u53c2\u6570\u6bd4\u80a9 6710 \u4ebf\u53c2\u6570 DeepSeek-R1-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/09c8be8760d342b.jpg\" alt=\"\u5c0f\u6a21\u578b\uff0c\u5927\u80fd\u91cf\uff1aQwQ-32B \u4ee5 320 \u4ebf\u53c2\u6570\u6bd4\u80a9 6710 \u4ebf\u53c2\u6570 DeepSeek-R1-1\" \/><\/p>\n<p>\u4ece\u7ed3\u679c\u6765\u770b\uff0cQwQ-32B \u5728\u591a\u4e2a\u5173\u952e benchmark \u4e0a\u90fd\u5c55\u73b0\u51fa\u4e86\u4e0e\u9876\u7ea7\u6a21\u578b\u76f8\u8fd1\u751a\u81f3\u66f4\u4f18\u7684\u6027\u80fd\u3002\u5c24\u5176\u503c\u5f97\u5173\u6ce8\u7684\u662f\uff0c\u5728\u4e0e\u53c2\u6570\u91cf\u8fdc\u8d85\u81ea\u8eab\u7684 DeepSeek-R1 \u5bf9\u6bd4\u4e2d\uff0cQwQ-32B \u4f9d\u7136\u80fd\u591f\u4fdd\u6301\u7ade\u4e89\u529b\uff0c\u8fd9\u8fdb\u4e00\u6b65\u5370\u8bc1\u4e86\u5f3a\u5316\u5b66\u4e60\u5728\u63d0\u5347\u6a21\u578b\u6548\u80fd\u65b9\u9762\u7684\u5de8\u5927\u6f5c\u529b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u5f3a\u5316\u5b66\u4e60 (Reinforcement Learning)<\/h2>\n<p>QwQ-32B \u7684\u5353\u8d8a\u6027\u80fd\uff0c\u5f88\u5927\u7a0b\u5ea6\u4e0a\u5f52\u529f\u4e8e\u5176\u80cc\u540e\u91c7\u7528\u7684\u5f3a\u5316\u5b66\u4e60 (RL) \u6280\u672f\u3002 \u7b80\u5355\u6765\u8bf4\uff0c\u5f3a\u5316\u5b66\u4e60\u662f\u4e00\u79cd\u901a\u8fc7\u5956\u52b1\u6216\u60e9\u7f5a\u673a\u5236\uff0c\u5f15\u5bfc\u6a21\u578b\u5b66\u4e60\u5728\u7279\u5b9a\u73af\u5883\u4e2d\u505a\u51fa\u6700\u4f18\u51b3\u7b56\u7684\u65b9\u6cd5\u3002 \u4e0e\u4f20\u7edf\u7684\u76d1\u7763\u5b66\u4e60\u4e0d\u540c\uff0c\u5f3a\u5316\u5b66\u4e60\u4e0d\u4f9d\u8d56\u4e8e\u5927\u91cf\u7684\u6807\u6ce8\u6570\u636e\uff0c\u800c\u662f\u901a\u8fc7\u4e0e\u73af\u5883\u7684\u4ea4\u4e92\uff0c\u4e0d\u65ad\u8bd5\u9519\u548c\u5b66\u4e60\uff0c\u6700\u7ec8\u638c\u63e1\u5b8c\u6210\u4efb\u52a1\u6240\u9700\u7684\u7b56\u7565\u3002<\/p>\n<p>\u5728 QwQ-32B \u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u963f\u91cc\u4e91\u7684\u7814\u7a76\u56e2\u961f\u4ece\u4e00\u4e2a\u51b7\u542f\u52a8\u7684 Checkpoint \u5f00\u59cb\uff0c\u5b9e\u65bd\u4e86\u57fa\u4e8e\u7ed3\u679c\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\u6269\u5c55\u65b9\u6cd5\u3002\u5728\u521d\u59cb\u9636\u6bb5\uff0c\u4ed6\u4eec\u4e3b\u8981\u9488\u5bf9\u6570\u5b66\u548c\u4ee3\u7801\u4efb\u52a1\u8fdb\u884c RL \u6269\u5c55\u3002\u4e0e\u4f9d\u8d56\u4f20\u7edf\u5956\u52b1\u6a21\u578b\u4e0d\u540c\uff0c\u7814\u7a76\u56e2\u961f\u91c7\u7528\u4e86\u9488\u5bf9\u6570\u5b66\u95ee\u9898\u7684\u51c6\u786e\u6027\u9a8c\u8bc1\u5668 (Accuracy Verifier)\uff0c\u4ee5\u786e\u4fdd\u6700\u7ec8\u7b54\u6848\u7684\u6b63\u786e\u6027\uff1b\u5e76\u4f7f\u7528\u4ee3\u7801\u6267\u884c\u670d\u52a1\u5668 (Code Execution Server) \u6765\u8bc4\u4f30\u751f\u6210\u7684\u4ee3\u7801\u662f\u5426\u6210\u529f\u901a\u8fc7\u9884\u5b9a\u4e49\u7684\u6d4b\u8bd5\u7528\u4f8b\u3002<\/p>\n<p>\u968f\u7740\u8bad\u7ec3\u7684\u6df1\u5165\uff0c\u6a21\u578b\u5728\u6570\u5b66\u548c\u4ee3\u7801\u9886\u57df\u7684\u6027\u80fd\u90fd\u5448\u73b0\u51fa\u6301\u7eed\u63d0\u5347\u7684\u8d8b\u52bf\u3002\u5728\u7b2c\u4e00\u9636\u6bb5\u4e4b\u540e\uff0c\u7814\u7a76\u56e2\u961f\u53c8\u589e\u52a0\u4e86\u9488\u5bf9\u901a\u7528\u80fd\u529b\u7684 RL \u8bad\u7ec3\u9636\u6bb5\u3002 \u8fd9\u4e00\u9636\u6bb5\u7684\u8bad\u7ec3\uff0c\u91c7\u7528\u4e86\u6765\u81ea\u901a\u7528\u5956\u52b1\u6a21\u578b\u548c\u4e00\u4e9b\u57fa\u4e8e\u89c4\u5219\u7684\u9a8c\u8bc1\u5668\u7684\u5956\u52b1\u4fe1\u53f7\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u901a\u8fc7\u5c11\u91cf\u6b65\u9aa4\u7684 RL \u8bad\u7ec3\uff0c\u5373\u53ef\u6709\u6548\u63d0\u5347\u6a21\u578b\u5728\u6307\u4ee4\u9075\u5faa\u3001\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u4ee5\u53ca Agent \u6027\u80fd\u7b49\u65b9\u9762\u7684\u901a\u7528\u80fd\u529b\uff0c\u4e14\u4e0d\u4f1a\u5bf9\u6570\u5b66\u548c\u4ee3\u7801\u80fd\u529b\u9020\u6210\u660e\u663e\u7684\u6027\u80fd\u4e0b\u964d\u3002<\/p>\n<p>\u8fd9\u91cc\u6709\u4e00\u7bc7\u5173\u4e8eQwen-2.5-3B\u4e3a\u4ec0\u4e48\u5177\u6709\u4f18\u79c0\u63a8\u7406\u80fd\u529b\u7684\u6587\u7ae0\uff1a<a href=\"https:\/\/www.kdjingpai.com\/damoxingruhebianbian\/\">\u5927\u6a21\u578b\u5982\u4f55\u53d8\u5f97\u66f4\u201c\u806a\u660e\u201d\uff1f\u65af\u5766\u798f\u5927\u5b66\u63ed\u79d8\u81ea\u6211\u6539\u8fdb\u7684\u5173\u952e\uff1a\u56db\u79cd\u8ba4\u77e5\u884c\u4e3a<\/a><\/p>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u6307\u5357 (Usage Guidelines)<\/h2>\n<p>\u4e3a\u4e86\u83b7\u5f97\u6700\u4f73\u6027\u80fd\uff0c\u5efa\u8bae\u91c7\u7528\u4ee5\u4e0b\u8bbe\u7f6e\uff1a<\/p>\n<ol>\n<li><strong>\u5f3a\u5236\u6a21\u578b\u8fdb\u884c\u601d\u8003\u8f93\u51fa<\/strong>: \u786e\u4fdd\u6a21\u578b\u4ee5\u00a0<code>&lt;think&gt;\\n<\/code>\u00a0\u5f00\u5934\uff0c\u4ee5\u907f\u514d\u751f\u6210\u7a7a\u7684\u601d\u8003\u5185\u5bb9\uff0c\u8fd9\u53ef\u80fd\u4f1a\u964d\u4f4e\u8f93\u51fa\u8d28\u91cf\u3002\u5982\u679c\u4f7f\u7528\u00a0<code>apply_chat_template<\/code>\u00a0\u5e76\u8bbe\u7f6e\u00a0<code>add_generation_prompt=True<\/code>\uff0c\u5219\u6b64\u9879\u5df2\u81ea\u52a8\u5b9e\u73b0\u3002\u4f46\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u54cd\u5e94\u5728\u5f00\u5934\u7f3a\u5c11\u00a0<code>&lt;think&gt;<\/code>\u00a0\u6807\u7b7e\uff0c\u5c5e\u4e8e\u6b63\u5e38\u73b0\u8c61\u3002<\/li>\n<li><strong>\u91c7\u6837\u53c2\u6570<\/strong>:\n<ul>\n<li>\u4f7f\u7528\u00a0<code>Temperature=0.6<\/code>\u00a0\u548c\u00a0<code>TopP=0.95<\/code>\u00a0\u4ee3\u66ff\u8d2a\u5a6a\u89e3\u7801 (Greedy decoding)\uff0c\u4ee5\u907f\u514d\u65e0\u4f11\u6b62\u7684\u91cd\u590d\u3002<\/li>\n<li>\u4f7f\u7528\u00a0<code>TopK<\/code>\u00a0\u5728 20 \u5230 40 \u4e4b\u95f4\uff0c\u4ee5\u8fc7\u6ee4\u6389\u7f55\u89c1\u7684 <a href=\"https:\/\/www.kdjingpai.com\/tokenization\/\">Token<\/a> \u51fa\u73b0\uff0c\u540c\u65f6\u4fdd\u6301\u751f\u6210\u8f93\u51fa\u7684\u591a\u6837\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6807\u51c6\u5316\u8f93\u51fa\u683c\u5f0f<\/strong>: \u5728\u8fdb\u884c\u57fa\u51c6\u6d4b\u8bd5\u65f6\uff0c\u5efa\u8bae\u4f7f\u7528 Prompt \u63d0\u793a\u6765\u6807\u51c6\u5316\u6a21\u578b\u8f93\u51fa\u3002\n<ul>\n<li><strong>\u6570\u5b66\u95ee\u9898<\/strong>: \u5728 Prompt \u4e2d\u52a0\u5165 \u201cPlease reason step by step, and put your final answer within \\boxed{}. \u201d (\u8bf7\u9010\u6b65\u63a8\u7406\uff0c\u5e76\u5c06\u6700\u7ec8\u7b54\u6848\u653e\u5728 \\boxed{} \u4e2d)\u3002<\/li>\n<li><strong>\u591a\u9879\u9009\u62e9\u9898<\/strong>: \u5728 Prompt \u4e2d\u6dfb\u52a0\u4ee5\u4e0b JSON \u7ed3\u6784\uff0c\u4ee5\u6807\u51c6\u5316\u54cd\u5e94: \u201cPlease show your choice in the\u00a0<code>answer<\/code>\u00a0field with only the choice letter, e.g.,<code>\\\"answer\\\": \\\"C\\\"<\/code>. \u201d (\u8bf7\u5728\u00a0<code>answer<\/code>\u00a0\u5b57\u6bb5\u4e2d\u663e\u793a\u60a8\u7684\u9009\u62e9\uff0c\u4ec5\u5305\u542b\u9009\u62e9\u5b57\u6bcd\uff0c\u4f8b\u5982\uff0c<code>\\\"answer\\\": \\\"C\\\"<\/code>)\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5904\u7406\u957f\u8f93\u5165<\/strong>: \u5bf9\u4e8e\u8d85\u8fc7 32,768 \u4e2a Tokens \u7684\u8f93\u5165\uff0c\u542f\u7528\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2309.00071\">YaRN<\/a>\u00a0\u6280\u672f\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u6709\u6548\u6355\u83b7\u957f\u5e8f\u5217\u4fe1\u606f\u7684\u80fd\u529b\u3002<\/li>\n<\/ol>\n<p>\u5bf9\u4e8e\u652f\u6301\u7684\u6846\u67b6\uff0c\u53ef\u4ee5\u5c06\u4ee5\u4e0b\u5185\u5bb9\u6dfb\u52a0\u5230\u00a0<code>config.json<\/code>\u00a0\u6587\u4ef6\u4e2d\u4ee5\u542f\u7528 YaRN\uff1a<\/p>\n<pre><code>{\r\n...,\r\n\"rope_scaling\": {\r\n\"factor\": 4.0,\r\n\"original_max_position_embeddings\": 32768,\r\n\"type\": \"yarn\"\r\n}\r\n}\r\n<\/code><\/pre>\n<p>\u5728\u90e8\u7f72\u65b9\u9762\uff0c\u5efa\u8bae\u4f7f\u7528 vLLM\u3002\u5982\u679c\u4e0d\u719f\u6089 vLLM\uff0c\u8bf7\u53c2\u8003\u00a0<a href=\"https:\/\/qwen.readthedocs.io\/en\/latest\/deployment\/vllm.html\">\u5b98\u65b9\u6587\u6863<\/a>\u00a0\u4ee5\u83b7\u53d6\u4f7f\u7528\u65b9\u6cd5\u3002\u76ee\u524d\uff0cvLLM \u4ec5\u652f\u6301\u9759\u6001 YARN\uff0c\u8fd9\u610f\u5473\u7740\u7f29\u653e\u56e0\u5b50\u5728\u8f93\u5165\u957f\u5ea6\u53d8\u5316\u65f6\u4fdd\u6301\u6052\u5b9a\uff0c<strong>\u8fd9\u53ef\u80fd\u4f1a\u5f71\u54cd\u6a21\u578b\u5728\u5904\u7406\u8f83\u77ed\u6587\u672c\u65f6\u7684\u6027\u80fd<\/strong>\u3002\u56e0\u6b64\uff0c\u5efa\u8bae\u4ec5\u5728\u9700\u8981\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u65f6\u6dfb\u52a0\u00a0<code>rope_scaling<\/code>\u00a0\u914d\u7f6e\u3002<\/p>\n<h3>\u5982\u4f55\u4f7f\u7528 QwQ-32B (Use QwQ-32B)<\/h3>\n<p>\u4ee5\u4e0b\u7b80\u8981\u793a\u4f8b\uff0c\u5c55\u793a\u4e86\u5982\u4f55\u901a\u8fc7 Hugging Face Transformers \u548c\u963f\u91cc\u4e91 DashScope API \u4f7f\u7528 QwQ-32B \u6a21\u578b\u3002<\/p>\n<p><strong>\u901a\u8fc7 Hugging Face Transformers:<\/strong><\/p>\n<pre><code>from transformers import AutoModelForCausalLM, AutoTokenizer\r\nmodel_name = \"Qwen\/QwQ-32B\"\r\nmodel = AutoModelForCausalLM.from_pretrained(\r\nmodel_name,\r\ntorch_dtype=\"auto\",\r\ndevice_map=\"auto\"\r\n)\r\ntokenizer = AutoTokenizer.from_pretrained(model_name)\r\nprompt = \"How many r's are in the word \\\"strawberry\\\"\"\r\nmessages = [\r\n{\"role\": \"user\", \"content\": prompt}\r\n]\r\ntext = tokenizer.apply_chat_template(\r\nmessages,\r\ntokenize=False,\r\nadd_generation_prompt=True\r\n)\r\nmodel_inputs = tokenizer([text], return_tensors=\"pt\").to(model.device)\r\ngenerated_ids = model.generate(\r\n**model_inputs,\r\nmax_new_tokens=32768\r\n)\r\ngenerated_ids = [\r\noutput_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)\r\n]\r\nresponse = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]\r\nprint(response)\r\n<\/code><\/pre>\n<p><strong>\u901a\u8fc7\u963f\u91cc\u4e91 DashScope API:<\/strong><\/p>\n<pre><code>from openai import OpenAI\r\nimport os\r\n# Initialize OpenAI client\r\nclient = OpenAI(\r\n# If the environment variable is not configured, replace with your API Key: api_key=\"sk-xxx\"\r\n# How to get an API Key\uff1ahttps:\/\/help.aliyun.com\/zh\/model-studio\/developer-reference\/get-api-key\r\napi_key=os.getenv(\"DASHSCOPE_API_KEY\"),\r\nbase_url=\"https:\/\/dashscope.aliyuncs.com\/compatible-mode\/v1\"\r\n)\r\nreasoning_content = \"\"\r\ncontent = \"\"\r\nis_answering = False\r\ncompletion = client.chat.completions.create(\r\nmodel=\"qwq-32b\",\r\nmessages=[\r\n{\"role\": \"user\", \"content\": \"Which is larger, 9.9 or 9.11?\"}\r\n],\r\nstream=True,\r\n# Uncomment the following line to return token usage in the last chunk\r\n# stream_options={\r\n#     \"include_usage\": True\r\n# }\r\n)\r\nprint(\"\\n\" + \"=\" * 20 + \"reasoning content\" + \"=\" * 20 + \"\\n\")\r\nfor chunk in completion:\r\n# If chunk.choices is empty, print usage\r\nif not chunk.choices:\r\nprint(\"\\nUsage:\")\r\nprint(chunk.usage)\r\nelse:\r\ndelta = chunk.choices[0].delta\r\n# Print reasoning content\r\nif hasattr(delta, 'reasoning_content') and delta.reasoning_content is not None:\r\nprint(delta.reasoning_content, end='', flush=True)\r\nreasoning_content += delta.reasoning_content\r\nelse:\r\nif delta.content != \"\" and is_answering is False:\r\nprint(\"\\n\" + \"=\" * 20 + \"content\" + \"=\" * 20 + \"\\n\")\r\nis_answering = True\r\n# Print content\r\nprint(delta.content, end='', flush=True)\r\ncontent += delta.content\r\n<\/code><\/pre>\n<p>&nbsp;<\/p>\n<h2>\u672a\u6765\u5c55\u671b (Future Work)<\/h2>\n<p>QwQ-32B \u7684\u53d1\u5e03\uff0c\u6807\u5fd7\u7740 \u201c\u5343\u95ee\u201d (Qwen) \u7cfb\u5217\u6a21\u578b\u5728\u6269\u5c55\u5f3a\u5316\u5b66\u4e60 (RL) \u4ee5\u589e\u5f3a\u63a8\u7406\u80fd\u529b\u65b9\u9762\u8fc8\u51fa\u4e86\u521d\u6b65\u4f46\u5173\u952e\u7684\u4e00\u6b65\u3002 \u901a\u8fc7\u8fd9\u4e00\u63a2\u7d22\uff0c\u963f\u91cc\u4e91\u4e0d\u4ec5\u89c1\u8bc1\u4e86\u5f3a\u5316\u5b66\u4e60\u6269\u5c55\u5e94\u7528\u7684\u5de8\u5927\u6f5c\u529b\uff0c\u4e5f\u8ba4\u8bc6\u5230\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5185\u90e8\u8574\u85cf\u7740\u5c1a\u672a\u88ab\u5145\u5206\u6316\u6398\u7684\u5de8\u5927\u6f5c\u529b\u3002<\/p>\n<p>\u5c55\u671b \u201c\u5343\u95ee\u201d \u4e0b\u4e00\u4ee3\u6a21\u578b\u7684\u7814\u53d1\uff0c\u963f\u91cc\u4e91\u5145\u6ee1\u4fe1\u5fc3\u3002\u4ed6\u4eec\u76f8\u4fe1\uff0c\u901a\u8fc7\u7ed3\u5408\u66f4\u5f3a\u5927\u7684\u57fa\u7840\u6a21\u578b\uff0c\u4ee5\u53ca\u7531\u89c4\u6a21\u5316\u8ba1\u7b97\u8d44\u6e90\u9a71\u52a8\u7684\u5f3a\u5316\u5b66\u4e60\u6280\u672f\uff0c\u5c06\u6709\u671b\u52a0\u901f\u5b9e\u73b0\u901a\u7528\u4eba\u5de5\u667a\u80fd (AGI) \u7684\u6700\u7ec8\u76ee\u6807\u3002\u6b64\u5916\uff0c\u963f\u91cc\u4e91\u8fd8\u5728\u79ef\u6781\u63a2\u7d22\u5c06 Agent \u4e0e RL \u8fdb\u884c\u6df1\u5ea6\u878d\u5408\uff0c\u4ee5\u5b9e\u73b0\u66f4\u957f\u7a0b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5e76\u81f4\u529b\u4e8e\u5728\u63a8\u7406\u65f6\u901a\u8fc7\u89c4\u6a21\u5316\u6269\u5c55\u6765\u91ca\u653e\u66f4\u5927\u7684\u667a\u80fd\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8fd1\u671f\uff0cAI \u9886\u57df\u6d8c\u73b0\u51fa\u4ee4\u4eba\u77a9\u76ee\u7684\u8fdb\u5c55\uff0c\u5c24\u5176\u662f\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u63a8\u7406\u80fd\u529b\u63d0\u5347\u65b9\u9762\u3002\u5176\u4e2d\uff0c\u5f3a\u5316\u5b66\u4e60 (Reinforcement Learning, RL) \u6b63\u9010\u6e10\u6210\u4e3a\u7a81\u7834\u4f20\u7edf\u6a21\u578b\u6027\u80fd\u74f6\u9888\u7684\u5173\u952e\u6280\u672f\u3002\u4e0d\u5c11\u7814\u7a76\u5df2\u8bc1\u5b9e\uff0cRL \u80fd\u591f\u663e\u8457\u589e&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[46],"tags":[],"class_list":["post-27775","post","type-post","status-publish","format-standard","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/27775","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=27775"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/27775\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=27775"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=27775"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=27775"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}