{"id":53793,"date":"2025-02-16T09:59:15","date_gmt":"2025-02-16T01:59:15","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53793\/"},"modified":"2025-02-16T09:59:15","modified_gmt":"2025-02-16T01:59:15","slug":"%e6%8e%a8%e7%90%86%e6%a8%a1%e5%9e%8b%e7%9a%84%e8%ae%ad%e7%bb%83%ef%bc%9a%e4%bb%8e%e5%8e%9f%e7%90%86%e5%88%b0%e5%ae%9e%e8%b7%b5","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53793\/","title":{"rendered":"\u63a8\u7406\u6a21\u578b\u7684\u8bad\u7ec3\uff1a\u4ece\u539f\u7406\u5230\u5b9e\u8df5"},"content":{"rendered":"<p>\u63a8\u7406\u6a21\u578b\u5e76\u4e0d\u5b8c\u5168\u662f\u65b0\u4e8b\u7269\uff0c\u4f46 DeepSeek \u7684 R1 \u6a21\u578b\u7684\u53d1\u5e03\u5728\u5f00\u6e90\u793e\u533a\u4e2d\u5f15\u8d77\u4e86\u6781\u5927\u7684\u5174\u594b\u3002\u5728\u9605\u8bfb DeepSeek-R1 \u8bba\u6587\u540e\uff0c\u6211\u5bf9\u4ed6\u4eec\u7684\u540e\u8bad\u7ec3\u6280\u672f\u7279\u522b\u611f\u5174\u8da3\uff0c\u636e\u62a5\u9053\uff0c\u8be5\u6280\u672f\u53ef\u4ee5\u663e\u7740\u63d0\u9ad8\u6027\u80fd\u3002<\/p>\n<p>\u4f20\u7edf\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u662f\u76d1\u7763\u5fae\u8c03 (SFT)\u3002\u5728 SFT \u4e2d\uff0c\u4e3a\u6a21\u578b\u63d0\u4f9b\u63d0\u793a\u548c\u671f\u671b\u7684\u8f93\u51fa\uff0c\u4ee5\u4e86\u89e3\u7528\u6237\u504f\u597d\u3001\u7ed3\u6784\u5bf9\u9f50\u3001\u5b89\u5168\u534f\u8bae\u7b49\u3002\u672c\u8d28\u4e0a\uff0cSFT \u5c55\u793a\u4e86\u5f00\u53d1\u4eba\u5458\u6216\u7528\u6237\u5bf9\u6a21\u578b\u7684\u671f\u671b\uff0c\u5e76\u5f15\u5bfc\u5b83\u4e0e\u8fd9\u4e9b\u793a\u4f8b\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n<p>\u65e2\u7136\u5df2\u7ecf\u63d0\u5230\u4e86\u540e\u8bad\u7ec3\uff0c\u90a3\u4e48\u5728\u6df1\u5165\u7814\u7a76 R1 \u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u4e4b\u524d\uff0c\u503c\u5f97\u7b80\u8981\u8ba8\u8bba\u4e00\u4e0b\u9884\u8bad\u7ec3\uff0c\u56e0\u4e3a\u4e24\u8005\u662f\u76f8\u5173\u7684\u3002\u8bad\u7ec3\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u521d\u59cb\u6b65\u9aa4\u4e4b\u4e00\u662f\u56e0\u679c\u8bed\u8a00\u5efa\u6a21\u3002\u5728\u6b64\u9636\u6bb5\uff0c\u6a21\u578b\u901a\u8fc7\u5173\u6ce8\u524d\u9762\u7684\u6807\u8bb0\u6765\u5b66\u4e60\u9884\u6d4b\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u4e2a\u6807\u8bb0\u3002\u6b64\u8fc7\u7a0b\u4ea7\u751f\u201c\u57fa\u7840\u201d\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u4ee5\u534a\u76d1\u7763\u7684\u65b9\u5f0f\u5728\u5927\u91cf\u516c\u5171\u6587\u672c\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6ca1\u6709\u660e\u786e\u7684\u6807\u7b7e\uff0c\u4ee5\u9884\u6d4b\u540e\u7eed\u6807\u8bb0\u3002\u8bf4\u5230\u516c\u5171\u6587\u672c\u6570\u636e\uff0c\u6211\u4eec\u5373\u5c06\u8017\u5c3d\u5b83\uff08\u5728\u67d0\u79cd\u7a0b\u5ea6\u4e0a\uff0c\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u5ffd\u7565\u5176\u4ed6\u6a21\u5f0f\uff09\u3002\u8fd9\u79cd\u7a00\u7f3a\u6027\u662f\u7814\u7a76\u4eba\u5458\u4e13\u6ce8\u4e8e LLM \u540e\u8bad\u7ec3\u9636\u6bb5\u7684\u4e3b\u8981\u9a71\u52a8\u529b\u4e4b\u4e00\u3002<\/p>\n<p>\u56de\u5230\u540e\u8bad\u7ec3\uff0c\u5b58\u5728\u8bb8\u591a\u6280\u672f\uff0c\u4f8b\u5982 SFT\u3001PPO \u548c DPO \u7b49\u3002\u867d\u7136\u67d0\u4e9b\u65b9\u6cd5\u5728\u7279\u5b9a\u9886\u57df\u8868\u73b0\u66f4\u597d\uff0c\u4f46\u5b83\u4eec\u901a\u5e38\u4f1a\u9650\u5236\u6307\u4ee4\u5bf9\u9f50\u7684\u521b\u4f5c\u81ea\u7531\uff0c\u6216\u8005\u5728\u957f\u65f6\u95f4\u8fd0\u884c\u4e2d\u53d8\u5f97\u8fc7\u4e8e\u6602\u8d35\u3002\u76f8\u53cd\uff0c\u57fa\u7840\u6a21\u578b\u6ca1\u6709\u8fd9\u4e9b\u9650\u5236\uff0c\u4f46\u4e0d\u53ef\u9884\u6d4b\u3001\u96be\u4ee5\u6307\u5bfc\uff0c\u5e76\u4e14\u7ecf\u5e38\u4ea7\u751f\u4eba\u7c7b\u96be\u4ee5\u7406\u89e3\u7684\u8f93\u51fa\u3002<\/p>\n<p>R1 \u7684\u4f5c\u8005\u53d1\u73b0\u4e86\u4e00\u79cd\u4e24\u5168\u5176\u7f8e\u7684\u65b9\u6cd5\uff1a\u5c06\u539f\u59cb\u6a21\u578b\u4e0e\u6307\u4ee4\u5bf9\u9f50\uff0c\u540c\u65f6\u4fdd\u7559\u4e00\u5b9a\u7a0b\u5ea6\u7684\u521b\u4f5c\u81ea\u7531\u3002\u6211\u4e0d\u4f1a\u5728\u8fd9\u91cc\u8be6\u7ec6\u4ecb\u7ecd\u6574\u4e2a\u8fc7\u7a0b\uff0c\u4f46\u4ece\u6839\u672c\u4e0a\u8bb2\uff0c\u5b83\u4eec\u5141\u8bb8\u57fa\u7840\u6a21\u578b\u4ee5\u76f8\u5f53\u5927\u7684\u81ea\u7531\u5ea6\u201c\u63a8\u7406\u201d\uff0c\u540c\u65f6\u8bbe\u5b9a\u4e13\u6ce8\u4e8e\u51c6\u786e\u6027\u548c\u683c\u5f0f\u7684\u7b80\u5355\u76ee\u6807\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u4ed6\u4eec\u91c7\u7528\u4e86\u5f3a\u5316\u5b66\u4e60\uff0c\u5b9e\u65bd\u4e86\u4e00\u79cd\u79f0\u4e3a\u201cGRPO\u201d\uff08\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff09\u7684\u4f18\u5316\u6280\u672f\u3002 GRPO \u6d88\u9664\u4e86\u5bf9\u6279\u8bc4\u6a21\u578b\uff08\u4f20\u7edf\u4e0a\u4e0e\u7b56\u7565\u6a21\u578b\u5927\u5c0f\u76f8\u540c\u5e76\u7528\u4e8e PPO\/DPO \u7b49\u65b9\u6cd5\uff09\u7684\u9700\u6c42\uff0c\u800c\u662f\u4f7f\u7528\u805a\u5408\u7ec4\u5206\u6570\u8ba1\u7b97\u57fa\u7ebf\u3002<\/p>\n<p>\u5c06\u8fd9\u79cd\u65b9\u6cd5\u76f4\u63a5\u5e94\u7528\u4e8e\u57fa\u7840\u6a21\u578b\u4ea7\u751f\u4e86\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4ece\u800c\u4ea7\u751f\u4e86 R1 \u6a21\u578b\u7684\u7b2c\u4e00\u4e2a\u7248\u672c\uff1aDeepSeek-R1-Zero\u3002\u5c3d\u7ba1\u8be5\u6a21\u578b\u63d0\u9ad8\u4e86\u57fa\u7840\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5176\u8f93\u51fa\u901a\u5e38\u96be\u4ee5\u88ab\u4eba\u7c7b\u9605\u8bfb\u6216\u89e3\u91ca\u3002\u5982\u524d\u6240\u8ff0\uff0cSFT \u65b9\u6cd5\u6709\u52a9\u4e8e\u6a21\u578b\u9075\u5faa\u4eba\u7c7b\u7684\u504f\u597d\u548c\u53ef\u8bfb\u6027\u3002\u7136\u800c\uff0c\u6211\u4eec\u8fd8\u8ba8\u8bba\u4e86 SFT \u5982\u4f55\u627c\u6740\u521b\u4f5c\u81ea\u7531\u548c\u6cdb\u5316\uff0c\u5982\u8bba\u6587\u201cSFT \u8bb0\u5fc6\uff0cRL \u6cdb\u5316\uff1a\u57fa\u7840\u6a21\u578b\u540e\u8bad\u7ec3\u7684\u6bd4\u8f83\u7814\u7a76\u201d\u4e2d\u6240\u8ff0\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u4f5c\u8005\u5728\u5f3a\u5316\u5b66\u4e60\u6b65\u9aa4\u4e4b\u524d\u4f7f\u7528\u4e86\u4e00\u4e2a\u5c0f\u578b\u3001\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u96c6\u6765\u201c\u51b7\u542f\u52a8\u201d\u4ed6\u4eec\u7684\u6a21\u578b\u3002\u8fd9\u79cd\u65b9\u6cd5\u5728\u8457\u540d\u7684 DeepSeek-R1 \u6a21\u578b\u7684\u521b\u5efa\u4e2d\u53d1\u6325\u4e86\u91cd\u8981\u4f5c\u7528\u3002\u5728\u4e0b\u4e00\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8\u5176\u4e2d\u4e00\u4e9b\u6280\u672f\u7684\u5b9e\u73b0\uff0c\u4ee5\u5b9e\u73b0\u7c7b\u4f3c\u7684\u7ed3\u679c\u3002<\/p>\n<h2>1\u3001\u5f3a\u5236\u63a8\u7406<\/h2>\n<p>\u5728\u4ed6\u4eec\u7684\u62a5\u544a\u4e2d\uff0cDeepSeek \u63d0\u5230\u4e86\u4e00\u7ec4\u65b0\u7684\u84b8\u998f\u6a21\u578b\u3002\u8fd9\u4e9b\u660e\u663e\u8f83\u5c0f\u7684\u6a21\u578b\u4f7f\u7528 SFT \u7684\u84b8\u998f\u6280\u672f\u8fdb\u884c\u8bad\u7ec3\uff0c\u5229\u7528\u4e86 DeepSeek-R1 \u7684\u8f93\u51fa\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e9b\u63d0\u70bc\u6a21\u578b\u6ca1\u6709\u7ecf\u8fc7\u4efb\u4f55\u5f3a\u5316\u5b66\u4e60\u6b65\u9aa4\uff0c\u8fd9\u662f\u4e00\u4e2a\u6709\u8da3\u7684\u89c2\u70b9\uff0c\u4e5f\u662f\u6211\u4eec\u503c\u5f97\u63a2\u7d22\u7684\u9886\u57df\u3002\u56e0\u6b64\uff0c\u5982\u679c\u6211\u4eec\u4e0d\u63d0\u70bc\u6a21\u578b\uff0c\u800c\u662f\u76f4\u63a5\u5728\u539f\u59cb\u6a21\u578b\u4e0a\u5b9e\u73b0 GRPO\uff0c\u4f1a\u600e\u4e48\u6837\uff1f<\/p>\n<p>\u6211\u4eec\u53ef\u4ee5\u7ee7\u7eed\u4f7f\u7528\u57fa\u7840\u6a21\u578b\u6216\u6307\u5bfc\u6a21\u578b\uff0c\u5047\u8bbe\u6307\u5bfc\u6a21\u578b\u5df2\u7ecf\u201c\u51b7\u542f\u52a8\u201d\u3002\u9274\u4e8e\u8fd9\u662f\u5728\u6211\u914d\u5907 RTX 3090 \u7684\u672c\u5730\u6e38\u620f\u88c5\u5907\u4e0a\u8fdb\u884c\u7684\u5b9e\u9a8c\uff0c\u6211\u9009\u62e9\u4e86 Qwen2.5\u20130.5B-Instruct \u6a21\u578b\u4f5c\u4e3a\u5f00\u59cb\u3002\u5bf9\u4e8e LLM \u6765\u8bf4\uff0c\u8fd9\u662f\u4e00\u4e2a\u53c2\u6570\u6570\u91cf\u975e\u5e38\u5c11\u7684\u6a21\u578b\uff0c\u867d\u7136\u5b83\u7684\u5c40\u9650\u6027\u5c1a\u4e0d\u786e\u5b9a\uff0c\u4f46\u503c\u5f97\u5c1d\u8bd5\u3002<\/p>\n<h2>2\u3001\u8bbe\u7f6e\u5956\u52b1<\/h2>\n<p>\u5728\u539f\u59cb\u8bba\u6587\u4e2d\uff0c\u4f5c\u8005\u4f7f\u7528\u4e86\u4e00\u4e2a\u5177\u6709 671B \u4e2a\u53c2\u6570\u7684\u5927\u578b\u6df7\u5408\u4e13\u5bb6 (MoE) \u6a21\u578b\u3002\u8fd9\u79cd\u76f8\u5f53\u5927\u7684\u6a21\u578b\u89c4\u6a21\u4f7f\u4ed6\u4eec\u80fd\u591f\u4e3a\u6a21\u578b\u63d0\u4f9b\u76f8\u5f53\u5927\u7684\u81ea\u7531\u5ea6\uff0c\u4fc3\u8fdb\u81ea\u7531\u3001\u7a33\u5065\u6027\u548c\u521b\u9020\u529b\u3002\u7136\u800c\uff0c\u5c31\u6211\u800c\u8a00\uff0c\u4f7f\u7528 0.5B \u53c2\u6570\u6a21\u578b\uff08\u662f\u7684\uff0c\u4f60\u6ca1\u770b\u9519\uff01\uff09\uff0c\u6211\u611f\u89c9\u6709\u5fc5\u8981\u5b9e\u65bd\u66f4\u4e25\u683c\u7684\u89c4\u5219\u6765\u6709\u6548\u5730\u6307\u5bfc\u6a21\u578b\u3002\u539f\u59cb\u7684 DeepSeek-R1 \u6a21\u578b\u91c7\u7528\u4e86\u4e24\u79cd\u4e3b\u8981\u7684\u5956\u52b1\u6765\u5efa\u6a21\u5956\u52b1\uff1a\u51c6\u786e\u6027\u548c\u683c\u5f0f\u3002\u8fd9\u4e24\u4e2a\u5956\u52b1\u65e8\u5728\u8bc4\u4f30\u6700\u7ec8\u54cd\u5e94\u7684\u6b63\u786e\u6027\u4ee5\u53ca\u683c\u5f0f\u662f\u5426\u5305\u542b &nbsp;\u548c &nbsp;\u6807\u7b7e\u5185\u7684\u63a8\u7406\u6b65\u9aa4\u3002<\/p>\n<p>  \u8bad\u7ec3\u5956\u52b1\u56fe\u8868&nbsp; <\/p>\n<p>\u5728\u6211\u66f4\u53d7\u9650\u5236\u7684\u573a\u666f\u4e2d\uff0c\u6211\u52a0\u5165\u4e86\u4e00\u5957\u66f4\u5e7f\u6cdb\u7684\u89c4\u5219\u6765\u5956\u52b1\u6216\u60e9\u7f5a\u6a21\u578b\uff1a<\/p>\n<ul>\n<li>\u7b54\u6848\u662f\u5426\u6b63\u786e\uff1f<\/li>\n<li>\u5b83\u662f\u5426\u9075\u5faa\u5efa\u8bae\u7684\u683c\u5f0f\uff1f<\/li>\n<li>\u5b83\u662f\u5426\u5305\u542b\u63a8\u7406\u6807\u7b7e\uff1f<\/li>\n<li>\u5982\u679c\u5b58\u5728\u63a8\u7406\u6807\u7b7e\uff0c\u601d\u8003\u8fc7\u7a0b\u7684\u957f\u5ea6\u662f\u591a\u5c11\uff1f<\/li>\n<li>\u5b83\u662f\u5426\u5305\u542b\u9a8c\u8bc1\u6807\u7b7e\uff1f<\/li>\n<li>\u5982\u679c\u5b58\u5728\u9a8c\u8bc1\u6807\u7b7e\uff0c\u81ea\u6211\u6279\u8bc4\u7684\u957f\u5ea6\u662f\u591a\u5c11\uff1f<\/li>\n<li>\u5b83\u662f\u5426\u4f7f\u7528\u201c\u554a\u54c8\uff01\u201d\u65f6\u523b\u6807\u8bb0\uff1f<\/li>\n<\/ul>\n<p>\u6211\u8ba4\u8bc6\u5230\u5176\u4e2d\u4e00\u4e9b\u5956\u52b1\u662f\u4e3b\u89c2\u7684\uff0c\u53ef\u80fd\u4e0d\u5177\u6709\u5e7f\u6cdb\u7684\u666e\u904d\u6027\u3002\u4f46\u662f\uff0c\u6b63\u5982\u6240\u63d0\u5230\u7684\uff0c\u6211\u76f8\u4fe1\u5176\u4e2d\u51e0\u4e2a\u5bf9\u4e8e\u8f83\u5c0f\u7684\u6a21\u578b\u548c\u4fc3\u8fdb\u66f4\u5feb\u7684\u8bad\u7ec3\u662f\u5fc5\u4e0d\u53ef\u5c11\u7684\u3002\u4e3a\u4ee3\u7801\u4e2d\u771f\u6b63\u7684\u4efb\u610f\u5956\u52b1\u70b9\u505a\u597d\u51c6\u5907\uff01\u6b64\u5916\uff0c\u8003\u8651\u6211\u4eec\u6b63\u5728\u4f7f\u7528\u7684\u6570\u636e\u96c6\u4e5f\u5f88\u91cd\u8981\uff0c\u56e0\u4e3a\u5176\u4e2d\u4e00\u4e9b\u89c4\u5219\u662f\u6839\u636e\u5176\u7279\u5b9a\u7279\u5f81\u91cf\u8eab\u5b9a\u5236\u7684\u3002\u6570\u636e\u96c6\u662f GSM8K\uff08\u5c0f\u5b66\u6570\u5b66 8K\uff09\uff0c\u5b83\u5305\u542b\u8bed\u8a00\u591a\u6837\u7684\u5c0f\u5b66\u6570\u5b66\u5e94\u7528\u9898\u3002\u8be5\u6570\u636e\u96c6\u4e13\u95e8\u7528\u4e8e\u652f\u6301\u9700\u8981\u591a\u6b65\u63a8\u7406\u7684\u57fa\u672c\u6570\u5b66\u95ee\u9898\u7684\u95ee\u7b54\u4efb\u52a1\u3002\u56e0\u6b64\uff0c\u5176\u4e2d\u4e00\u4e9b\u89c4\u5219\u662f\u7279\u5b9a\u4e8e\u6b64\u6570\u636e\u96c6\u7684\u4efb\u52a1\uff0c\u53ef\u80fd\u65e0\u6cd5\u5f88\u597d\u5730\u6982\u62ec\u3002\u4e0d\u8fc7\uff0c\u8bf7\u5141\u8bb8\u6211\u89e3\u91ca\u4e00\u4e0b\u8fd9\u4e9b\u5956\u52b1\u9009\u62e9\u80cc\u540e\u7684\u4e00\u4e9b\u7406\u7531\u3002<\/p>\n<ul>\n<li>\u5bf9\u4e8e\u51c6\u786e\u6027\u5956\u52b1\uff0c\u6211\u4eec\u5c06\u63d0\u53d6 <code>&lt;answer&gt;&lt;\/answer&gt;<\/code> \u6807\u7b7e\u4e4b\u95f4\u7684\u6700\u7ec8\u7b54\u6848\u4f5c\u4e3a\u9884\u6d4b\uff0c\u5e76\u4f7f\u7528\u57fa\u672c\u4e8b\u5b9e\u7b54\u6848\u68c0\u67e5\u5b83\u4eec\u7684\u51c6\u786e\u6027\u3002<\/li>\n<li>\u683c\u5f0f\u5316\u90e8\u5206\u662f\u68c0\u67e5\u63a8\u7406\u3001\u6279\u8bc4\u548c\u7b54\u6848\u90e8\u5206\u7684\u6807\u7b7e\u662f\u5426\u9075\u5faa\u4e00\u822c\u7ed3\u6784\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e24\u4e2a\u4e0e\u539f\u59cb\u8bba\u6587\u76f8\u4f3c\uff0c\u4f46\u5176\u4f59\u7684\u53d6\u51b3\u4e8e\u4f60\u7684\u4efb\u52a1\u548c\u521b\u9020\u529b\u3002<\/p>\n<p>\u4e3a\u4e86\u9f13\u52b1\u66f4\u957f\u7684\u601d\u8003\u8fc7\u7a0b\uff0c\u6211\u8ba1\u7b97\u4e86\u63a8\u7406\u548c\u9a8c\u8bc1\u6807\u7b7e\u4e4b\u95f4\u7684\u5355\u8bcd\u6570\u91cf\uff0c\u5e76\u5c06\u5176\u4f5c\u4e3a\u5956\u52b1\u6dfb\u52a0\u5230\u4e00\u5b9a\u8303\u56f4\u5185\u3002\uff08\u4ee5\u9632\u6b62\u6a21\u578b\u5229\u7528\u65e0\u7528\u7684\u957f\u7b54\u6848\uff09<\/p>\n<p>\u539f\u59cb\u4f5c\u54c1\u63d0\u5230\u201c\u554a\u54c8\uff01\u201d\u65f6\u523b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u81ea\u7136\u53d1\u751f\u3002\u867d\u7136\u66f4\u5927\u7684\u6a21\u578b\u548c\u66f4\u957f\u7684\u8bad\u7ec3\u7a0b\u5e8f\u4e5f\u6709\u53ef\u80fd\u51fa\u73b0\u8fd9\u79cd\u60c5\u51b5\uff0c\u4f46\u5728\u6211\u7684\u573a\u666f\u4e2d\uff0c\u6211\u8bbe\u7f6e\u4e86\u4e00\u4e2a\u4fd7\u6c14\u7684\u6e38\u620f\u89c4\u5219\u6765\u5f3a\u5236\u8fd9\u4e00\u65f6\u523b\u3002<\/p>\n<p>\u5728\u6700\u8fd1\u7684\u4e00\u7bc7\u8bba\u6587 \u300as1\uff1a\u7b80\u5355\u7684\u6d4b\u8bd5\u65f6\u95f4\u7f29\u653e\u4e2d\u300b\uff0c\u4f5c\u8005\u5728\u6a21\u578b\u51c6\u5907\u7ed3\u675f\u9884\u6d4b\u5e8f\u5217\u65f6\u63d2\u5165\u4e86\u4e00\u4e2a\u201c\u7b49\u5f85\u201d\u6807\u8bb0\uff0c\u4ee5\u8feb\u4f7f\u5b83\u4eec\u91cd\u65b0\u601d\u8003\u3002\u8fd9\u542c\u8d77\u6765\u5f88\u50cf\u201c\u554a\u54c8\uff01\u201d\u65f6\u523b\uff0c\u6240\u4ee5\u6211\u5c31\u628a\u8fd9\u4e2a\u60f3\u6cd5\u63d2\u5165\u5230\u6211\u7684\u5956\u52b1\u51fd\u6570\u4e2d\u3002<\/p>\n<p>\u57fa\u672c\u4e0a\uff0c\u5982\u679c\u6709\u4e00\u4e2a\u201c\u554a\u54c8\uff01\u201d\u65f6\u523b\uff0c\u6bd4\u5982\u201c\u7b49\u5f85\uff0c\u4f46\u6216\u91cd\u65b0\u601d\u8003\u201d\uff0c\u7136\u540e\u662f\u6700\u5c11\u7684\u5b57\u7b26\uff0c\u6211\u4f1a\u8bbe\u7f6e\u5de8\u5927\u7684\u5956\u52b1\u6765\u9f13\u52b1\u8fd9\u79cd\u601d\u8003\uff0c\u5e76\u5c06\u5176\u8bbe\u7f6e\u4e3a\u7cfb\u7edf\u63d0\u793a\u3002<\/p>\n<h2>3\u3001\u5267\u900f\u8b66\u544a\uff01<\/h2>\n<p>\u4ee5\u4e0b\u662f\u8bad\u7ec3\u96c6\u9884\u6d4b\u7684\u4e00\u4e2a\u4f8b\u5b50\uff1a<\/p>\n<p>\u597d\u5427\uff0c\u770b\u770b\u8fd9\u4e2a\u8f93\u51fa\uff0c\u4f3c\u4e4e\u6a21\u578b\u6b63\u5728\u5c1d\u8bd5\u9075\u5faa\u6211\u5efa\u7acb\u7684\u89c4\u5219\u3002\u5c3d\u7ba1\u5b83\u5229\u7528\u4e86\u4e00\u4e9b\u65b9\u9762\uff08\u8fd9\u662f\u9884\u6599\u4e4b\u4e2d\u7684\uff09\uff0c\u4f46\u5b83\u4f3c\u4e4e\u662f\u4e00\u4e2a\u6709\u524d\u9014\u7684\u63a8\u7406\u6a21\u578b\uff0c\u5e76\u5e26\u6709\u4e00\u4e2a\u81ea\u6211\u53d1\u8d77\u7684\u201c\u7b49\u7b49\uff0c\u8ba9\u6211\u4eec\u91cd\u65b0\u601d\u8003\u201d\u7684\u65f6\u523b\u3002<\/p>\n<h2>4\u3001\u4ee3\u7801\u65f6\u95f4<\/h2>\n<p>\u597d\u7684\uff0c\u6211\u8ba4\u4e3a\u6211\u4eec\u5df2\u7ecf\u6db5\u76d6\u4e86\u8db3\u591f\u7684\u7406\u8bba\u80cc\u666f\u3002\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u6df1\u5165\u7814\u7a76\u5b9e\u9645\u7684\u4ee3\u7801\u5b9e\u73b0\u3002\u6211\u4eec\u5c06\u9996\u5148\u4e3a\u51e0\u4e2a\u5173\u952e\u7ec4\u4ef6\u521b\u5efa\u5b9e\u7528\u51fd\u6570\uff1a\u8bbe\u7f6e\u7cfb\u7edf\u63d0\u793a\u3001\u4ece\u6570\u636e\u96c6\u4e2d\u63d0\u53d6\u7b54\u6848\u3001\u63d0\u53d6\u6307\u5b9a\u6807\u7b7e\u4e4b\u95f4\u7684\u5185\u5bb9\u4ee5\u53ca\u52a0\u8f7d GSM8K \u6570\u636e\u96c6\u672c\u8eab\u3002<\/p>\n<pre><code>import re\nimport torch\nfrom datasets import load_dataset, Dataset\nfrom trl import GRPOConfig, GRPOTrainer\n\nfrom nltk.tokenize import word_tokenize\n\nSYSTEM_PROMPT = \"\"\"\nRespond in the following format:\n&lt;reasoning&gt;\n[Break and reason your answer here]\n&lt;\/reasoning&gt;\n&lt;validate&gt;\n[Criticize your reason and think about it. If you see somethin start with 'wait']\n&lt;\/validate&gt;\n&lt;answer&gt;\n[Final integer answer justified by validate]\n&lt;\/answer&gt;\n\"\"\"\n\ndef extract_hash_answer(text: str) -&gt; str | None:\n    if \"####\" not in text:\n        return None\n    return text.split(\"####\")[1].strip()\n\ndef extract_xml_answer(response, tag=\"answer\"):\n    import re\n    match = re.search(rf'&lt;{tag}&gt;(.*?)&lt;\/{tag}&gt;', response, re.DOTALL)\n    if match:\n        return match.group(1).strip()\n    return \"\"\n\ndef get_gsm8k_questions(split = \"train\") -&gt; Dataset:\n    data = load_dataset('openai\/gsm8k', 'main')[split]\n    data = data.map(lambda x: {\n        'prompt': [\n            {'role': 'system', 'content': SYSTEM_PROMPT},\n            {'role': 'user', 'content': x['question']}\n        ],\n        'answer': extract_hash_answer(x['answer'])\n    })\n    return data\n\ndataset = get_gsm8k_questions()<\/code><\/pre>\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u5b9e\u73b0\u6700\u91cd\u8981\u7684\u90e8\u5206\u4e4b\u4e00\uff1a\u5956\u52b1\u51fd\u6570\u3002\u6211\u6839\u636e\u8981\u70b9\u521b\u5efa\u4e86\u5b83\u3002\u6211\u5728\u4e0a\u4e00\u7ae0\u4e2d\u63d0\u5230\u8fc7\uff0c\u4f46\u4f60\u53ef\u4ee5\u968f\u610f\u4fee\u6539\u6216\u7f16\u5199\u81ea\u5df1\u7684\u4ee3\u7801\u3002<\/p>\n<pre><code>def custom_reward_func(prompts, completions, answer, min_reasoning_length=10, **kwargs) -&gt; list[float]:\n    responses = [completion[0]['content'] for completion in completions]\n    q = prompts[0][-1]['content']\n    extracted_responses_answer = [extract_xml_answer(r, tag=\"answer\") for r in responses]\n    extracted_responses_reasoning = [extract_xml_answer(r, tag=\"reasoning\") for r in responses]\n    extracted_responses_validate = [extract_xml_answer(r, tag=\"validate\") for r in responses]\n\n    rewards = []\n    for original_response, extracted_answer, extracted_reasoning, extracted_validate in zip(\n        responses, extracted_responses_answer, extracted_responses_reasoning, extracted_responses_validate\n    ):\n        is_correct = (extracted_answer == answer[0])\n        is_int = extracted_answer.isdigit()\n        has_answer_tags = \"&lt;answer&gt;\" in original_response and \"&lt;\/answer&gt;\" in original_response\n        has_reasoning_tags = \"&lt;reasoning&gt;\" in original_response and \"&lt;\/reasoning&gt;\" in original_response\n        has_validate_tags = \"&lt;validate&gt;\" in original_response and \"&lt;\/validate&gt;\" in original_response\n        reasoning_length = len(word_tokenize(extracted_reasoning.lower()))\n        validate_length = len(word_tokenize(extracted_validate.lower()))\n\n        reward = 0.0\n        reasoning_reward = 0.0\n\n        if is_correct:\n            reward += 5.0\n        if is_int:\n            reward += 0.5\n\n        if has_validate_tags:\n            reward *= 1.25\n            if validate_length &gt;= 5:\n                min_validate_length = 5\n                max_validate_length = 256\n                max_validate_bonus = 3.0\n                if validate_length &gt;= min_validate_length:\n                    if validate_length &gt;= max_validate_length:\n                        validate_bonus = max_validate_bonus\n                    else:\n                        validate_bonus = ((validate_length - min_validate_length) \/ (max_validate_length - min_validate_length)) * max_validate_bonus\n                else:\n                    validate_bonus = 0.0\n            else:\n                validate_bonus = 0.0\n        else:\n            validate_bonus = 0.0\n\n        if has_reasoning_tags:\n            reward *= 1.25\n            if reasoning_length &gt;= 5:\n                min_scaling_length = 5\n                max_scaling_length = 1024\n                max_scaling_bonus = 10\n                if reasoning_length &lt;= min_scaling_length:\n                    reasoning_reward = 0.0\n                elif reasoning_length &gt;= max_scaling_length:\n                    reasoning_reward = 5.0\n                else:\n                    reasoning_reward = ((reasoning_length - min_scaling_length) \/ (max_scaling_length - min_scaling_length)) * max_scaling_bonus\n            else:\n                reasoning_reward = 0.0\n        else:\n            reasoning_reward = 0.0\n\n        total_reward = reward + reasoning_reward + validate_bonus\n\n        if has_validate_tags:\n            validate_lower = extracted_validate.lower()\n            if re.search(r\"(wait|but|rethink)(?=.{20,})\", validate_lower, re.DOTALL):\n                total_reward *= 10.0\n\n        rewards.append(total_reward)\n\n    return rewards<\/code><\/pre>\n<p>\u597d\u7684\uff0c\u6211\u4eec\u5b8c\u6210\u4e86\u5b9e\u7528\u51fd\u6570\u548c\u5956\u52b1\u51fd\u6570\u3002\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u8fdb\u5165\u8bad\u7ec3\u811a\u672c\u3002\u4e3a\u4e86\u5728\u6709\u9650\u7684\u8ba1\u7b97\u80fd\u529b\u4e0b\u8fdb\u884c\u9ad8\u6548\u3001\u5feb\u901f\u7684\u8bad\u7ec3\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u4e00\u4e9b\u5e93\u548c\u6280\u672f\uff0c\u4f8b\u5982 Unsloth\u3001trl\u3001vLLM \u548c LoRA \u7b49\u3002<\/p>\n<p>\u6211\u4e0d\u4f1a\u5728\u672c\u6587\u4e2d\u6df1\u5165\u4ecb\u7ecd\u8fd9\u4e9b\u5185\u5bb9\uff0c\u4f46\u4f1a\u63d0\u5230\u5b83\u4eec\u5728\u6b64\u6d41\u7a0b\u4e2d\u7684\u4f5c\u7528\u3002<\/p>\n<p>\u57fa\u672c\u4e0a\uff0cUnsloth \u5141\u8bb8\u4f18\u5316\u7684\u6a21\u578b\u4ee5\u66f4\u5c0f\u7684\u5185\u5b58\u5360\u7528\u548c\u66f4\u5feb\u7684\u8fed\u4ee3\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n<pre><code>from unsloth import FastLanguageModel, PatchFastRL\nPatchFastRL(\"GRPO\", FastLanguageModel)\n\n\nmax_seq_length = 1024\nlora_rank = 32\n\nmodel, tokenizer = FastLanguageModel.from_pretrained(\n    model_name = \"unsloth\/Qwen2.5-0.5B-Instruct-unsloth-bnb-4bit\",\n    max_seq_length = max_seq_length,\n    load_in_4bit = False,\n    fast_inference = True, # Enable vLLM fast inference\n    max_lora_rank = lora_rank,\n    gpu_memory_utilization = 0.8,\n)\n\nmodel = FastLanguageModel.get_peft_model(\n    model,\n    r = lora_rank,\n    target_modules = [\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n                      \"gate_proj\", \"up_proj\", \"down_proj\"],\n    lora_alpha = lora_rank*2,\n    use_gradient_checkpointing = \"unsloth\",\n    random_state = 1907,\n)<\/code><\/pre>\n<p>\u901a\u8fc7\u5e94\u7528 LoRA\uff0c\u6211\u4eec\u6b63\u5728\u8bad\u7ec3\u4e00\u5c0f\u90e8\u5206\u6a21\u578b\u53c2\u6570\u3002\u7136\u540e\u6211\u4eec\u8bbe\u7f6e trl \u8bad\u7ec3\u53c2\u6570\u548c\u8bad\u7ec3\u5668\u672c\u8eab\u3002<\/p>\n<pre><code>output_dir=\"outputs\/Qwen-.5B-GRPO\"\n\ntraining_args = GRPOConfig(\n    output_dir=output_dir,\n    run_name=run_name,\n    learning_rate=3e-5,\n    adam_beta1 = 0.9,\n    adam_beta2 = 0.99,\n    weight_decay = 0.1,\n    warmup_steps=5,\n    lr_scheduler_type='cosine',\n    logging_steps=1,\n    bf16=True,\n    optim=\"paged_adamw_8bit\",\n    per_device_train_batch_size=1,\n    gradient_accumulation_steps=32,\n    num_generations=2,\n    max_prompt_length=256,\n    max_completion_length=512,\n    num_train_epochs=1,\n    save_steps=100,\n    max_grad_norm=1.0,\n    report_to=\"none\",\n    log_on_each_node=False,\n    use_vllm=True,\n)\n\ntrainer = GRPOTrainer(\n    model=model,\n    processing_class=tokenizer,\n    reward_funcs=[custom_reward_func],\n    args=training_args,\n    train_dataset=dataset,\n)\ntrainer.train()<\/code><\/pre>\n<p>\u8ba9\u5947\u8ff9\u53d1\u751f\u5427\uff01\u8fd9\u4e9b\u8d85\u53c2\u6570\u53ef\u80fd\u8fdc\u975e\u6700\u4f73\uff08\u5176\u4ed6\u7ec4\u4ef6\u4e5f\u662f\u5982\u6b64\uff09\uff0c\u4f46\u6211\u76f8\u4fe1\u5f00\u6e90\u793e\u533a\u5f88\u5feb\u5c31\u4f1a\u627e\u5230\u66f4\u597d\u7684\u65b9\u6cd5\u3002\uff08\u6211\u76ee\u524d\u7684\u8bbe\u7f6e\u662f\u57fa\u4e8e GRPO Llama-1B \u6784\u5efa\u7684\uff09\u3002<\/p>\n<p>  \u5b8c\u6210\u957f\u5ea6\u56fe\u8868 <\/p>\n<p>\u5982\u4e0a\u6240\u793a\uff0c\u8be5\u6a21\u578b\u901a\u8fc7\u5ef6\u957f\u5176\u63a8\u7406\u548c\u9a8c\u8bc1\u6b65\u9aa4\u4e0d\u65ad\u589e\u52a0\u5176\u8f93\u51fa\u3002\u4f46\u5728\u67d0\u4e2a\u65f6\u5019\uff0c\u5b83\u505c\u6b62\u4e86\u589e\u957f\uff0c\u539f\u56e0\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\u7f3a\u4e4f max_sequence \u9650\u5236\uff1a\u968f\u7740\u63a8\u7406\u6b65\u9aa4\u53d8\u957f\uff0c\u6a21\u578b\u5f00\u59cb\u524a\u51cf\u6700\u7ec8\u7b54\u6848\uff0c\u56e0\u6b64\u65e0\u6cd5\u83b7\u5f97\u91cd\u8981\u7684\u51c6\u786e\u6027\u5956\u52b1\u3002\uff08\u6211\u5728\u8bad\u7ec3\u5668\u4e2d\u8bbe\u7f6e\u7684\u9650\u5236\uff09<\/li>\n<li>\u7f3a\u4e4f\u6a21\u578b\u80fd\u529b\uff1a\u8fd9\u542c\u8d77\u6765\u53ef\u80fd\u50cf\u662f\u4e00\u4e2a\u501f\u53e3\uff0c\u4f46\u5bf9\u4e8e 0.5B \u53c2\u6570\u6a21\u578b\u6765\u8bf4\uff0c\u8fd9\u786e\u5b9e\u5f88\u91cd\u8981\u3002<\/li>\n<li>\u5956\u52b1\u51fd\u6570\u63a8\u7406\u957f\u5ea6\u7684\u4e0a\u9650\u3002<\/li>\n<\/ul>\n<h2>5\u3001\u8bc4\u4f30\u65f6\u95f4\uff01<\/h2>\n<p>\u5982\u679c\u4f60\u8ba4\u4e3a\u6211\u4f1a\u5728\u4e0d\u8bc4\u4f30\u6d4b\u8bd5\u96c6\u7ed3\u679c\u7684\u60c5\u51b5\u4e0b\u5f97\u51fa\u7ed3\u8bba\uff0c\u90a3\u4f60\u5c31\u9519\u4e86\uff0c\u54c8\u54c8\uff01\u5e78\u8fd0\u7684\u662f\uff0cGSM8K \u5305\u542b\u4e00\u4e2a\u6d4b\u8bd5\u62c6\u5206\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u8bc4\u4f30\u7ed3\u679c\u5e76\u8fdb\u884c\u6bd4\u8f83\u3002\u9996\u5148\uff0c\u8ba9\u6211\u4eec\u5728\u539f\u59cb\u6307\u4ee4\u6a21\u578b\u4e0a\u4f7f\u7528\u76f8\u540c\u7684\u7cfb\u7edf\u63d0\u793a\u6765\u5efa\u7acb\u57fa\u7ebf\u7ed3\u679c\u3002\u6211\u5c06\u91c7\u6837\u8bbe\u7f6e\u4e3a False\uff0c\u4ee5\u786e\u4fdd\u4e24\u7ec4\u9884\u6d4b\u7684\u7ed3\u679c\u90fd\u662f\u786e\u5b9a\u6027\u7684\u3002<\/p>\n<p>\u539f\u59cb\u6a21\u578b\u7684\u51c6\u786e\u5ea6\u5f97\u5206\uff1a0.0887035633055345<\/p>\n<p>\u55ef\u2026\u2026\u8fd9\u4e0d\u662f\u5f88\u597d\u2026\u2026\u4f46\u662f\uff0c\u91cd\u8981\u7684\u662f\u8981\u8003\u8651\u5230\u6a21\u578b\u6ca1\u6709\u660e\u786e\u8bad\u7ec3\u4e3a\u4ec5\u63d0\u4f9b\u6574\u6570\u7b54\u6848\uff0c\u9664\u4e86\u7cfb\u7edf\u63d0\u793a\u6240\u6697\u793a\u7684\u5185\u5bb9\u3002\u56e0\u6b64\uff0c\u6709\u65f6\u5373\u4f7f\u683c\u5f0f\u4e0d\u5b8c\u7f8e\uff0c\u9884\u6d4b\u4e2d\u4e5f\u53ef\u80fd\u5b58\u5728\u6b63\u786e\u7b54\u6848\u3002\u4e0d\u8fc7\uff0c\u6211\u4eec\u53ef\u4ee5\u89c2\u5bdf\u5230\u6a21\u578b\u8bd5\u56fe\u9075\u5faa\u6307\u4ee4\uff0c\u4f46\u5e76\u6ca1\u6709\u5b8c\u5168\u7406\u89e3\u5b83\u4eec\u7684\u542b\u4e49\uff0c\u6216\u8005\u53ea\u662f\u91cd\u590d\u63d0\u793a\u4e2d\u7684\u4e00\u6b21\u6027\u793a\u4f8b\u2026\u2026<\/p>\n<p>  \u539f\u59cb\u6a21\u578b\u9884\u6d4b <\/p>\n<p>\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u4f7f\u7528\u76f8\u540c\u7684\u8bbe\u7f6e\u6d4b\u8bd5 GRPO \u8c03\u6574\u6a21\u578b\uff1a<\/p>\n<p>\u8c03\u6574\u6a21\u578b\u7684\u51c6\u786e\u5ea6\u5f97\u5206\uff1a0.6527672479150872<\/p>\n<p>\u54c7\uff01\u5bf9\u4e8e\u8fd9\u4e48\u5c0f\u7684\u6a21\u578b\uff0c\u8fd9\u6bd4\u6211\u9884\u671f\u7684\u8981\u597d\u3002\u6211\u751a\u81f3\u4e0d\u5f97\u4e0d\u4ed4\u7ec6\u68c0\u67e5\u662f\u5426\u5b58\u5728\u6f5c\u5728\u7684\u6570\u636e\u6cc4\u6f0f\u3002\u5373\u4f7f\u6ca1\u6709\u6cc4\u6f0f\uff0c\u6211\u4eec\u4e5f\u53ef\u80fd\u5df2\u7ecf\u5fae\u8c03\u4e86\u8fd9\u4e2a\u6a21\u578b\u4ee5\u8fc7\u5ea6\u62df\u5408\u8fd9\u79cd\u7279\u5b9a\u7c7b\u578b\u7684\u4efb\u52a1\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u672c\u8d28\u4e0a\u62e5\u6709\u7684\u4e0d\u662f\u901a\u7528\u6a21\u578b\uff0c\u800c\u662f\u4e13\u95e8\u7684 GSM8K \u8ba1\u7b97\u5668\u6a21\u578b\u3002\u4f46\u8fd9\u4ecd\u7136\u662f\u4e00\u4e2a\u91cd\u5927\u7684\u8fdb\u6b65\uff0c\u4e3a\u8bb8\u591a\u4e13\u95e8\u7684\u6a21\u578b\u8bad\u7ec3\u5de5\u4f5c\uff08\u6216\u901a\u7528\u5927\u6a21\u578b\uff09\u6253\u5f00\u4e86\u5927\u95e8\uff0c\u8fd9\u4e5f\u662f\u6211\u975e\u5e38\u611f\u5174\u8da3\u7684\u53e6\u4e00\u4e2a\u8bdd\u9898\u2026\u2026<\/p>\n<p>  \u5fae\u8c03\u6a21\u578b\u9884\u6d4b <\/p>\n<h2>6\u3001\u7ed3\u675f\u8bed<\/h2>\n<p>\u6240\u4ee5\uff0c\u5c31\u662f\u8fd9\u6837\uff01\u6211\u4eec\u91c7\u7528\u4e86\u4e00\u4e2a\u5fae\u5c0f\u7684 0.5B \u53c2\u6570\u6a21\u578b\uff0c\u5728\u6211\u4eec\u7684\u5bb6\u5ead\u5b9e\u9a8c\u5ba4\uff08\u6216\u8005\uff0c\u4f60\u77e5\u9053\uff0c\u6e38\u620f\u88c5\u5907\uff01\uff09\u4e2d\u5411\u5b83\u6295\u5165\u4e86\u4e00\u4e9b GRPO\uff0c\u5e76\u8bbe\u6cd5\u6559\u4f1a\u5b83\u4e00\u4e9b\u76f8\u5f53\u4e0d\u9519\u7684\u63a8\u7406\u6280\u80fd\uff0c\u81f3\u5c11\u5bf9\u4e8e\u6570\u5b66\u95ee\u9898\u6765\u8bf4\u662f\u8fd9\u6837\u3002\u8c01\u4f1a\u60f3\u5230\u5462\uff0c\u5bf9\u5427\uff1f\u867d\u7136\u6211\u4eec\u7684\u5c0f Qwen-GRPO \u4e0d\u4f1a\u5f88\u5feb\u53d6\u4ee3 DeepSeek-R1\uff0c\u4f46\u6211\u4eec\u770b\u5230\u7684\u51c6\u786e\u5ea6\u7684\u98de\u8dc3\u786e\u5b9e\u4ee4\u4eba\u5174\u594b\u3002\u5b83\u8bc1\u660e\uff0c\u5373\u4f7f\u8d44\u6e90\u6709\u9650\uff0c\u6211\u4eec\u4e5f\u53ef\u4ee5\u5f00\u59cb\u4f7f\u7528\u8fd9\u4e9b\u5148\u8fdb\u7684\u6280\u672f\uff0c\u5e76\u7a81\u7834\u5c0f\u578b\u6a21\u578b\u7684\u6781\u9650\u3002<\/p>\n<p>\u6574\u4e2a\u5b9e\u9a8c\u786e\u5b9e\u5f3a\u8c03\u4e86\u4e00\u4e2a\u5173\u952e\u70b9\uff1a\u63a8\u7406\u6a21\u578b\u4e0d\u4ec5\u9002\u7528\u4e8e\u62e5\u6709\u65e0\u9650\u8ba1\u7b97\u80fd\u529b\u7684\u5927\u578b\u5b9e\u9a8c\u5ba4\u3002\u4f60\u53ef\u4ee5\u5728\u5bb6\u91cc\u6446\u5f04\u8fd9\u4e9b\u4e1c\u897f\uff0c\u5b66\u5230\u5f88\u591a\u4e1c\u897f\uff0c\u751a\u81f3\u5f97\u5230\u4ee4\u4eba\u60ca\u8bb6\u7684\u597d\u7ed3\u679c\u3002\u5f53\u7136\uff0c\u8fd8\u6709\u5f88\u957f\u7684\u8def\u8981\u8d70\u2014\u2014\u6cdb\u5316\u3001\u66f4\u5f3a\u5927\u7684\u5956\u52b1\uff0c\u751a\u81f3\u53ef\u80fd\u5f04\u6e05\u695a\u5982\u4f55\u8ba9\u90a3\u4e2a\u201c\u554a\u54c8\uff01\u201d\u65f6\u523b\u4e0d\u90a3\u4e48\u4fd7\u6c14\u2014\u2014\u4f46\u8fd9\u53ea\u662f\u4e00\u4e2a\u5f00\u59cb\u3002\u5f00\u6e90\u793e\u533a\u5145\u6ee1\u4e86\u5404\u79cd\u60f3\u6cd5\uff0c\u6211\u8feb\u4e0d\u53ca\u5f85\u5730\u60f3\u770b\u770b\u6211\u4eec\u80fd\u4e00\u8d77\u6784\u5efa\u4ec0\u4e48\u4ee4\u4eba\u60ca\u53f9\u7684\u63a8\u7406\u6a21\u578b\uff0c\u4e00\u7816\u4e00\u74e6\uff0c\u6216\u8005\u4e00\u884c\u4e00\u884c\u5730\u7f16\u5199\u4ee3\u7801\uff01\u5728\u5bb6\u63a8\u7406\u7684\u672a\u6765\uff1f\u5b83\u770b\u8d77\u6765\u51fa\u5947\u5730\u5149\u660e\uff0c\u800c\u6211\uff0c\u5c31\u6211\u800c\u8a00\uff0c\u5df2\u7ecf\u7cfb\u597d\u5b89\u5168\u5e26\uff0c\u51c6\u5907\u597d\u8e0f\u4e0a\u65c5\u7a0b\u4e86\uff01<\/p>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>\u63a8\u7406\u6a21\u578b\u5e76\u4e0d\u5b8c\u5168\u662f\u65b0\u4e8b\u7269\uff0c\u4f46 DeepSeek \u7684 R1 \u6a21\u578b\u7684\u53d1\u5e03\u5728\u5f00\u6e90\u793e\u533a\u4e2d\u5f15\u8d77\u4e86\u6781\u5927\u7684\u5174\u594b\u3002\u5728\u9605\u8bfb DeepSeek-R1 \u8bba\u6587\u540e\uff0c\u6211\u5bf9\u4ed6\u4eec\u7684\u540e\u8bad\u7ec3\u6280\u672f\u7279\u522b\u611f\u5174\u8da3\uff0c\u636e\u62a5\u9053\uff0c\u8be5\u6280\u672f\u53ef\u4ee5\u663e\u7740\u63d0\u9ad8\u6027\u80fd\u3002 \u4f20\u7edf\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u662f\u76d1\u7763\u5fae\u8c03 (SFT)\u3002\u5728 SFT \u4e2d\uff0c\u4e3a\u6a21\u578b\u63d0\u4f9b\u63d0\u793a\u548c\u671f\u671b\u7684\u8f93\u51fa\uff0c\u4ee5\u4e86\u89e3\u7528\u6237\u504f\u597d\u3001\u7ed3\u6784\u5bf9\u9f50\u3001\u5b89\u5168\u534f\u8bae\u7b49\u3002\u672c\u8d28\u4e0a\uff0cSFT \u5c55\u793a\u4e86\u5f00\u53d1\u4eba\u5458\u6216\u7528\u6237\u5bf9\u6a21\u578b\u7684\u671f\u671b\uff0c\u5e76\u5f15\u5bfc\u5b83\u4e0e\u8fd9\u4e9b\u793a\u4f8b\u4fdd\u6301\u4e00\u81f4\u3002 \u65e2\u7136\u5df2\u7ecf\u63d0\u5230\u4e86\u540e\u8bad\u7ec3\uff0c\u90a3\u4e48\u5728\u6df1\u5165\u7814\u7a76 R1 \u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u4e4b\u524d\uff0c\u503c\u5f97\u7b80\u8981\u8ba8\u8bba\u4e00\u4e0b\u9884\u8bad\u7ec3\uff0c\u56e0\u4e3a\u4e24\u8005\u662f\u76f8\u5173\u7684\u3002\u8bad\u7ec3\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u521d\u59cb\u6b65\u9aa4\u4e4b\u4e00\u662f\u56e0\u679c\u8bed\u8a00\u5efa\u6a21\u3002\u5728\u6b64\u9636\u6bb5\uff0c\u6a21\u578b\u901a\u8fc7\u5173\u6ce8\u524d\u9762\u7684\u6807\u8bb0\u6765\u5b66\u4e60\u9884\u6d4b\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u4e2a\u6807\u8bb0\u3002\u6b64\u8fc7\u7a0b\u4ea7\u751f\u201c\u57fa\u7840\u201d\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u4ee5\u534a\u76d1\u7763\u7684\u65b9\u5f0f\u5728\u5927\u91cf\u516c\u5171\u6587\u672c\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6ca1\u6709\u660e\u786e\u7684\u6807\u7b7e\uff0c\u4ee5\u9884\u6d4b\u540e\u7eed\u6807\u8bb0\u3002\u8bf4\u5230\u516c\u5171\u6587\u672c\u6570\u636e\uff0c\u6211\u4eec\u5373\u5c06\u8017\u5c3d\u5b83\uff08\u5728\u67d0\u79cd\u7a0b\u5ea6\u4e0a\uff0c\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u5ffd\u7565\u5176\u4ed6\u6a21\u5f0f\uff09\u3002\u8fd9\u79cd\u7a00\u7f3a\u6027\u662f\u7814\u7a76\u4eba\u5458\u4e13\u6ce8\u4e8e LLM \u540e\u8bad\u7ec3\u9636\u6bb5\u7684\u4e3b\u8981\u9a71\u52a8\u529b\u4e4b\u4e00\u3002 \u56de\u5230\u540e\u8bad\u7ec3\uff0c\u5b58\u5728\u8bb8\u591a\u6280\u672f\uff0c\u4f8b\u5982 SFT\u3001PPO \u548c DPO \u7b49\u3002\u867d\u7136\u67d0\u4e9b\u65b9\u6cd5\u5728\u7279\u5b9a\u9886\u57df\u8868\u73b0\u66f4\u597d\uff0c\u4f46\u5b83\u4eec\u901a\u5e38\u4f1a\u9650\u5236\u6307\u4ee4\u5bf9\u9f50\u7684\u521b\u4f5c\u81ea\u7531\uff0c\u6216\u8005\u5728\u957f\u65f6\u95f4\u8fd0\u884c\u4e2d\u53d8\u5f97\u8fc7\u4e8e\u6602\u8d35\u3002\u76f8\u53cd\uff0c\u57fa\u7840\u6a21\u578b\u6ca1\u6709\u8fd9\u4e9b\u9650\u5236\uff0c\u4f46\u4e0d\u53ef\u9884\u6d4b\u3001\u96be\u4ee5\u6307\u5bfc\uff0c\u5e76\u4e14\u7ecf\u5e38\u4ea7\u751f\u4eba\u7c7b\u96be\u4ee5\u7406\u89e3\u7684\u8f93\u51fa\u3002 R1 \u7684\u4f5c\u8005\u53d1\u73b0\u4e86\u4e00\u79cd\u4e24\u5168\u5176\u7f8e\u7684\u65b9\u6cd5\uff1a\u5c06\u539f\u59cb\u6a21\u578b\u4e0e\u6307\u4ee4\u5bf9\u9f50\uff0c\u540c\u65f6\u4fdd\u7559\u4e00\u5b9a\u7a0b\u5ea6\u7684\u521b\u4f5c\u81ea\u7531\u3002\u6211\u4e0d\u4f1a\u5728\u8fd9\u91cc\u8be6\u7ec6\u4ecb\u7ecd\u6574\u4e2a\u8fc7\u7a0b\uff0c\u4f46\u4ece\u6839\u672c\u4e0a\u8bb2\uff0c\u5b83\u4eec\u5141\u8bb8\u57fa\u7840\u6a21\u578b\u4ee5\u76f8\u5f53\u5927\u7684\u81ea\u7531\u5ea6\u201c\u63a8\u7406\u201d\uff0c\u540c\u65f6\u8bbe\u5b9a\u4e13\u6ce8\u4e8e\u51c6\u786e\u6027\u548c\u683c\u5f0f\u7684\u7b80\u5355\u76ee\u6807\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u4ed6\u4eec\u91c7\u7528\u4e86\u5f3a\u5316\u5b66\u4e60\uff0c\u5b9e\u65bd\u4e86\u4e00\u79cd\u79f0\u4e3a\u201cGRPO\u201d\uff08\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff09\u7684\u4f18\u5316\u6280\u672f\u3002 GRPO \u6d88\u9664\u4e86\u5bf9\u6279\u8bc4\u6a21\u578b\uff08\u4f20\u7edf\u4e0a\u4e0e\u7b56\u7565\u6a21\u578b\u5927\u5c0f\u76f8\u540c\u5e76\u7528\u4e8e PPO\/DPO \u7b49\u65b9\u6cd5\uff09\u7684\u9700\u6c42\uff0c\u800c\u662f\u4f7f\u7528\u805a\u5408\u7ec4\u5206\u6570\u8ba1\u7b97\u57fa\u7ebf\u3002 \u5c06\u8fd9\u79cd\u65b9\u6cd5\u76f4\u63a5\u5e94\u7528\u4e8e\u57fa\u7840\u6a21\u578b\u4ea7\u751f\u4e86\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4ece\u800c\u4ea7\u751f\u4e86 R1 \u6a21\u578b\u7684\u7b2c\u4e00\u4e2a\u7248\u672c\uff1aDeepSeek-R1-Zero\u3002\u5c3d\u7ba1\u8be5\u6a21\u578b\u63d0\u9ad8\u4e86\u57fa\u7840\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5176\u8f93\u51fa\u901a\u5e38\u96be\u4ee5\u88ab\u4eba\u7c7b\u9605\u8bfb\u6216\u89e3\u91ca\u3002\u5982\u524d\u6240\u8ff0\uff0cSFT \u65b9\u6cd5\u6709\u52a9\u4e8e\u6a21\u578b\u9075\u5faa\u4eba\u7c7b\u7684\u504f\u597d\u548c\u53ef\u8bfb\u6027\u3002\u7136\u800c\uff0c\u6211\u4eec\u8fd8\u8ba8\u8bba\u4e86 SFT \u5982\u4f55\u627c\u6740\u521b\u4f5c\u81ea\u7531\u548c\u6cdb\u5316\uff0c\u5982\u8bba\u6587\u201cSFT \u8bb0\u5fc6\uff0cRL \u6cdb\u5316\uff1a\u57fa\u7840\u6a21\u578b\u540e\u8bad\u7ec3\u7684\u6bd4\u8f83\u7814\u7a76\u201d\u4e2d\u6240\u8ff0\u3002 \u56e0\u6b64\uff0c\u4f5c\u8005\u5728\u5f3a\u5316\u5b66\u4e60\u6b65\u9aa4\u4e4b\u524d\u4f7f\u7528\u4e86\u4e00\u4e2a\u5c0f\u578b\u3001\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u96c6\u6765\u201c\u51b7\u542f\u52a8\u201d\u4ed6\u4eec\u7684\u6a21\u578b\u3002\u8fd9\u79cd\u65b9\u6cd5\u5728\u8457\u540d\u7684 DeepSeek-R1 \u6a21\u578b\u7684\u521b\u5efa\u4e2d\u53d1\u6325\u4e86\u91cd\u8981\u4f5c\u7528\u3002\u5728\u4e0b\u4e00\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8\u5176\u4e2d\u4e00\u4e9b\u6280\u672f\u7684\u5b9e\u73b0\uff0c\u4ee5\u5b9e\u73b0\u7c7b\u4f3c\u7684\u7ed3\u679c\u3002 1\u3001\u5f3a\u5236\u63a8\u7406 \u5728\u4ed6\u4eec\u7684\u62a5\u544a\u4e2d\uff0cDeepSeek \u63d0\u5230\u4e86\u4e00\u7ec4\u65b0\u7684\u84b8\u998f\u6a21\u578b\u3002\u8fd9\u4e9b\u660e\u663e\u8f83\u5c0f\u7684\u6a21\u578b\u4f7f\u7528 SFT \u7684\u84b8\u998f\u6280\u672f\u8fdb\u884c\u8bad\u7ec3\uff0c\u5229\u7528\u4e86 DeepSeek-R1 \u7684\u8f93\u51fa\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e9b\u63d0\u70bc\u6a21\u578b\u6ca1\u6709\u7ecf\u8fc7\u4efb\u4f55\u5f3a\u5316\u5b66\u4e60\u6b65\u9aa4\uff0c\u8fd9\u662f\u4e00\u4e2a\u6709\u8da3\u7684\u89c2\u70b9\uff0c\u4e5f\u662f\u6211\u4eec\u503c\u5f97\u63a2\u7d22\u7684\u9886\u57df\u3002\u56e0\u6b64\uff0c\u5982\u679c\u6211\u4eec\u4e0d\u63d0\u70bc\u6a21\u578b\uff0c\u800c\u662f\u76f4\u63a5\u5728\u539f\u59cb\u6a21\u578b\u4e0a\u5b9e\u73b0 GRPO\uff0c\u4f1a\u600e\u4e48\u6837\uff1f \u6211\u4eec\u53ef\u4ee5\u7ee7\u7eed\u4f7f\u7528\u57fa\u7840\u6a21\u578b\u6216\u6307\u5bfc\u6a21\u578b\uff0c\u5047\u8bbe\u6307\u5bfc\u6a21\u578b\u5df2\u7ecf\u201c\u51b7\u542f\u52a8\u201d\u3002\u9274\u4e8e\u8fd9\u662f\u5728\u6211\u914d\u5907 RTX 3090 \u7684\u672c\u5730\u6e38\u620f\u88c5\u5907\u4e0a\u8fdb\u884c\u7684\u5b9e\u9a8c\uff0c\u6211\u9009\u62e9\u4e86 Qwen2.5\u20130.5B-Instruct \u6a21\u578b\u4f5c\u4e3a\u5f00\u59cb\u3002\u5bf9\u4e8e [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53793","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53793","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53793"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53793\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53793"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53793"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53793"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}