{"id":39150,"date":"2025-03-03T00:01:38","date_gmt":"2025-03-02T16:01:38","guid":{"rendered":"https:\/\/17aitech.com\/?p=39150"},"modified":"2025-03-03T00:01:38","modified_gmt":"2025-03-02T16:01:38","slug":"deepseek%e5%85%b3%e9%94%aerl%e7%ae%97%e6%b3%95grpo%ef%bc%8c%e6%9c%89%e4%ba%ba%e4%bb%8e%e5%a4%b4%e8%b7%91%e9%80%9a%e4%ba%86%ef%bc%8c%e8%b4%a1%e7%8c%ae%e5%ae%8c%e6%95%b4%e4%bb%a3%e7%a0%81","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=39150","title":{"rendered":"DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-02-3\" target=\"_blank\">DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801<\/a><\/p>\n<section>GRPO\uff08Group Relative Policy Optimization\uff09\u662f DeepSeek-R1 \u6210\u529f\u7684\u57fa\u7840\u6280\u672f\u4e4b\u4e00\uff0c\u6211\u4eec\u4e4b\u524d\u4e5f\u591a\u6b21\u62a5\u9053\u8fc7\u8be5\u6280\u672f\uff0c\u6bd4\u5982\u300a<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650953996&amp;idx=2&amp;sn=97226369385027d76133dd1c6b39c497&amp;scene=21#wechat_redirect\" target=\"_blank\">DeepSeek \u7528\u7684 GRPO \u5360\u7528\u5927\u91cf\u5185\u5b58\uff1f\u6709\u4eba\u7ed9\u51fa\u4e86\u4e9b\u7834\u89e3\u65b9\u6cd5<\/a>\u300b\u3002<\/section>\n<section><\/section>\n<section>\u7b80\u5355\u6765\u8bf4\uff0cGRPO \u7b97\u6cd5\u4e22\u5f03\u4e86 critic model\uff0c\u653e\u5f03\u4e86\u4ef7\u503c\u51fd\u6570\u8fd1\u4f3c\uff0c\u8f6c\u800c\u901a\u8fc7\u7ec4\u5185\u6837\u672c\u7684\u76f8\u5bf9\u6bd4\u8f83\u6765\u8ba1\u7b97\u7b56\u7565\u68af\u5ea6\uff0c\u4ece\u800c\u6709\u6548\u964d\u4f4e\u4e86\u8bad\u7ec3\u7684\u4e0d\u7a33\u5b9a\u6027\uff0c\u540c\u65f6\u63d0\u9ad8\u4e86\u5b66\u4e60\u6548\u7387\u3002<\/section>\n<section><\/section>\n<section>\u65e2\u7136 GRPO \u5982\u6b64\u6709\u6548\uff0c\u90a3\u4e48\uff0c\u4f60\u77e5\u9053\u5982\u4f55\u4ece\u5934\u5f00\u59cb\u5b9e\u73b0 GRPO \u5417\uff1f<\/section>\n<section><\/section>\n<section>\u8fd1\u65e5\uff0cAI \u5de5\u7a0b\u5e08\u548c\u6280\u672f\u4f5c\u5bb6 Andriy Burkov \u53d1\u5e03\u4e86\u4e00\u4efd\u300c\u4ece\u5934\u5f00\u59cb\u5199 GRPO \u4ee3\u7801\u300d\u7684\u6559\u7a0b\uff0c\u5176\u4e2d\u4ecb\u7ecd\u4e86\u5982\u4f55\u57fa\u4e8e Qwen2.5-1.5B-Instruct \u6a21\u578b\u6784\u5efa\u4e00\u4e2a\u4f7f\u7528 GRPO \u7684\u5206\u5e03\u5f0f\u5f3a\u5316\u5b66\u4e60\u6d41\u7a0b\u3002<\/section>\n<section><\/section>\n<section>\u4e0d\u8fc7\uff0c\u5728\u6211\u4eec\u6df1\u5165\u8fd9\u4efd\u6559\u7a0b\u4e4b\u524d\uff0c\u5148\u7b80\u5355\u4ecb\u7ecd\u4e00\u4e0b\u5b83\u7684\u4f5c\u8005\u3002Andriy Burkov \u7b97\u5f97\u4e0a\u662f AI \u9886\u57df\u7684\u4e00\u4f4d\u8457\u540d\u79d1\u666e\u4f5c\u5bb6\uff0c\u5728\u52a0\u62ff\u5927\u62c9\u74e6\u5c14\u5927\u5b66\u53d6\u5f97\u4e86\u8ba1\u7b97\u673a\u79d1\u5b66\u535a\u58eb\u5b66\u4f4d\uff0c\u8fd8\u66fe\u53d1\u8868\u8fc7\u4e24\u672c\u9887\u53d7\u6b22\u8fce\u7684 AI \u4e3b\u9898\u8457\u4f5c\uff1a\u300a100 \u9875\u8bed\u8a00\u6a21\u578b\u4e66\u300b\u548c\u300a100 \u9875\u673a\u5668\u5b66\u4e60\u4e66\u300b\uff1b\u4e66\u4e2d\u4e00\u6b65\u6b65\u8be6\u5b9e\u5730\u4ecb\u7ecd\u4e86\u76f8\u5173\u6982\u5ff5\uff0c\u5e76\u9644\u5e26\u4e86\u7b80\u660e\u7684\u5b9e\u73b0\u4ee3\u7801\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png\"><\/a><\/section>\n<section>\u63a5\u4e0b\u6765\u6211\u4eec\u5c31\u6765\u770b\u770b\u8fd9\u4efd GRPO \u4ece\u5934\u5b9e\u73b0\u6559\u7a0b\u5427\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-628493f3bb0beb64d8d3a858d4add525.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-628493f3bb0beb64d8d3a858d4add525.png\"><\/a><\/section>\n<section>\u6559\u7a0b\u5730\u5740\uff1ahttps:\/\/github.com\/aburkov\/theLMbook\/blob\/main\/GRPO_From_Scratch_Multi_GPU_DataParallel_Qwen_2_5_1_5B_Instruct.ipynb<\/section>\n<section><\/section>\n<section><strong>\u4ece\u5934\u7f16\u5199 GRPO \u4ee3\u7801<\/strong><\/section>\n<section><strong>\u4f7f\u7528 Qwen2.5-1.5B-Instruct \u7684\u5206\u5e03\u5f0f\u5b9e\u73b0<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u6559\u7a0b\u5c06\u5c55\u793a\u5982\u4f55\u4f7f\u7528 GRPO \u65b9\u6cd5\u6784\u5efa\u5206\u5e03\u5f0f\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6d41\u7a0b\uff0c\u4ece\u800c\u53ef\u4ee5\u9488\u5bf9\u6570\u5b66\u3001\u903b\u8f91\u548c\u7f16\u7a0b\u4efb\u52a1\u5bf9\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002<\/section>\n<section><\/section>\n<section>\u9996\u5148\u9700\u8981\u660e\u786e\uff0c\u8fd9\u4e9b\u4efb\u52a1\u90fd\u5b58\u5728\u4e00\u4e2a\u552f\u4e00\u4e14\u6b63\u786e\u7684 ground truth \u7b54\u6848\uff0c\u53ef\u901a\u8fc7\u7b80\u5355\u7684\u5b57\u7b26\u4e32\u6bd4\u8f83\u8f7b\u677e\u52a0\u4ee5\u9a8c\u8bc1\u3002<\/section>\n<section><\/section>\n<section>GRPO \u7684\u53d1\u660e\u8005\u662f DeepSeek\uff0c\u6700\u65e9\u662f\u88ab\u7528\u4e8e\u5fae\u8c03 DeepSeek \u7684 R1 \u548c R1-Zero \u6a21\u578b \u2014\u2014 \u5b83\u4eec\u53ef\u901a\u8fc7\u5b66\u4e60\u751f\u6210\u601d\u7ef4\u94fe\uff08CoT\uff09\u6765\u66f4\u597d\u5730\u89e3\u51b3\u6570\u5b66\u548c\u903b\u8f91\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section>\u672c\u6559\u7a0b\u7684\u76ee\u6807\u662f\u5c06\u901a\u7528\u8bed\u8a00\u6a21\u578b Qwen2.5-1.5B-Instruct \u8f6c\u6362\u4e3a\u6570\u5b66\u95ee\u9898\u6c42\u89e3\u5668\u3002\u6211\u4eec\u5c06\u4ece\u5934\u5f00\u59cb\u7f16\u5199 GRPO \u4ee3\u7801\uff0c\u7136\u540e\u5c06\u5176\u4e0e\u51e0\u4e2a\u6d41\u884c\u7684\u5e93\u548c\u5de5\u5177\u96c6\u6210\u8d77\u6765\uff0c\u4ee5\u5b9e\u73b0\u5206\u5e03\u5f0f\u8bad\u7ec3\u7ba1\u9053\u6d41\u7a0b\uff0c\u5305\u62ec\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>PyTorch\uff1a\u7528\u4e8e\u5f20\u91cf\u8fd0\u7b97\u548c\u5206\u5e03\u5f0f\u8bad\u7ec3\u3002<\/section>\n<\/li>\n<li>\n<section>Hugging Face Transformers\uff1a\u7528\u4e8e\u52a0\u8f7d\u9884\u8bad\u7ec3\u7684\u8bed\u8a00\u6a21\u578b\u548c tokenizer\u3002<\/section>\n<\/li>\n<li>\n<section>FlashAttention2\uff1a\u4f18\u5316\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u6709\u52a9\u4e8e\u51cf\u5c11\u5185\u5b58\u4f7f\u7528\u91cf\u5e76\u63d0\u9ad8\u8bad\u7ec3\u901f\u5ea6\u3002<\/section>\n<\/li>\n<li>\n<section>Weights &amp; Biases (wandb)\uff1a\u7528\u4e8e\u5b9e\u9a8c\u8ddf\u8e2a\u3001\u53ef\u89c6\u5316\u548c\u6a21\u578b\u7248\u672c\u63a7\u5236\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u672c\u6559\u7a0b\u5206\u4e3a\u51e0\u4e2a\u90e8\u5206\u3002\u9996\u5148\u662f\u57fa\u672c\u8bbe\u7f6e\u548c\u5bfc\u5165\uff0c\u7136\u540e\u662f\u6570\u636e\u683c\u5f0f\u5316\u548c\u7b54\u6848\u63d0\u53d6\u3001\u6570\u636e\u96c6\u51c6\u5907\u3001\u8bc4\u4f30\u51fd\u6570\u3001\u5956\u52b1\u51fd\u6570\u3001\u8bad\u7ec3\u8bbe\u7f6e\u548c\u6267\u884c\uff0c\u6700\u540e\u52a0\u8f7d\u548c\u6d4b\u8bd5\u6a21\u578b\u3002\u6b64\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u4ece\u5934\u5b9e\u73b0 GRPO \u7b97\u6cd5\u3002<\/section>\n<section><\/section>\n<section><strong>Part 1\uff1a\u57fa\u7840\u8bbe\u7f6e\u548c\u5bfc\u5165<\/strong><\/section>\n<section><\/section>\n<section>\u9996\u5148\u662f\u5b89\u88c5\u5e76\u5bfc\u5165\u6240\u6709\u5fc5\u8981\u7684\u6a21\u5757\u3002\u4e0b\u9762\u662f\u5bfc\u5165\u5e93\u7684\u4e00\u6bb5\u4ee3\u7801\u622a\u56fe\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-83292f708f3333e18232309e5f42f7b1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-83292f708f3333e18232309e5f42f7b1.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u90e8\u5206\u4ee3\u7801\u622a\u56fe\u3002\u5b8c\u6574\u4ee3\u7801\u5757\u53c2\u89c1 GitHub\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u8fd0\u884c\u4e0a\u8ff0\u4ee3\u7801\uff08\u53c2\u8003\u9879\u76ee\u5b8c\u6574\u4ee3\u7801\uff09\uff0c\u53ef\u4ee5\u6267\u884c\u4ee5\u4e0b\u4efb\u52a1\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u8bbe\u7f6e\u968f\u673a\u79cd\u5b50\uff1aset_random_seed \u51fd\u6570\u901a\u8fc7\u4e3a Python \u7684\u968f\u673a\u6a21\u5757\u3001NumPy \u548c PyTorch \u8bbe\u7f6e\u79cd\u5b50\uff0c\u786e\u4fdd\u53ef\u590d\u73b0\u6027\uff1b<\/section>\n<\/li>\n<li>\n<section>\u73af\u5883\u53d8\u91cf\u914d\u7f6e\uff1a\u8bbe\u7f6e WANDB_API_KEY \u548c WANDB_PROJECT \u73af\u5883\u53d8\u91cf\uff0c\u4ee5\u542f\u7528\u4e0e Weights &amp; Biases \u7684\u5b9e\u9a8c\u8ddf\u8e2a\uff1b<\/section>\n<\/li>\n<li>\n<section>\u5bfc\u5165\u5fc5\u8981\u7684\u5e93\uff0c\u5305\u62ec random\u3001copy\u3001re\u3001torch \u7b49\u7b49\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>Part 2\uff1a\u6570\u636e\u683c\u5f0f\u4ee5\u53ca\u7b54\u6848\u63d0\u53d6<\/strong><\/section>\n<section><\/section>\n<section>\u63a5\u4e0b\u6765\uff0c\u9879\u76ee\u5b9a\u4e49\u4e86\u6570\u636e\u683c\u5f0f\uff0c\u4ee5\u53ca\u6a21\u578b\u5982\u4f55\u4ece\u8f93\u51fa\u548c\u6570\u636e\u96c6\u4e2d\u63d0\u53d6\u7b54\u6848\u6bb5\u843d\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u786e\u4fdd\u6a21\u578b\u8f93\u51fa\u683c\u5f0f\u4e00\u81f4\uff0c\u9879\u76ee\u8fd8\u5b9a\u4e49\u4e86\u4e00\u4e2a\u7cfb\u7edf\u63d0\u793a\u3002\u8be5\u63d0\u793a\u6307\u793a\u6a21\u578b\u751f\u6210\u5305\u542b &lt; reasoning &gt; \u548c &lt; answer &gt; \u6807\u7b7e\u7684\u8f93\u51fa\u3002\u8fd9\u4e00\u6b65\u901a\u8fc7\u4e24\u4e2a\u51fd\u6570\u5b8c\u6210\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>extract_answer_from_model_output\uff1a\u6b64\u51fd\u6570\u83b7\u53d6\u6a21\u578b\u7684\u8f93\u51fa\u6587\u672c\uff0c\u5e76\u63d0\u53d6 &lt; answer &gt; \u6807\u7b7e\u5185\u7684\u5185\u5bb9\uff1b<\/section>\n<\/li>\n<li>\n<section>extract_answer_from_dataset\uff1a\u6b64\u51fd\u6570\u4ece GSM8K \u6570\u636e\u96c6\u4e2d\u63d0\u53d6\u9884\u671f\u7b54\u6848\uff0c\u8be5\u6570\u636e\u96c6\u4f7f\u7528 \u201c####\u201d \u5206\u9694\u7b26\u6765\u5206\u9694\u7b54\u6848\uff1a<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-624276c72edb046dcd667497d81dc991.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-624276c72edb046dcd667497d81dc991.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u90e8\u5206\u4ee3\u7801\u622a\u56fe\u3002\u5b8c\u6574\u4ee3\u7801\u5757\u53c2\u89c1 GitHub\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Part 3\uff1a\u6570\u636e\u51c6\u5907<\/strong><\/section>\n<section><\/section>\n<section>\u8be5\u9879\u76ee\u4f7f\u7528 GSM8K \u6570\u636e\u96c6\u8fdb\u884c\u8bad\u7ec3\u3002\u9879\u76ee\u4f7f\u7528\u4e86\u8be5\u6570\u636e\u96c6\u4e2d\u7684\u793a\u4f8b\u6765\u8bad\u7ec3\u6a21\u578b\uff0c\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8bad\u7ec3\u8303\u5f0f\uff0c\u8ba9\u6a21\u578b\u751f\u6210\u591a\u4e2a\u95ee\u9898\u89e3\u7b54\u6837\u672c\uff0c\u4e4b\u540e\u4f5c\u8005\u5c06\u8fd9\u4e9b\u89e3\u7b54\u4e0e GSM8K \u793a\u4f8b\u4e2d\u7684\u6807\u51c6\u7b54\u6848\u8fdb\u884c\u5bf9\u6bd4\uff0c\u5982\u679c\u5339\u914d\uff0c\u5c31\u4e3a RL \u7b97\u6cd5\uff08GRPO\uff09\u63d0\u4f9b\u9ad8\u5956\u52b1\uff0c\u7136\u540e\u66f4\u65b0\u6a21\u578b\u6743\u91cd\uff0c\u4ee5\u589e\u52a0\u6a21\u578b\u4e0b\u6b21\u83b7\u5f97\u9ad8\u5956\u52b1\u7684\u53ef\u80fd\u6027\u3002<\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u8fc7\u7a0b\u662f\u8fd9\u6837\u7684\u3002\u9996\u5148\u4ece Hugging Face \u52a0\u8f7d\u6570\u636e\u96c6\uff0c\u7136\u540e\u683c\u5f0f\u5316\u6bcf\u4e2a\u793a\u4f8b\uff0c\u5305\u62ec\u7cfb\u7edf\u63d0\u793a\u548c\u7528\u6237\u63d0\u793a\u3002\u8fd9\u6bb5\u5b9e\u73b0\u4ee3\u7801\u4e2d\u8fd8\u5b9a\u4e49\u4e86\u4e24\u4e2a\u8f85\u52a9\u51fd\u6570\uff1aprepare_dataset \u4ee5\u53ca build_prompt\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f9573950ec5f46cdb303da39aa29a27c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f9573950ec5f46cdb303da39aa29a27c.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u90e8\u5206\u4ee3\u7801\u622a\u56fe\u3002\u5b8c\u6574\u4ee3\u7801\u5757\u53c2\u89c1 GitHub\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Part 4\uff1a\u8bc4\u4f30\u51fd\u6570<\/strong><\/section>\n<section><\/section>\n<section>\u8bc4\u4f30\u5bf9\u4e8e\u8ddf\u8e2a\u6a21\u578b\u7684\u8fdb\u5c55\u81f3\u5173\u91cd\u8981\u3002\u56e0\u6b64\u4f5c\u8005\u5b9a\u4e49\u4e86\u4e00\u4e9b\u51fd\u6570\uff0c\u4ece\u800c\u53ef\u4ee5\u5728\u4e00\u7ec4\u793a\u4f8b\u4e0a\u5bf9\u6a21\u578b\u8fdb\u884c\u8bc4\u4f30\u3002\u8be5\u9879\u76ee\u7684\u8bc4\u4f30\u51fd\u6570\u6267\u884c\u4ee5\u4e0b\u4efb\u52a1\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>token \u5316\u63d0\u793a\u5e76\u751f\u6210\u54cd\u5e94\uff1a\u6a21\u578b\u7684\u8f93\u51fa\u662f\u5728 token \u5316\u63d0\u793a\u7684\u57fa\u7840\u4e0a\u751f\u6210\u7684\u3002<\/section>\n<\/li>\n<li>\n<section>\u63d0\u53d6\u9884\u6d4b\u7b54\u6848\uff1a\u4ece\u751f\u6210\u7684\u54cd\u5e94\u4e2d\u63d0\u53d6\u7b54\u6848\u3002<\/section>\n<\/li>\n<li>\n<section>\u5c06\u9884\u6d4b\u7b54\u6848\u4e0e\u9884\u671f\u7b54\u6848\u8fdb\u884c\u6bd4\u8f83\uff1a\u8fd9\u79cd\u6bd4\u8f83\u662f\u901a\u8fc7\u7cbe\u786e\u5339\u914d\u4ee5\u53ca\u6570\u503c\u7b49\u4ef7\u68c0\u67e5\u6765\u5b8c\u6210\u7684\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5728\u8fd9\u6bb5\u4ee3\u7801\u4e2d\uff0c\u4e24\u4e2a\u8f85\u52a9\u51fd\u6570 _extract_last_number \u548c _extract_single_number \u88ab\u7528\u6765\u4ece\u6587\u672c\u4e2d\u63d0\u53d6\u6570\u5b57\u3002\u8bc4\u4f30\u51fd\u6570 evaluate_model \u4f7f\u7528\u8fd9\u4e9b\u8f85\u52a9\u51fd\u6570\u6765\u786e\u5b9a\u9884\u6d4b\u7b54\u6848\u662f\u5426\u6b63\u786e\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e6cdbe061d94ecde73f2e60b54ae3432.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e6cdbe061d94ecde73f2e60b54ae3432.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u90e8\u5206\u4ee3\u7801\u622a\u56fe\u3002\u5b8c\u6574\u4ee3\u7801\u5757\u53c2\u89c1 GitHub\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Part 5\uff1a\u5956\u52b1\u51fd\u6570<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u5956\u52b1\u51fd\u6570\u662f\u5fc5\u4e0d\u53ef\u7f3a\u7684\uff0c\u4f5c\u8005\u5b9a\u4e49\u4e86\u4e24\u4e2a\u5956\u52b1\u51fd\u6570\uff1a<\/section>\n<section><\/section>\n<section>correctness_reward\uff1a\u8fd9\u4e2a\u51fd\u6570\u6839\u636e\u751f\u6210\u7684\u7b54\u6848\u662f\u5426\u6b63\u786e\u6765\u5206\u914d\u5956\u52b1\u3002\u91c7\u7528\u4e24\u79cd\u65b9\u5f0f\uff1a\u7cbe\u786e\u7684\u5b57\u7b26\u4e32\u5339\u914d\u548c\u6570\u503c\u7b49\u4ef7\u68c0\u67e5\uff0c\u5c06\u6a21\u578b\u8f93\u51fa\u7684\u7b54\u6848\u4e0e\u9884\u671f\u7b54\u6848\u8fdb\u884c\u6bd4\u8f83\u3002\u5b8c\u5168\u5339\u914d\u4f1a\u83b7\u5f97\u66f4\u9ad8\u7684\u5956\u52b1\uff082.0\uff09\uff0c\u800c\u57fa\u4e8e\u6570\u503c\u7b49\u4ef7\u7684\u5339\u914d\u4f1a\u83b7\u5f97\u8f83\u5c0f\u7684\u5956\u52b1\uff081.5\uff09\u3002<\/section>\n<section><\/section>\n<section>format_reward\uff1a\u8fd9\u4e2a\u51fd\u6570\u9f13\u52b1\u6a21\u578b\u9075\u5faa\u6240\u9700\u7684\u7c7b\u4f3c XML \u7684\u8f93\u51fa\u683c\u5f0f\u3002\u5b83\u4e3a\u751f\u6210\u6587\u672c\u4e2d\u5b58\u5728 &lt; reasoning&gt;\u3001\u3001 \u548c  \u6807\u7b7e\u63d0\u4f9b\u5c0f\u989d\u5956\u52b1\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-cf60c8be76cab916317037a4a8188761.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-cf60c8be76cab916317037a4a8188761.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u90e8\u5206\u4ee3\u7801\u622a\u56fe\u3002\u5b8c\u6574\u4ee3\u7801\u5757\u53c2\u89c1 GitHub\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Part 6\uff1a\u4ece\u5934\u5f00\u59cb\u5b9e\u73b0 DataParallel GRPO<\/strong><\/section>\n<section><\/section>\n<section>\u8fd9\u4e00\u8282\uff0c\u6211\u4eec\u5c06\u4ece\u5934\u5b9e\u73b0 GRPO \u7b97\u6cd5\u7684\u6240\u6709\u6784\u5efa\u6a21\u5757\u3002\u9996\u5148\uff0c\u8fd9\u91cc\u5047\u8bbe\u8fd0\u884c\u4ee3\u7801\u7684\u673a\u5668\u81f3\u5c11\u6709 2 \u53f0 GPU\u3002\u4e3a\u6b64\uff0c\u8fd9\u91cc\u8981\u4f7f\u7528 PyTorch \u7684 DataParallel API \u6765\u5c06\u7b56\u7565\u6a21\u578b\u653e\u5728\u591a\u4e2a GPU \u6838\u5fc3\u4e0a\uff0c\u6bcf\u4e2a GPU \u6838\u5fc3\u90fd\u6709\u8be5\u6a21\u578b\u7684\u4e00\u4e2a\u526f\u672c\u3002\u7136\u540e\u5c06\u6279\u91cf\u6570\u636e\u5206\u6563\u5728\u8fd9\u4e9b GPU \u6838\u5fc3\u4e0a\u5b8c\u6210\u5904\u7406\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-6f964ceb1e32a178064020054fb57d16.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-6f964ceb1e32a178064020054fb57d16.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u90e8\u5206\u4ee3\u7801\u622a\u56fe\u3002\u5b8c\u6574\u4ee3\u7801\u5757\u53c2\u89c1 GitHub\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Part 7\uff1a\u8bad\u7ec3\u8bbe\u7f6e\u548c\u6267\u884c<\/strong><\/section>\n<section><\/section>\n<section>\u8fd9\u4e00\u8282\uff0c\u6211\u4eec\u5c06\u6240\u6709\u7ec4\u4ef6\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u5b8c\u6210\u8bbe\u7f6e\u5e76\u5f00\u59cb\u8bad\u7ec3\u3002<\/section>\n<section><\/section>\n<section>\u9996\u5148\uff0c\u52a0\u8f7d\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u548c tokenizer\uff0c\u51c6\u5907\u8bc4\u4f30\u6570\u636e\uff0c\u7136\u540e\u4f7f\u7528\u4e0a\u9762\u4ece\u5934\u5b9e\u73b0\u7684 train_with_grpo \u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u3002<\/section>\n<section><\/section>\n<section>\u5173\u952e\u6b65\u9aa4\u5305\u62ec\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u6a21\u578b\u548c tokenizer \u521d\u59cb\u5316\uff1a\u4f7f\u7528\u4f18\u5316\u8bbe\u7f6e\uff08\u4f7f\u7528 torch.bfloat16 \u548c FlashAttention2\uff09\u52a0\u8f7d\u6a21\u578b Qwen\/Qwen2.5-1.5B-Instruct\u3002tokenizer \u4e5f\u8981\u52a0\u8f7d\uff0c\u5176\u586b\u5145 token \u8bbe\u7f6e\u4e3a\u5e8f\u5217\u672b\u5c3e token\u3002\u4f7f\u7528 torch.bfloat16 \u52a0\u8f7d\u6a21\u578b\u4f1a\u5c06\u5176\u53c2\u6570\u8f6c\u6362\u4e3a\u6bcf\u4e2a\u6570\u503c\u4f7f\u7528 16 \u4f4d\u800c\u4e0d\u662f 32 \u4f4d\u7684\u5f62\u5f0f\uff0c\u8fd9\u53ef\u5c06\u6a21\u578b\u7684\u5185\u5b58\u4f7f\u7528\u91cf\u51cf\u5c11\u4e00\u534a\uff0c\u5e76\u4e14\u53ef\u52a0\u5feb\u5728\u73b0\u4ee3 GPU \u4e0a\u7684\u8bad\u7ec3\u901f\u5ea6\u3002<\/section>\n<\/li>\n<li>\n<section>\u521d\u6b65\u8bc4\u4f30\uff1a\u5728\u5fae\u8c03\u4e4b\u524d\uff0c\u6839\u636e\u51e0\u4e2a\u793a\u4f8b\u5bf9\u6a21\u578b\u8fdb\u884c\u8bc4\u4f30\uff0c\u4ee5\u786e\u5b9a\u57fa\u51c6\u6027\u80fd\u3002<\/section>\n<\/li>\n<li>\n<section>\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\uff1a\u4e3a\u4ece\u5934\u5f00\u59cb\u5b9e\u73b0 GRPO \u7684\u8bad\u7ec3\u51fd\u6570 train_with_grpo \u914d\u7f6e\u9002\u5f53\u7684\u8bad\u7ec3\u53c2\u6570\u548c\u5956\u52b1\u51fd\u6570\u3002\u7136\u540e\uff0c\u5728\u5269\u4f59\u7684\u8bad\u7ec3\u6570\u636e\u4e0a\u6267\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002<\/section>\n<\/li>\n<li>\n<section>\u6700\u7ec8\u8bc4\u4f30\u548c\u6a21\u578b\u4fdd\u5b58\uff1a\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u540e\uff0c\u518d\u6b21\u8bc4\u4f30\u6a21\u578b\uff0c\u5e76\u4fdd\u5b58\u6700\u7ec8\u6a21\u578b\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u4e0b\u9762\u7684\u4ee3\u7801\u4f1a\u6267\u884c\u4ee5\u4e0b\u529f\u80fd\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u786e\u5b9a\u8bbe\u5907\uff08\u5982\u679c\u6709 GPU \u5c31\u7528 GPU\uff0c\u5426\u5219\u5c31\u7528 CPU\uff09\u3002<\/section>\n<\/li>\n<li>\n<section>\u52a0\u8f7d\u9884\u8bad\u7ec3\u7248 Qwen2.5-1.5B-Instruct \u6a21\u578b\u548c tokenizer\u3002tokenizer \u7684 pad token \u8bbe\u7f6e\u4e3a eos_token\u3002<\/section>\n<\/li>\n<li>\n<section>\u4fdd\u7559\u4e00\u5c0f\u90e8\u5206\u6570\u636e\u96c6\u7528\u4e8e\u8bc4\u4f30\uff0c\u4ee5\u63d0\u4f9b\u57fa\u7ebf\u3002<\/section>\n<\/li>\n<li>\n<section>\u901a\u8fc7\u542f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9\u548c\u7981\u7528 KV \u7f13\u5b58\uff0c\u4f18\u5316\u6a21\u578b\u7684\u5185\u5b58\u6548\u7387\u3002<\/section>\n<\/li>\n<li>\n<section>\u6b65\u9aa4 1\uff1a\u5728\u5fae\u8c03\u4e4b\u524d\u8bc4\u4f30\u6a21\u578b\uff0c\u4ee5\u5efa\u7acb\u57fa\u7ebf\u51c6\u786e\u6027\u3002<\/section>\n<\/li>\n<li>\n<section>\u6b65\u9aa4 2\uff1a\u4f7f\u7528 train_with_grpo \u51fd\u6570\u548c\u6211\u4eec\u5b9a\u4e49\u7684\u5956\u52b1\u51fd\u6570\uff08format_reward \u548c correctness_reward\uff0c\u5408\u5e76\u4e3a combined_reward\uff09\u6267\u884c\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u3002\u8fd9\u91cc\u4f7f\u7528\u4e86\u591a\u53f0 GPU \u8bad\u7ec3\u6a21\u578b\u3002<\/section>\n<\/li>\n<li>\n<section>\u6b65\u9aa4 3\uff1a\u5c06\u6700\u7ec8\u7684\u5fae\u8c03\u6a21\u578b\u548c tokenizer \u4fdd\u5b58\u5230\u78c1\u76d8\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>GRPO \u8bad\u7ec3\u6d41\u7a0b\u4f7f\u7528\u7684\u8d85\u53c2\u6570\u5982\u4e0b\u3002<\/section>\n<section><\/section>\n<section><strong>\u8bad\u7ec3\u914d\u7f6e<\/strong><\/section>\n<section><\/section>\n<section>\u4ee5\u4e0b\u53c2\u6570\u8bbe\u5b9a\u4e86\u4f7f\u7528\u4e0a\u9762\u7684 GRPO \u7b97\u6cd5\u5b9e\u73b0\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u8fd0\u884c\u7684\u914d\u7f6e\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>num_iterations=1\uff1a\u4ece\u5f53\u524d\u7b56\u7565\u6a21\u578b\u521b\u5efa\u65b0\u53c2\u8003\u6a21\u578b\u7684\u5916\u90e8\u8fed\u4ee3\u6b21\u6570\u3002\u4e00\u6b21\u8fed\u4ee3\u662f\u6307\u5728\u6574\u4e2a\u6570\u636e\u96c6\u4e0a\u6267\u884c\u4e00\u6b21\u901a\u8fc7\u3002<\/section>\n<\/li>\n<li>\n<section>num_steps=500\uff1a\u8bad\u7ec3\u5faa\u73af\u5c06\u6267\u884c\u6700\u591a 500 \u4e2a\u6b65\u9aa4\uff0c\u6bcf\u4e2a\u6b65\u9aa4\u5904\u7406\u4e00\u6279\u6837\u672c\u3002<\/section>\n<\/li>\n<li>\n<section>batch_size=7\uff1a\u5728 8 \u53f0 GPU \u7684\u60c5\u51b5\u4e0b\uff0c\u6bcf\u4e2a\u6b65\u9aa4\u6bcf\u6279\u5904\u7406 7 \u4e2a\u6837\u672c\uff0c\u6bcf\u53f0 GPU \u4e0a\u653e\u7f6e 1 \u4e2a\u6837\u672c\u3002\u4f7f\u7528\u4e00\u4e2a GPU (0) \u88ab DataParallel \u7528\u4f5c\u4e3b\u8282\u70b9\u6765\u805a\u5408\u68af\u5ea6\u5e76\u6536\u96c6\u8f93\u51fa\u3002<\/section>\n<\/li>\n<li>\n<section>num_generations=14\uff1a\u5bf9\u4e8e\u8bad\u7ec3\u6570\u636e\u4e2d\u7684\u6bcf\u4e2a\u63d0\u793a\u8bcd\uff0c\u8bad\u7ec3\u5668\u5c06\u751f\u6210 14 \u4e2a\u4e0d\u540c\u7684\u5b8c\u6210\u7ed3\u679c\u3002\u8fd9\u4e9b\u751f\u6210\u7ed3\u679c\u5c06\u88ab\u7528\u4e8e\u8ba1\u7b97\u6307\u5bfc\u5f3a\u5316\u5b66\u4e60\u66f4\u65b0\u7684\u76f8\u5bf9\u4f18\u52bf\uff08\u6216\u5956\u52b1\u4fe1\u53f7\uff09\u3002\u5982\u679c\u4f60\u7684 GPU \u7684 VRAM \u8f83\u5c11\uff0c\u8bf7\u51cf\u5c11\u6b64\u6570\u5b57\u3002<\/section>\n<\/li>\n<li>\n<section>max_completion_length=400\uff1a\u5728\u751f\u6210\u5b8c\u6210\u7ed3\u679c\uff08\u5e8f\u5217\u7684 response \u90e8\u5206\uff09\u65f6\uff0c\u751f\u6210\u4e0a\u9650\u4e3a 400 \u4e2a token\u3002\u8fd9\u9650\u5236\u4e86\u6a21\u578b\u5728 RL \u9636\u6bb5\u751f\u6210\u7684\u8f93\u51fa\u7684\u957f\u5ea6\u3002\u5982\u679c\u4f60\u7684 GPU \u7684 VRAM \u8f83\u5c11\uff0c\u8bf7\u51cf\u5c11\u6b64\u6570\u5b57\u3002<\/section>\n<\/li>\n<li>\n<section>beta=0.04\uff1aGRPO \u635f\u5931\u51fd\u6570\u4e2d KL \u6563\u5ea6\u60e9\u7f5a\u7684\u7cfb\u6570\u3002\u8fd9\u63a7\u5236\u7684\u662f\u6a21\u578b\u4e0e\u53c2\u8003\u6a21\u578b\u7684\u504f\u5dee\u7a0b\u5ea6\u3002<\/section>\n<\/li>\n<li>\n<section>learning_rate=5e-6\uff1aRL \u5fae\u8c03\u7684\u5b66\u4e60\u7387\u3002\u4e3a\u4e86\u5b9e\u73b0\u7a33\u5b9a\u7684\u7b56\u7565\u66f4\u65b0\uff0c\u8fd9\u91cc\u4f7f\u7528\u4e86\u76f8\u5bf9\u8f83\u4f4e\u7684\u5b66\u4e60\u7387\u3002<\/section>\n<\/li>\n<li>\n<section>mu=1\uff1a\u5bf9\u6bcf\u6279 rollout \u6570\u636e\u6267\u884c\u7684\u7b56\u7565\u66f4\u65b0\u6b21\u6570\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u6bcf\u6279\u53ea\u6267\u884c\u4e00\u6b21\u66f4\u65b0\u3002<\/section>\n<\/li>\n<li>\n<section>epsilon=0.1\uff1aGRPO \u7684 PPO \u7ec4\u4ef6\u7684 clipping \u53c2\u6570\u3002\u8fd9\u53ef\u4ee5\u9632\u6b62\u7b56\u7565\u5728\u5355\u6b21\u66f4\u65b0\u4e2d\u53d1\u751f\u592a\u5927\u7684\u53d8\u5316\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5728\u5fae\u8c03\u4e4b\u524d\u548c\u4e4b\u540e\u90fd\u4f1a\u5bf9\u6a21\u578b\u8fdb\u884c\u8bc4\u4f30\uff0c\u4ee5\u8861\u91cf\u51c6\u786e\u7387\u7684\u63d0\u9ad8\u60c5\u51b5\u3002\u6700\u540e\uff0c\u5c06\u5fae\u8c03\u540e\u7684\u6a21\u578b\u4fdd\u5b58\u5230 grpo_finetuned_model \u76ee\u5f55\u4e2d\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8f9034d89d2c604724f90c2b3cb5164a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8f9034d89d2c604724f90c2b3cb5164a.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u90e8\u5206\u4ee3\u7801\u622a\u56fe\u3002\u5b8c\u6574\u4ee3\u7801\u5757\u53c2\u89c1 GitHub\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6559\u7a0b\u4e2d\u8fd8\u7ed9\u51fa\u4e86\u8be6\u7ec6\u7684\u6267\u884c\u60c5\u51b5\uff0c\u53ef\u4f5c\u53c2\u8003\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-88b74b2f122eb928bd290dadc67a9a8a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-88b74b2f122eb928bd290dadc67a9a8a.png\"><\/a><\/section>\n<section>\u4e0b\u9762\u6211\u4eec\u4e5f\u7b80\u5355\u770b\u770b\u5176\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section>\u9996\u5148\uff0c\u521d\u59cb\u914d\u7f6e\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u8fd0\u884c GRPO \u4e4b\u524d\u7684\u51c6\u786e\u5ea6\u4e3a 23.33%\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f96adfa1188bc18585d4684c1fe32b36.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f96adfa1188bc18585d4684c1fe32b36.png\"><\/a><\/section>\n<section>\u7136\u540e\u7ecf\u8fc7 500 \u6b65\u7684 1 \u8f6e GRPO \u8fed\u4ee3\uff0c\u4e0b\u56fe\u5c55\u793a\u4e86\u76f8\u5173\u7684\u8bad\u7ec3\u52a8\u6001\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-b06727b5465df01178a1b40ce053fac2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-b06727b5465df01178a1b40ce053fac2.png\"><\/a><\/section>\n<section>\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u81ea\u7136\u8fd8\u9700\u8981\u5bf9\u6a21\u578b\u8fdb\u884c\u65b0\u4e00\u8f6e\u7684\u8bc4\u4f30\u3002\u8fd9\u91cc\u91c7\u7528\u4e86 30 \u4e2a\u8bc4\u4f30\u6837\u672c\u6765\u8fdb\u884c\u8bc4\u4f30\uff0c\u4ee5\u4e0b\u5c55\u793a\u4e86\u5176\u4e2d\u4e00\u4e2a\u6a21\u578b\u56de\u7b54\u6b63\u786e\u7684\u793a\u4f8b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-01c7e49bbc7c31bd4203ab34a859d12e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-01c7e49bbc7c31bd4203ab34a859d12e.png\"><\/a><\/section>\n<section>\u6574\u4f53\u8868\u73b0\u5982\u4f55\u5462\uff1f\u53ef\u4ee5\u770b\u5230\uff0c\u7ecf\u8fc7\u4e00\u8f6e GRPO \u4e4b\u540e\uff0cQwen-2.5-1.5B-Instruct \u6a21\u578b\u7b54\u5bf9\u4e86 30 \u95ee\u9898\u4e2d\u7684 27 \u9898\uff0c\u5b9e\u73b0\u4e86 90% \u7684\u51c6\u786e\u5ea6\u3002\u76f8\u8f83\u4e8e GRPO \u4e4b\u524d\u7684 23.33%\uff0c\u53ef\u8bf4\u662f\u5b9e\u73b0\u4e86\u6027\u80fd\u98de\u8dc3\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-3a6fd07674e2043ed54250d593f6f07f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-3a6fd07674e2043ed54250d593f6f07f.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-21481d0e8d5a7091962d8d96170e156d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-21481d0e8d5a7091962d8d96170e156d.png\"><\/a><\/section>\n<section>\u4e0a\u9762\u4e24\u5f20\u56fe\u5c55\u793a\u4e86\u6a21\u578b\u7684\u5b66\u4e60\u8fc7\u7a0b\u52a8\u6001\uff0c\u53ef\u4ee5\u770b\u5230\uff1a\u5e73\u5747\u5956\u52b1\u5728 2.25 \u5de6\u53f3\u5c31\u8d8b\u4e8e\u7a33\u5b9a\u4e86\uff08\u7406\u8bba\u6700\u5927\u503c\u4e3a 0.8 + 2.0 = 2.8\uff09\u3002\u76f8\u6bd4\u4e8e\u53e6\u4e00\u5904\u5fae\u8c03\u7684 Qwen-2.5-0.5B-Instruct\uff08\u83b7\u5f97\u7684\u5e73\u5747\u5956\u52b1\u4e3a 1.4\uff09\uff0c\u8fd9\u4e2a\u6570\u5b57\u76f8\u5f53\u9ad8\u4e86\uff0c\u53c2\u9605\uff1ahttps:\/\/github.com\/aburkov\/theLMbook\/blob\/main\/GRPO_Qwen_0_5_Instruct.ipynb<\/section>\n<section><\/section>\n<section>\u5982\u679c\u4f7f\u7528\u66f4\u5927\u7684\u6a21\u578b\u5e76\u5141\u8bb8\u66f4\u957f\u7684\u751f\u6210\u65f6\u95f4\uff0c\u6a21\u578b\u6b63\u786e\u89e3\u7b54\u95ee\u9898\u7684\u80fd\u529b\u8fd8\u5c06\u8fdb\u4e00\u6b65\u63d0\u5347\u3002\u4f46\u662f\uff0c\u5982\u679c\u8981\u8bad\u7ec3\u66f4\u5927\u7684\u6a21\u578b\uff0c\u4e0d\u4ec5\u9700\u8981\u5c06\u6570\u636e\u5206\u5e03\u5728\u591a\u53f0 GPU \u4e0a\uff0c\u8fd8\u9700\u8981\u5c06\u6a21\u578b\u5206\u5f00\u653e\u5728\u591a\u53f0 GPU \u4e0a\uff0c\u8fd9\u9700\u8981\u7528\u5230 DeepSpeed \u6216 FSDP\uff08\u5b8c\u5168\u5206\u7247\u6570\u636e\u5e76\u884c\uff09\u7b49\u6a21\u578b\u5e76\u884c\u5de5\u5177\u3002<\/section>\n<section><\/section>\n<section>\u4e0b\u9762\u52a0\u8f7d\u548c\u6d4b\u8bd5\u5df2\u7ecf\u5fae\u8c03\u7684\u6a21\u578b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ddd929ed4ce93bf0be7d876394d3dc83.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ddd929ed4ce93bf0be7d876394d3dc83.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u5b8c\u6574\u4ee3\u7801\u89c1\u539f\u7b14\u8bb0\u672c<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u52a0\u8f7d\u5b8c\u6210\u540e\u6d4b\u8bd5\u4e00\u4e0b\uff0c\u9996\u5148\u95ee\u95ee 1+1 \u7b49\u4e8e\u51e0\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d5489363fa286ad8b7144e2ec912f613.gif\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d5489363fa286ad8b7144e2ec912f613.gif\"><\/a><\/section>\n<section>\u53ef\u4ee5\u770b\u5230\uff0c\u6a21\u578b\u53cd\u590d\u601d\u8003\u4e86\u5f88\u591a\u6b21\uff0c\u7ec8\u4e8e\u8ba4\u5b9a\u786e\u5b9e\u7b49\u4e8e 2\u3002<\/section>\n<section><\/section>\n<section>\u591a\u6b21\u6d4b\u8bd5\u540e\u8fd8\u53ef\u4ee5\u53d1\u73b0\uff0c\u8be5\u6a21\u578b\u6ca1\u6709\u5b66\u4f1a\u751f\u6210\u5e8f\u5217\u7ed3\u675f\uff08EOS\uff09token\uff0c\u56e0\u6b64\u5373\u4f7f\u5728  token \u4e4b\u540e\uff0c\u8f93\u51fa\u5e8f\u5217\u4ecd\u4f1a\u7ee7\u7eed\u3002\u8fd9\u662f\u9884\u671f\u7684\u884c\u4e3a\uff0c\u56e0\u4e3a\u6211\u4eec\u4f7f\u7528\u7684\u5956\u52b1\u51fd\u6570\u4e2d\u6ca1\u6709\u5305\u542b\u4e00\u4e2a\u7528\u4e8e\u505c\u6b62\u751f\u6210\u7684\u5956\u52b1\u3002\u6211\u4eec\u4e5f\u6ca1\u6709\u6267\u884c\u76d1\u7763\u5fae\u8c03\u6b65\u9aa4 \u2014\u2014 \u8be5\u6b65\u9aa4\u53ef\u4ee5\u8ba9\u6a21\u578b\u5b66\u4f1a\u5728  \u4e4b\u540e\u7acb\u5373\u751f\u6210 EOS\u3002<\/section>\n<section><\/section>\n<section>\u4f60\u5bf9\u8fd9\u7bc7\u4ee3\u7801\u5bc6\u96c6\u7684\u6559\u7a0b\u600e\u4e48\u770b\uff1f\u6709\u6ca1\u6709\u8ba9\u4f60\u4ea7\u751f\u5728\u81ea\u5df1\u7684\u7535\u8111\u4e0a\u5b9e\u73b0 GRPO \u7684\u60f3\u6cd5\uff1f<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-02-3\" target=\"_blank\">DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:DeepSeek\u5173\u952eR [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[52,55],"class_list":["post-39150","post","type-post","status-publish","format-standard","hentry","category-news","tag-numpy","tag-55"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=39150\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=39150\",\"url\":\"https:\/\/17aitech.com\/?p=39150\",\"name\":\"DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=39150#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=39150#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png\",\"datePublished\":\"2025-03-02T16:01:38+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=39150#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=39150\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=39150#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=39150#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=39150","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=39150","url":"https:\/\/17aitech.com\/?p=39150","name":"DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=39150#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=39150#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png","datePublished":"2025-03-02T16:01:38+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=39150#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=39150"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=39150#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1658e8512f29c20f2367ff87491c0020.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=39150#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"DeepSeek\u5173\u952eRL\u7b97\u6cd5GRPO\uff0c\u6709\u4eba\u4ece\u5934\u8dd1\u901a\u4e86\uff0c\u8d21\u732e\u5b8c\u6574\u4ee3\u7801"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39150","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=39150"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39150\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39150"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39150"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39150"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}