{"id":27311,"date":"2025-03-02T15:48:48","date_gmt":"2025-03-02T07:48:48","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=27311"},"modified":"2025-03-02T15:48:48","modified_gmt":"2025-03-02T07:48:48","slug":"weiruanyuanwizardlmtuan","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/weiruanyuanwizardlmtuan\/","title":{"rendered":"\u5fae\u8f6f\u539fWizardLM\u56e2\u961f\uff1a\u4ee3\u7801\u5927\u6a21\u578bWarriorCoder\uff0c\u6027\u80fd\u65b0SOTA"},"content":{"rendered":"<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/0f593908debb7a2.png\" \/><\/p>\n<p>&nbsp;<\/p>\n<ul>\n<li>\u8bba\u6587\u6807\u9898\uff1aWarriorCoder: Learning from Expert Battles to Augment Code Large Language Models<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2412.17395<\/li>\n<\/ul>\n<h2>\n01 \u80cc\u666f<\/h2>\n<p>\u8fd1\u5e74\u6765\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u4ee3\u7801\u76f8\u5173\u7684\u4efb\u52a1\u4e0a\u5c55\u73b0\u4e86\u60ca\u4eba\u7684\u8868\u73b0\uff0c\u5404\u79cd\u4ee3\u7801\u5927\u6a21\u578b\u5c42\u51fa\u4e0d\u7a77\u3002\u8fd9\u4e9b\u6210\u529f\u7684\u6848\u4f8b\u8868\u660e\uff0c\u5728\u5927\u89c4\u6a21\u4ee3\u7801\u6570\u636e\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u6838\u5fc3\u7f16\u7a0b\u80fd\u529b\u3002\u9664\u4e86\u9884\u8bad\u7ec3\u5916\uff0c\u4e00\u4e9b\u901a\u8fc7 instruction \u6570\u636e\u5bf9 LLM \u8fdb\u884c post-training \u7684\u65b9\u6cd5\uff0c\u4e5f\u4f7f\u5f97\u6a21\u578b\u5728\u5bf9\u6307\u4ee4\u7684\u7406\u89e3\u548c\u56de\u7b54\u7684\u8d28\u91cf\u7b49\u65b9\u9762\u53d6\u5f97\u4e86\u663e\u8457\u63d0\u9ad8\u3002\u7136\u800c\uff0cpost-training \u7684\u6548\u679c\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4f9d\u8d56\u4e8e\u53ef\u7528\u7684\u9ad8\u8d28\u91cf\u6570\u636e\uff0c\u4f46\u662f\u6570\u636e\u7684\u6536\u96c6\u548c\u6ce8\u91ca\u5b58\u5728\u7740\u4e0d\u5c0f\u7684\u6311\u6218\u3002<br \/>\n\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u6311\u6218\uff0c\u4e00\u4e9b\u65b9\u6cd5\u8bbe\u8ba1\u4e86\u5404\u79cd\u6570\u636e\u98de\u8f6e\u6765\u751f\u6210 instruction \u6570\u636e\uff0c\u5982 Self-Instruct\uff0cEvol-Instruct \u7b49\u3002\u8fd9\u4e9b\u65b9\u6cd5\u901a\u8fc7\u591a\u79cd\u6570\u636e\u589e\u5f3a\u624b\u6bb5\u6765\u6784\u5efa instruction \u6570\u636e\uff0c\u5728\u8fd9\u4e9b\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\u53ef\u4ee5\u6709\u6548\u63d0\u5347\u6a21\u578b\u7684\u4ee3\u7801\u751f\u6210\u80fd\u529b\u3002\u7136\u800c\uff0c\u5982\u56fe 1 \u6240\u793a\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u4ecd\u7136\u4f9d\u8d56\u4e8e\u5bf9\u73b0\u5b58\u6570\u636e\u96c6\u8fdb\u884c\u6269\u5c55\u5e76\u9700\u8981\u8c03\u7528\u79c1\u6709 LLM\uff08\u5982 GPT-3.5\u3001GPT-4 \u7b49\uff09\uff0c\u4f7f\u5f97\u6570\u636e\u6536\u96c6\u6210\u672c\u8f83\u9ad8\u3002\u6b64\u5916\uff0c\u6709\u9650\u7684\u6570\u636e\u6765\u6e90\u548c\u7528\u4e8e\u6ce8\u91ca\u7684 LLM \u4e5f\u9650\u5236\u4e86\u6570\u636e\u7684\u591a\u6837\u6027\uff0c\u5e76\u7ee7\u627f\u4e86\u6709\u9650\u7684\u79c1\u6709 LLM \u672c\u8eab\u56fa\u6709\u7684\u7cfb\u7edf\u504f\u89c1\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/e4f49f6a8034802.png\" \/> <em>\u56fe 1<\/em><\/p>\n<p>\u672c\u6587\u63d0\u51fa\u4e86 <strong>WarriorCoder\uff0c\u4e00\u79cd\u5168\u65b0\u7684\u4ee3\u7801\u5927\u6a21\u578b\u7684\u6570\u636e\u98de\u8f6e\u8bad\u7ec3\u8303\u5f0f\uff0c\u6a21\u578b\u901a\u8fc7\u5b66\u4e60\u4e13\u5bb6\u5bf9\u6297\u7684\u65b9\u5f0f\u6765\u96c6\u6210\u5404\u4e2a\u4ee3\u7801\u4e13\u5bb6\u5927\u6a21\u578b\u7684\u4f18\u70b9<\/strong>\u3002\u5982\u56fe 1 \u6240\u793a\uff0c\u5404\u4e2a\u4ee3\u7801\u4e13\u5bb6\u5927\u6a21\u578b\u4e24\u4e24\u5bf9\u6218\uff0c\u653b\u51fb\u8005\u5728\u5176\u81ea\u8eab\u7684\u4e13\u4e1a\u9886\u57df\u5185\u6311\u6218\u5bf9\u624b\uff0c\u76ee\u6807\u6a21\u578b\u5219\u5411\u8fd9\u4e9b\u5bf9\u6218\u4e2d\u7684\u80dc\u8005\u5b66\u4e60\u3002\u4e0e\u4e4b\u524d\u7684\u65b9\u6cd5\u4e0d\u540c\uff0c\u4e4b\u524d\u7684\u65b9\u6cd5\u5927\u591a\u90fd\u4f9d\u8d56\u73b0\u6709\u5f00\u6e90\u6570\u636e\u96c6\uff0c\u5c06\u8fd9\u4e9b\u6570\u636e\u96c6\u505a\u4e3a\u79cd\u5b50\u6570\u636e\u53bb\u5408\u6210\u548c\u589e\u5f3a\uff0c\u800c warriorCoder \u662f\u4ece 0 \u5230 1 \u7684\u751f\u6210\u6570\u636e\u4e0d\u9700\u8981\u79cd\u5b50\u6570\u636e\uff0c\u5e76\u4e14\u8be5\u65b9\u6cd5\u53ef\u4ee5\u878d\u5408\u591a\u4e2a\u4ee3\u7801\u4e13\u5bb6\u5927\u6a21\u578b\u7684\u4f18\u52bf\uff0c\u800c\u4e0d\u662f\u4ec5\u4ec5\u84b8\u998f\u4e2a\u522b\u6a21\u578b\u7684\u4f18\u52bf\u3002\u6b64\u5916\uff0c\u672c\u6587\u63d0\u51fa\u7684\u65b9\u6cd5\u6d88\u9664\u4e86\u5728\u6570\u636e\u6536\u96c6\u8fc7\u7a0b\u4e2d\u5bf9\u4eba\u5de5\u53c2\u4e0e\u548c\u79c1\u6709 LLM \u7684\u4f9d\u8d56\uff0c\u53ef\u4ee5\u4ee5\u6781\u4f4e\u6210\u672c\u6536\u96c6\u9ad8\u8d28\u91cf\u3001\u591a\u6837\u5316\u7684\u8bad\u7ec3\u6570\u636e\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cwarriorCoder \u4e0d\u4ec5\u4ec5\u662f\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e2d\u8fbe\u5230\u4e86\u5f53\u524d\u7684 SOTA\uff0c\u8fd8\u5728 code reasoning \u548c libraries using \u7b49 benchmark \u4e0a\u4e5f\u53d6\u5f97\u4e86\u5353\u8d8a\u7684\u6210\u7ee9\uff0c\u53ef\u8c13\u662f\u4ee3\u7801\u516d\u8fb9\u5f62\u6218\u58eb\u3002<\/p>\n<h2>\n02 \u65b9\u6cd5<\/h2>\n<p>\u672c\u6587\u6784\u5efa\u4e86\u4e00\u4e2a\u4ee3\u7801\u5927\u6a21\u578b\u7684\u7ade\u6280\u573a\u3002\u5728\u8fd9\u91cc\uff0c\u6700\u5148\u8fdb\u7684\u4ee3\u7801\u4e13\u5bb6\u5927\u6a21\u578b\u76f8\u4e92\u5bf9\u6297\uff0c\u6bcf\u4e2a\u6a21\u578b\u5229\u7528\u5176\u5df2\u7ecf\u638c\u63e1\u7684\u77e5\u8bc6\u6311\u6218\u5176\u4ed6\u6a21\u578b\uff0c\u800c\u5176\u4f59\u6a21\u578b\u5219\u62c5\u4efb\u88c1\u5224\u8bc4\u4f30\u5bf9\u6297\u7ed3\u679c\u3002\u76ee\u6807\u6a21\u578b\u968f\u540e\u4ece\u8fd9\u4e9b\u5bf9\u6297\u4e2d\u7684\u80dc\u8005\u5b66\u4e60\uff0c\u9010\u6b65\u6574\u5408\u6240\u6709\u7ade\u4e89\u8005\u7684\u4f18\u52bf\u3002\u672c\u6587\u5c06\u53c2\u8d5b\u8005\uff08\u4ee3\u7801\u4e13\u5bb6\u5927\u6a21\u578b\uff09\u89c6\u4e3a\u4e00\u4e2a\u7ec4\uff0c\u901a\u8fc7\u7ec4\u5185\u76f8\u5bf9\u4f18\u52bf\u7b54\u6848\u6765\u4f18\u5316\u6a21\u578b\uff0c\u8fd9\u4e00\u70b9\u4e0e <a href=\"https:\/\/www.kdjingpai.com\/grpo-ruhezaishi\/\">GRPO<\/a> \u6709\u7740\u5f02\u66f2\u540c\u5de5\u4e4b\u5999\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/fa517c2792c386f.png\" \/> <em>\u56fe 2<\/em><\/p>\n<p><strong>2.1 Competitors Setting<\/strong><\/p>\n<p>\u53c2\u8d5b\u8005\u7684\u80fd\u529b\u51b3\u5b9a\u4e86 WarriorCoder \u7684\u6700\u7ec8\u8868\u73b0\u3002\u7406\u8bba\u4e0a\uff0c\u4ece\u66f4\u5927\u3001\u66f4\u5f3a\u7684\u53c2\u8d5b\u8005\u6c60\u4e2d\u83b7\u5f97\u7684\u8bad\u7ec3\u6570\u636e\u591a\u6837\u6027\u66f4\u5f3a\u3001\u8d28\u91cf\u66f4\u9ad8\uff0c\u6700\u7ec8\u8bad\u7ec3\u51fa\u6765\u7684\u6a21\u578b\u7684\u6027\u80fd\u4e5f\u5c31\u8d8a\u597d\u3002\u5728\u6bcf\u4e00\u8f6e\u7ade\u6280\u573a\u4e2d\uff0c\u53ea\u6709\u4e00\u5bf9\u4ee3\u7801\u4e13\u5bb6\u88ab\u9009\u4e3a\u7ade\u4e89\u8005\uff0c\u5176\u4ed6\u7684\u5219\u4f5c\u4e3a\u88c1\u5224\u3002\u672c\u6587\u4ece BigCodeBench \u6392\u884c\u699c\u4e2d\u9009\u53d6\u4e86\u4e94\u4e2a 75B \u4ee5\u5185\u7684\u5148\u8fdb\u5927\u6a21\u578b \u2014\u2014Athene-V2-Chat\u3001DeepSeek-Coder-V2-Lite-Instruct\u3001Llama-3.3-70B-Instruct\u3001Qwen2.5-72B-Instruct \u548c QwQ-32B-Preview\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e94\u4e2a\u5927\u6a21\u578b\u5747\u4e3a\u5f00\u6e90\u5927\u6a21\u578b\uff0cWarriorCoder \u4ec5\u57fa\u4e8e\u8fd9\u4e9b\u5f00\u6e90\u5927\u6a21\u578b\u7684\u5bf9\u6297\u5c31\u5f97\u5230\u4e86\u4f18\u5f02\u7684\u6027\u80fd\u3002\u5f53\u7136\uff0cWarriorCoder \u4e5f\u80fd\u591f\u4ece\u5f3a\u5927\u7684\u79c1\u6709\u5927\u6a21\u578b\u4e2d\u5b66\u4e60\u3002<\/p>\n<p><strong>2.2 Instruction Mining from Scratch<\/strong><\/p>\n<p>\u5bf9\u4e8e\u4e00\u5bf9\u5bf9\u624b \u2014\u2014A \u548c B\uff08\u5176\u4e2d A \u4e3a\u653b\u51fb\u8005\uff0cB \u4e3a\u9632\u5b88\u8005\uff09\uff0c\u5bf9\u6297\u7684\u7b2c\u4e00\u6b65\u662f\u5728 A \u64c5\u957f\u7684\u9886\u57df\u6311\u6218 B\uff0c\u8fd9\u5c31\u9700\u8981\u4e86\u89e3 A \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5b66\u5230\u4e86\u4ec0\u4e48\u3002\u7136\u800c\uff0c\u51e0\u4e4e\u5f53\u524d\u6240\u6709\u7684\u5f00\u6e90\u5927\u6a21\u578b\u90fd\u672a\u516c\u5e03\u5176\u6838\u5fc3\u8bad\u7ec3\u6570\u636e\uff0c\u8fd9\u4f7f\u5f97\u653b\u51fb\u8005\u64c5\u957f\u7684\u77e5\u8bc6\u53d8\u5f97\u6781\u4e3a\u56f0\u96be\u3002\u53d7 Magpie \u7684\u542f\u53d1\uff0c\u672c\u6587\u8bbe\u8ba1\u4e86\u4e00\u79cd\u57fa\u4e8e\u5bf9\u8bdd\u8865\u5168\u7684\u65b9\u5f0f\u6765\u6316\u6398\u5927\u6a21\u578b\u5df2\u638c\u63e1\u7684\u80fd\u529b\u3002\u4ee5 Qwen2.5 \u4e3a\u4f8b\uff0c\u5982\u679c\u8981\u5176\u751f\u6210\u4e00\u4e2a\u5feb\u901f\u6392\u5e8f\u7b97\u6cd5\uff0c\u5219\u5b8c\u6574\u7684 prompt \u683c\u5f0f\u5982\u56fe 3 \u6240\u793a\u3002Prompt \u5e94\u5305\u62ec system content\u3001user content \u4ee5\u53ca\u4e0e\u683c\u5f0f\u6709\u5173\u7684\u7279\u6b8a token\uff0c\u5982 \u201c&lt;|im_start|&gt;\u201d\u3001\u201c&lt;|im_end|&gt;\u201d \u7b49\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/01f0dd785ca53ed.png\" \/> <em>\u56fe 3<\/em><\/p>\n<p>\u800c\u5982\u679c\u4ec5\u5c06\u524d\u7f00\u90e8\u5206\uff08\u672c\u8eab\u65e0\u4efb\u4f55\u5177\u4f53\u610f\u4e49\uff0c\u5982\u56fe 4 \u6240\u793a\uff09\u8f93\u5165\u6a21\u578b\uff0c\u5229\u7528\u6a21\u578b\u7684\u8865\u5168\u80fd\u529b\u5c31\u53ef\u4ee5\u5f97\u5230\u7528\u6237\u6307\u4ee4\uff08user content\uff09\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/9b44546e8f1adf4.png\" \/> <em>\u56fe 4<\/em><\/p>\n<p>\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u5728\u4e0d\u540c\u7684\u751f\u6210\u53c2\u6570\u914d\u7f6e\u4e0b\uff08\u4f8b\u5982\u4e0d\u540c\u7684\u6e29\u5ea6\u503c\u548c top-p \u503c\uff09\u5c31\u53ef\u4ee5\u6536\u96c6\u5230\u6a21\u578b\u5df2\u7ecf\u5b66\u4e60\u5230\u7684 instruction \u6570\u636e\u3002\u4e0e\u4f20\u7edf\u7684\u6570\u636e\u5408\u6210\u4e0d\u540c\uff0c\u672c\u6587\u6536\u96c6\u7684 instruction \u6570\u636e\u4e0d\u662f\u7531\u6a21\u578b\u5408\u6210\u7684\uff0c\u800c\u662f\u76f4\u63a5\u4ece\u6a21\u578b\u7684\u5206\u5e03\u4e2d\u8fdb\u884c\u91c7\u6837\u5f97\u5230\u7684\uff0c\u8fd9\u907f\u514d\u4e86\u6a21\u5f0f\u8fc7\u62df\u5408\u3001\u8f93\u51fa\u5206\u5e03\u504f\u79fb\u7b49\u95ee\u9898\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u6307\u4ee4\u53ef\u80fd\u4f1a\u91cd\u590d\u3001\u6709\u6b67\u4e49\u3001\u4e0d\u6e05\u6670\u6216\u8fc7\u4e8e\u7b80\u5355\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u6211\u4eec\u5bf9\u6570\u636e\u8fdb\u884c\u53bb\u91cd\uff0c\u5e76\u91c7\u7528\u88c1\u5224\u6a21\u578b\u6765\u8bc4\u4f30\u5176\u96be\u5ea6\u3002\u672c\u6587\u5c06\u96be\u5ea6\u5206\u4e3a\u56db\u4e2a\u7b49\u7ea7\uff1aExcellent\u3001Good\u3001Average\u3001Poor\u3002\u6700\u7ec8\u4ec5\u4f7f\u7528 Excellent \u548c Good \u4e24\u4e2a\u7b49\u7ea7\u7684\u6307\u4ee4\uff0c\u5e76\u4f7f\u7528 KcenterGreedy \u7b97\u6cd5\u5bf9 instruction \u6570\u636e\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u538b\u7f29\u3002<\/p>\n<p><strong>2.3 Win-Loss Decision<\/strong><\/p>\n<p>\u6311\u6218\u8005\u548c\u9632\u5fa1\u8005\u90fd\u8981\u6839\u636e instruction \u6570\u636e\u751f\u6210\u56de\u7b54\uff0c\u5e76\u7531\u88c1\u5224\uff08\u5269\u4f59\u7684\u6a21\u578b\uff09\u6295\u7968\u51b3\u5b9a\u8f93\u8d62\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d9af9bc6daf21e4.png\" \/> \u7136\u800c\uff0c\u4ec5\u4f9d\u8d56\u4e8e textit {\u5c40\u90e8\u5f97\u5206} \u6765\u9009\u62e9\u83b7\u80dc\u8005\u53ef\u80fd\u4f1a\u5e26\u6765\u5076\u7136\u6027\u95ee\u9898\u3002\u7531\u4e8e\u6295\u7968\u4f1a\u53d7\u5230\u968f\u673a\u6027\u6216\u8bc4\u5ba1\u8005\u504f\u89c1\u7b49\u56e0\u7d20\u7684\u5f71\u54cd\uff0c\u5728\u67d0\u4e9b\u6307\u4ee4\u4e0b\u8f83\u5f31\u7684\u6a21\u578b\u53ef\u80fd\u4f1a\u6bd4\u8f83\u5f3a\u7684\u6a21\u578b\u83b7\u5f97\u66f4\u591a\u7684\u6295\u7968\uff0c\u5373\u4fbf\u5176\u56de\u7b54\u5e76\u6ca1\u6709\u771f\u6b63\u6bd4\u8f83\u5f3a\u7684\u6a21\u578b\u66f4\u597d\u3002<br \/>\n\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u672c\u6587\u5728\u51b3\u7b56\u8fc7\u7a0b\u4e2d\u540c\u65f6\u8003\u8651\u5c40\u90e8\u5076\u7136\u6027\u548c\u5168\u5c40\u4e00\u81f4\u6027\u3002\u672c\u6587\u5f15\u5165\u4e86\u5168\u5c40\u5f97\u5206\u7684\u6982\u5ff5 \u2014\u2014Elo \u8bc4\u7ea7\u3002\u5b83\u80fd\u66f4\u5168\u9762\u5730\u53cd\u6620\u6a21\u578b\u76f8\u5bf9\u8868\u73b0\u7684\u53d8\u5316\uff0c\u6db5\u76d6\u4e0d\u540c\u65f6\u95f4\u548c\u591a\u6b21\u8bc4\u4f30\u4e2d\u7684\u8868\u73b0\u3002\u901a\u8fc7\u5f15\u5165 Elo \u8bc4\u7ea7\uff0c\u53ef\u4ee5\u5728\u8bc4\u4f30\u8fc7\u7a0b\u4e2d\u540c\u65f6\u8003\u8651\u6a21\u578b\u5728\u5355\u573a\u6bd4\u8d5b\u4e2d\u7684\u5c40\u90e8\u8868\u73b0\u548c\u5728\u591a\u8f6e\u6bd4\u8d5b\u4e2d\u7684\u5168\u5c40\u8868\u73b0\uff0c\u4ece\u800c\u63d0\u4f9b\u4e00\u4e2a\u66f4\u4e3a\u7a33\u5065\u548c\u51c6\u786e\u7684\u6a21\u578b\u7efc\u5408\u80fd\u529b\u5ea6\u91cf\uff0c\u8fd9\u6709\u52a9\u4e8e\u964d\u4f4e\u8f83\u5f31\u6a21\u578b\u7531\u4e8e\u5076\u7136\u7684\u3001\u4e0d\u5177\u4ee3\u8868\u6027\u7684\u6295\u7968\u800c\u83b7\u80dc\u7684\u98ce\u9669\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/836f5bd3140e87b.png\" \/> <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/948b473cc132b32.png\" \/> \u6700\u540e\u7684 response \u5206\u6570\u7531 Elo \u8bc4\u7ea7\u548c\u88c1\u5224\u6295\u7968\u52a0\u6743\u5f97\u5230\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/6fcff6a88a53d15.png\" \/> \u6bcf\u4e00\u4e2a response \u90fd\u8981\u548c\u6240\u6709\u5bf9\u624b\u7684 response \u6bd4\u8f83\uff0c\u56e0\u6b64\u8be5\u5206\u6570\u4ee3\u8868\u4e86\u5f53\u524d response \u7684\u7ec4\u5185\u76f8\u5bf9\u4f18\u52bf\u3002<\/p>\n<p><strong>2.4 Final Training<\/strong><\/p>\n<p>\u672c\u6587\u5f97\u5230\u7684\u6570\u636e\u683c\u5f0f\u4e3a instruction\u3001\u6765\u81ea\u4e8e\u5404\u4e2a\u53c2\u8d5b\u8005\u7684 response\u3001\u5404 response \u5bf9\u5e94\u7684\u5206\u6570\u3002\u8fd9\u79cd\u6570\u636e\u683c\u5f0f\u53ef\u4ee5\u652f\u6301\u591a\u79cd post-training \u65b9\u6cd5\uff0c\u6bd4\u5982 SFT\u3001DPO\u3001KTO \u7b49\u7b49\u3002\u672c\u6587\u91c7\u7528 SFT\uff0c\u5c06\u7ec4\u5185\u5206\u6570\u6700\u9ad8\u7684 response \u4f5c\u4e3a gold output\uff0c\u4f7f\u5f97 WarriorCoder \u5728\u8bad\u7ec3\u4e2d\u53ef\u4ee5\u878d\u5408\u5404\u4e2a\u53c2\u8d5b\u8005\u7684\u4f18\u52bf\uff0c\u96c6\u767e\u5bb6\u4e4b\u957f\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>03 \u5b9e\u9a8c<\/h3>\n<p><strong>3.1 \u4e3b\u8981\u7ed3\u679c<\/strong><br \/>\n\u8868 1 \u663e\u793a\u4e86 WarriorCoder \u5728 code generation benchmark \u4e0a\u7684\u8868\u73b0\u3002\u4e0e\u540c\u7c7b\u5de5\u4f5c\u76f8\u6bd4\uff0cWarriorCoder \u5728 HumanEval\u3001HumanEval+\u3001MBPP \u548c MBPP + \u4e0a\u53d6\u5f97\u4e86 SOTA\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cWarriorCoder \u5b8c\u5168\u4e0d\u9700\u8981\u79c1\u6709\u5927\u6a21\u578b\uff08\u5982 GPT-4 \u7b49\uff09\u5c31\u53d6\u5f97\u4e86\u60ca\u8273\u7684\u6548\u679c\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/ffe46ab4c9ff904.png\" \/> <em>\u8868 1<\/em><\/p>\n<p>\u6b64\u5916\u5728 code reasoning benchmark \u548c libraries using benchmark \u4e0a\uff0cWarriorCoder \u4e5f\u53d6\u5f97\u4e86\u5353\u8d8a\u7684\u6210\u7ee9\u3002\u5982\u8868 2 \u548c\u8868 3 \u6240\u793a\uff0cWarriorCoder \u5728\u7edd\u5927\u591a\u6570\u6307\u6807\u4e0a\u8868\u73b0\u6700\u4f18\uff0c\u751a\u81f3\u8d85\u8d8a\u4e86 15B \u548c 34B \u7b49\u66f4\u5927\u91cf\u7ea7\u7684\u6a21\u578b\u3002\u8fd9\u4e5f\u8bc1\u660e\u4e86\u672c\u6587\u63d0\u51fa\u7684\u65b9\u6cd5\u5177\u6709\u826f\u597d\u7684\u6cdb\u5316\u6027\uff0c\u53ef\u4ee5\u8ba9\u6a21\u578b\u4ece\u591a\u4e2a\u4ee3\u7801\u4e13\u5bb6\u5927\u6a21\u578b\u5904\u83b7\u5f97\u591a\u79cd\u4e0d\u540c\u7684\u80fd\u529b\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/b5005f6fa19549e.png\" \/> <em>\u8868 2<\/em> <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/24adfda345f9e62.png\" \/> <em>\u8868 3<\/em><\/p>\n<p><strong>3.2 \u6570\u636e\u5206\u6790<\/strong><\/p>\n<p>\u672c\u6587\u8fd8\u5bf9\u6240\u6784\u9020\u7684\u8bad\u7ec3\u6570\u636e\u8fdb\u884c\u4e86\u5206\u6790\uff0c\u4ece Dependence\u3001Diversity\u3001Difficulty \u4e09\u4e2a\u89d2\u5ea6\u8fdb\u884c\u7814\u7a76\u3002<br \/>\nDependence<br \/>\n\u6b64\u524d\u7684\u5de5\u4f5c\u5f80\u5f80\u4f1a\u57fa\u4e8e\u4e00\u4e9b\u73b0\u6709\u7684\u4ee3\u7801\u6570\u636e\u96c6\uff08\u5982\uff09\u8fdb\u884c\u6269\u5c55\u3001\u6570\u636e\u589e\u5f3a\uff0c\u800c\u672c\u6587\u5219\u662f\u5b8c\u5168\u4ece\u96f6\u5f00\u59cb\u6784\u9020\u5168\u65b0\u7684\u6570\u636e\u3002\u5982\u56fe 5 \u6240\u793a\uff0c\u4f5c\u8005\u8ba1\u7b97\u4e86\u8bad\u7ec3\u6570\u636e\u4e0e\u4e24\u4e2a\u5e38\u7528\u4ee3\u7801\u6570\u636e\u96c6\u7684\u91cd\u53e0\u7a0b\u5ea6\uff08rouge \u6307\u6807\uff09\uff0c\u7edd\u5927\u591a\u6570\u6307\u4ee4\u4e0e codealpaca \u548c codeultrafeedback \u7684 ROUGE \u5f97\u5206\u4f4e\u4e8e 0.3\uff0c\u8868\u660e\u5b83\u4eec\u4e0e\u73b0\u6709\u6570\u636e\u96c6\u4e2d\u7684\u6307\u4ee4\u5728\u5185\u5bb9\u4e0a\u5b58\u5728\u8f83\u5927\u5dee\u5f02\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6316\u6398\u51fa\u7684\u6307\u4ee4\u4e2d\u6ca1\u6709\u4efb\u4f55\u4e00\u9879 ROUGE \u6307\u6807\u8d85\u8fc7 0.6\uff0c\u8fd9\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u8fd9\u4e9b\u6307\u4ee4\u6765\u6e90\u4e8e\u4e13\u5bb6\u5927\u6a21\u578b\u7684\u5185\u90e8\u5206\u5e03\uff0c\u800c\u975e\u73b0\u6709\u8bad\u7ec3\u6570\u636e\u7684\u7b80\u5355\u590d\u5236\u6216\u6269\u5c55\u3002\u56e0\u6b64\uff0c\u8fd9\u4e9b\u6307\u4ee4\u66f4\u65b0\u9896\u3001\u5177\u6709\u66f4\u9ad8\u7684\u72ec\u7acb\u6027\uff0c\u8fd9\u5bf9\u4e8e\u8bad\u7ec3\u5c24\u4e3a\u5b9d\u8d35\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/88e760aa213e81b.png\" \/> <em>\u56fe 5<\/em><\/p>\n<p>\u8868 4 \u5c55\u793a\u4e86\u8bad\u7ec3\u6570\u636e\u7684\u6784\u6210\uff0c\u6db5\u76d6\u4e86 7 \u79cd\u4e0d\u540c\u7684 code \u4efb\u52a1\uff0c\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 WarriorCoder \u80fd\u591f\u5728\u591a\u4e2a benchmark \u4e0a\u8868\u73b0\u4f18\u5f02\u7684\u539f\u56e0\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f code reasoning \u4ec5\u5360\u6bd4 2.9% \u5c31\u4f7f\u5f97 WarriorCoder \u5728\u76f8\u5173 benchmark \u4e0a\u5177\u6709\u60ca\u4eba\u8868\u73b0\uff0c\u8fd9\u8bf4\u660e\u4e86\u672c\u6587\u63d0\u51fa\u7684\u65b9\u6cd5\u5177\u6709\u5f88\u5927\u6f5c\u529b\uff0c\u5982\u679c\u9488\u5bf9\u6a21\u578b\u7684\u5f31\u70b9\u5b9a\u5411\u6316\u6398\u6570\u636e\u53ef\u4ee5\u8ba9\u6a21\u578b\u80fd\u529b\u66f4\u4e0a\u4e00\u5c42\u697c\u3002\u6b64\u5916\uff0c\u56fe 6 \u7684\u70ed\u529b\u56fe\u4e5f\u5c55\u793a\u4e86\u53c2\u8d5b\u8005\u5bf9\u6297\u7ed3\u679c\uff0c\u5373\u4fbf\u662f\u518d\u5f3a\u7684\u6a21\u578b\u4e5f\u7ec8\u7a76\u6709\u8868\u73b0\u4e0d\u597d\u7684\u65f6\u5019\uff0c\u800c WarriorCoder \u4ec5\u5411\u5f53\u524d\u6307\u4ee4\u4e0b\u5206\u6570\u6700\u9ad8\u7684 winner response \u5b66\u4e60\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/a6c17c55ce14b18.png\" \/> <em>\u8868 4<\/em> <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/d8806f14174efbb.png\" \/> <em>\u56fe 6<\/em><em><br \/>\n<\/em><br \/>\n\u56fe 7 \u5c55\u793a\u4e86\u4e0d\u540c\u6a21\u578b\u4ea7\u751f\u7684\u6307\u4ee4\u7684\u96be\u5ea6\u6bd4\u4f8b\u3002\u5927\u591a\u6570\u6307\u4ee4\u7684\u96be\u5ea6\u5904\u4e8e good \u7b49\u7ea7\uff0c\u5f97\u5206\u5728 6 \u5230 8 \u4e4b\u95f4\u3002\u88ab\u8bc4\u4e3a excellent\uff08\u5f97\u5206 9-10\uff09\u7684\u6307\u4ee4\u4ec5\u5360\u6570\u636e\u96c6\u7684\u4e00\u5c0f\u90e8\u5206\uff0c\u8868\u660e\u9ad8\u5ea6\u590d\u6742\u6216\u9ad8\u7ea7\u7684\u4efb\u52a1\u76f8\u5bf9\u8f83\u4e3a\u7a00\u5c11\u3002\u4f5c\u8005\u5c06\u5f97\u5206\u4f4e\u4e8e 6 \u7684\u6307\u4ee4\u88ab\u6392\u9664\u5728\u8bad\u7ec3\u96c6\u4e4b\u5916\uff0c\u56e0\u4e3a\u5b83\u4eec\u5f80\u5f80\u8981\u4e48\u8fc7\u4e8e\u7b80\u5355\uff0c\u8981\u4e48\u8fc7\u4e8e\u6a21\u7cca\uff0c\u8fd9\u6837\u7684\u6307\u4ee4\u4f1a\u5bf9\u8bad\u7ec3\u9636\u6bb5\u6709\u5bb3\uff0c\u751a\u81f3\u53ef\u80fd\u524a\u5f31\u6a21\u578b\u7684\u6027\u80fd\u548c\u6cdb\u5316\u80fd\u529b\u3002 <img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/0ce45cb809d7844.png\" \/> <em>\u56fe 7<\/em><\/p>\n<h2><em><br \/>\n<\/em>04 \u76f8\u5173\u8d44\u6e90<\/h2>\n<p>\u867d\u7136\u4f5c\u8005\u5f53\u524d\u5e76\u672a\u5f00\u6e90\u6a21\u578b\uff0c\u4f46\u662f\u6211\u4eec\u53d1\u73b0\u5df2\u7ecf\u6709\u4eba\u590d\u73b0\u4e86\u4f5c\u8005\u7684\u5de5\u4f5c\uff0c\u5730\u5740\u5982\u4e0b\uff1a<br \/>\n\u9879\u76ee\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/HuggingMicah\/warriorcoder_reproduce<\/p>\n","protected":false},"excerpt":{"rendered":"<p>&nbsp; \u8bba\u6587\u6807\u9898\uff1aWarriorCoder: Learning from Expert Battles to Augment Code Large Language Models \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/&#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-27311","post","type-post","status-publish","format-standard","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/27311","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=27311"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/27311\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=27311"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=27311"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=27311"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}