{"id":562,"date":"2024-02-06T19:23:09","date_gmt":"2024-02-06T11:23:09","guid":{"rendered":"http:\/\/lingbo.online\/?p=562"},"modified":"2024-03-06T10:24:44","modified_gmt":"2024-03-06T02:24:44","slug":"instructgpt","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/algorithm_learning\/llms\/instructgpt\/","title":{"rendered":"InstructGPT\u8bba\u6587\u7cbe\u8bfb"},"content":{"rendered":"<h1>\u80cc\u666f<\/h1>\n<p>ChatGPT\u548cInstructGPT\u662f\u4e00\u5bf9\u5b6a\u751f\u5144\u5f1f\uff0c\u5b83\u4eec\u5728\u6a21\u578b\u7ed3\u6784\u548c\u8bad\u7ec3\u65b9\u5f0f\u4e0a\u90fd\u5b8c\u5168\u4e00\u81f4\uff0c\u6838\u5fc3\u601d\u60f3\u5728\u4e8e\u4f7f\u7528\u6307\u793a\u5b66\u4e60\uff08Instruction Learning\uff09\u548c\u4eba\u5de5\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning from Human Feedback\uff0cRLHF\uff09\u6765\u6307\u5bfc\u6a21\u578b\u7684\u8bad\u7ec3\u3002InstructGPT\u4e0eGPT-3\u7684\u7ed3\u6784\u76f8\u540c\uff0c\u533a\u522b\u5728\u4e8e\u5f15\u5165\u4e86\u4eba\u5de5\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60RLHF\u8fdb\u884c\u5fae\u8c03\u3002\u5728\u4f5c\u8005\u7684\u63d0\u793a\u5206\u5e03\u7684\u4eba\u5de5\u8bc4\u4f30\u4e2d\uff0c\u5c3d\u7ba1\u53c2\u6570\u5c11\u4e86100\u500d\uff0c\u4f46\u6765\u81ea13b\u53c2\u6570\u7684InstructGPT\u6a21\u578b\u7684\u8f93\u51fa\u6bd4\u6765\u81ea175B\u53c2\u6570\u7684GPT-3\u7684\u8f93\u51fa\u66f4\u53d7\u6b22\u8fce\u3002<br \/>\nInstructGPT\u8bba\u6587\u7684\u9605\u8bfb\u5730\u5740\u4e3a\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2203.02155\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/abs\/2203.02155<\/a><\/p>\n<h1>\u6d41\u7a0b\u7b80\u4ecb<\/h1>\n<p>InctructGPT\u4e3b\u8981\u4f7f\u7528\u6765\u81ea\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60(\u5229\u7528\u4eba\u7c7b\u7684\u504f\u597d\u4f5c\u4e3a\u5956\u52b1\u4fe1\u53f7)\uff0c\u5bf9GPT-3\u8fdb\u884c\u5fae\u8c03\u3002\u5177\u4f53\u7684\u5b9e\u73b0\u6b65\u9aa4\u6982\u62ec\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li><strong>\u6536\u96c6\u793a\u8303\u6570\u636e\uff0c\u8bad\u7ec3\u6709\u76d1\u7763\u7b56\u7565\u3002<\/strong>\u96c7\u4f6340\u4e2a\u5916\u5305\u6807\u6ce8\u6570\u636e\uff0c\u6839\u636eprompts\uff0c\u4eba\u5de5\u64b0\u5199\u4e00\u7cfb\u5217demonstration\uff08\u6f14\u793a\uff09\u4f5c\u4e3a\u6a21\u578b\u7684\u671f\u671b\u8f93\u51fa\u3002<strong>\u4f7f\u7528\u8fd9\u90e8\u5206\u6570\u636e\u96c6\u5bf9\u9884\u8bad\u7ec3\u7684GPT-3\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff08\u76d1\u7763\u5b66\u4e60\u57fa\u7ebf\uff09<\/strong>\u3002<\/li>\n<li><strong>\u6536\u96c6\u6bd4\u8f83\u6570\u636e\uff0c\u8bad\u7ec3\u5956\u52b1\u6a21\u578b\u3002<\/strong>\u4e0a\u8ff0\u6a21\u578b\u4ea7\u751f\u4e00\u7cfb\u5217\u8f93\u51fa\uff0c\u6807\u6ce8\u8005(labelers)\u5bf9\u8fd9\u4e9b\u8f93\u51fa\u8fdb\u884c\u6bd4\u8f83\u548c\u6392\u5e8f\uff08\u7531\u597d\u5230\u574f\uff09\u3002\u57fa\u4e8e\u8fd9\u4e9b\u6570\u636e\u96c6\uff0c<strong>\u8bad\u7ec3\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\uff08Reward Model\uff09\uff0c\u7528\u4e8e\u9884\u6d4b\u6807\u6ce8\u8005\u66f4\u559c\u6b22\u54ea\u4e00\u4e2a\u8f93\u51fa<\/strong>\u3002<\/li>\n<li><strong>\u4f7f\u7528PPO\u9488\u5bf9\u5956\u52b1\u6a21\u578b\u4f18\u5316\u7b56\u7565\u3002<\/strong>\u4f7f\u7528\u7b2c\u4e8c\u6b65\u7684RM\u4f5c\u4e3a\u5956\u52b1\u51fd\u6570\uff0c\u5fae\u8c03\u7b2c\u4e00\u6b65\u7684\u76d1\u7763\u5b66\u4e60baseline\uff0c\u4f7f\u7528PPO\u7b97\u6cd5\u6700\u5927\u5316\u5956\u52b1\u3002<\/li>\n<\/ul>\n<p>\u6b65\u9aa42\u548c\u6b65\u9aa43\u53ef\u4ee5\u8fde\u7eed\u8fed\u4ee3\u3002\u5728\u5f53\u524d\u6700\u4f73\u7b56\u7565\u4e0a\uff0c\u6536\u96c6\u66f4\u591a\u7684\u6bd4\u8f83\u6570\u636e\uff0c\u7528\u6765\u8bad\u7ec3RM\u6a21\u578b\u548c\u65b0\u7b56\u7565\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u5927\u591a\u6570\u6bd4\u8f83\u6570\u636e\u6e90\u4e8e\u76d1\u7763\u7b56\u7565\uff0c\u90e8\u5206\u6765\u81ea\u4e8ePPO\u7b56\u7565\u3002<\/p>\n<p>\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e9b\u4e0d\u4e2d\u5c06\u4f1a\u628aGPT3\u7684\u884c\u4e3a\u4e0e\u90e8\u5206\u7279\u5b9a\u4eba\u7fa4\uff08\u6807\u6ce8\u8005\u548c\u7814\u7a76\u8005\uff09\u7684\u504f\u597d\u8fdb\u884c\u5bf9\u9f50\uff0c\u4ec5\u80fd\u4ee3\u8868\u8fd9\u4e9b\u4eba\u7684\u504f\u597d\uff0c\u800c\u4e0d\u662f\u201c\u4eba\u7c7b\u201d\u504f\u597d\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/intruct_GPT.jpg\" alt=\"\" \/><\/p>\n<h1>\u65b9\u6cd5\u548c\u5b9e\u9a8c\u7ec6\u8282<\/h1>\n<h2>\u6570\u636e\u96c6<\/h2>\n<p>prompt\u6570\u636e\u96c6\u6700\u521d\u7531\u63d0\u4ea4\u7ed9OpenAI API\u7684\u6587\u672c\u63d0\u793a\u7ec4\u6210\u3002\u4f7f\u7528Playround\u7684\u7528\u6237\u4f1a\u88ab\u544a\u77e5\uff0c\u53ea\u8981\u4f7f\u7528InstructGPT\u6a21\u578b\uff0c\u8fd9\u4e9b\u6570\u636e\u5c31\u4f1a\u88ab\u6536\u96c6\u8d77\u6765\uff0c\u4ee5\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u6a21\u578b\u8bad\u7ec3\u3002<br \/>\n\u5bf9\u6536\u96c6\u5230\u7684prompt\u505a\u540e\u5904\u7406\uff0c\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u68c0\u67e5\u957f\u516c\u6709\u524d\u7f00prompt\uff0c\u6d88\u9664\u91cd\u590dprompt\uff1b<\/li>\n<li>\u9650\u5236\u6bcf\u4e2a\u7528\u6237ID\u7684prompts\u6570\u91cf\u4e3a200\uff1b<\/li>\n<li>\u57fa\u4e8e\u7528\u6237ID\u521b\u5efa\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u548c\u6d4b\u8bd5\u96c6\uff0c\u4ee5\u4fbf\u9a8c\u8bc1\u96c6\u548c\u6d4b\u8bd5\u96c6\u4e2d\u4e0d\u5305\u542b\u8bad\u7ec3\u96c6\u7684\u6570\u636e\u3002<\/li>\n<li>\u5bf9\u6570\u636e\u8fdb\u884c\u8131\u654f\u5904\u7406\uff0c\u907f\u514d\u6a21\u578b\u5b66\u4e60\u5230\u6f5c\u5728\u7684\u654f\u611f\u5ba2\u6237\u7ec6\u8282<\/li>\n<\/ul>\n<p>\u5728\u8bad\u7ec3\u6700\u65e9\u7684IntructGPT\u65f6\uff0clabelers\u81ea\u5df1\u64b0\u5199prompts\uff0c\u5177\u4f53\u5305\u542b\u4e09\u79cd\uff1a<\/p>\n<ol>\n<li>\u7b80\u5355(Plain)\uff1a\u8981\u6c42labeler\u81ea\u5df1\u63d0\u51fa\u4efb\u610f\u7684\u4efb\u52a1\uff0c\u540c\u65f6\u786e\u4fdd\u4efb\u52a1\u5177\u6709\u591a\u6837\u6027\u3002<\/li>\n<li>\u5c11\u6837\u672c(Few-shot)\uff0c\u8981\u6c42labelers\u7ed9\u51fa\u4e00\u6761\u6307\u4ee4\uff0c\u540c\u65f6\u7ed9\u51fa\u8fd9\u6761\u6307\u4ee4\u5bf9\u5e94\u7684\u591a\u4e2a\u67e5\u8be2\/\u54cd\u5e94\u5bf9\u3002<\/li>\n<li>\u57fa\u4e8e\u7528\u6237(User-based)\uff1a\u8981\u6c42labelers\u7ed9\u51faOpenAI API\u5e94\u7528\u6848\u4f8b\u5bf9\u5e94\u7684prompts<\/li>\n<\/ol>\n<p>\u4f9d\u636e\u8fd9\u4e9b\u63d0\u793a\uff0cIntructGPT\u751f\u6210\u4e86\u7528\u4e8e\u5fae\u8c03\u8fc7\u7a0b\u7684\u662f\u4e09\u4e2a\u4e0d\u540c\u7684\u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li>SFT dataset\uff1a\u7531labelers\u6807\u6ce8\u7684\u4f8b\u5b50\uff0c\u7528\u4e8e\u8bad\u7ec3SFT model\uff0c\u5305\u542b13k\u4e2a\u8bad\u7ec3prompts\uff08\u6765\u6e90\u4e8eAPI\u548c\u4eba\u5de5\u64b0\u5199\uff09<\/li>\n<li>RM dataset\uff1a\u7531labelers\u5bf9\u6a21\u578b\u8f93\u51fa\u7ed3\u679c\u7684\u6392\u5e8f\u5bf9\u6570\u636e\uff0c\u7528\u4e8e\u8bad\u7ec3Reward Model\uff0c\u5305\u542b33k\u4e2a\u8bad\u7ec3prompts\uff08\u6765\u6e90\u4e8eAPI\u548c\u4eba\u5de5\u64b0\u5199\uff09<\/li>\n<li>PPO dataset\uff1a\u65e0\u4eba\u5de5\u6807\u7b7e\uff0c\u7528\u4e8e\u4f5c\u4e3aRLHF fine-tuning \u7684\u8f93\u5165\u6570\u636e\uff0c\u5305\u542b31k\u4e2a\u8bad\u7ec3prompts\uff08\u4ec5\u6765\u6e90\u4e8eAPI\uff09\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/data_number.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/data_distribution.jpg\" alt=\"\" \/><\/li>\n<\/ul>\n<h2>\u6a21\u578b\u8bad\u7ec3<\/h2>\n<p>Intruct GPT\u4eceGPT-3\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5f00\u59cb\u3002\u8fd9\u4e9b\u6a21\u578b\u662f\u5728\u5e7f\u6cdb\u5206\u5e03\u7684\u4e92\u8054\u7f51\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\u7684\uff0c\u53ef\u4ee5\u9002\u5e94\u5e7f\u6cdb\u7684\u4e0b\u6e38\u4efb\u52a1\uff0c\u4f46\u884c\u4e3a\u7279\u5f81\u4e0d\u4f73\u3002\u4ece\u8fd9\u4e9b\u6a21\u578b\u5f00\u59cb\uff0c\u6211\u4eec\u7528\u4e09\u79cd\u4e0d\u540c\u7684\u6280\u672f\u8bad\u7ec3\u6a21\u578b:<\/p>\n<h3>Supervised fine-tuning (SFT)<\/h3>\n<p>\u4f7f\u7528\u6709\u76d1\u7763\u5b66\u4e60\u5bf9GPT-3\u8fdb\u884c\u5fae\u8c03\u3002\u6587\u7ae0\u5171\u8bad\u7ec3\u4e8616\u4e2aepoch\uff0c\u4f7f\u7528\u4f59\u5f26\u5b66\u4e60\u8870\u51cf\u7387\uff0cresidual dropout\u4e3a0.2\u3002\u6587\u7ae0\u6839\u636e\u9a8c\u8bc1\u96c6\u4e0a\u7684RM\u5206\u6570\u8fdb\u884c\u6700\u7ec8\u7684SFT\u6a21\u578b\u9009\u62e9\u3002<\/p>\n<h3>Reward modeling (RM)<\/h3>\n<p><strong>RM\u662f\u5c06SFT\u6a21\u578b\u6700\u540e\u7684\u5d4c\u5165\u5c42\u53bb\u6389\u540e\u7684\u6a21\u578b\uff0c\u5b83\u7684\u8f93\u5165\u662fprompt\u548cresponse\uff0c\u8f93\u51fa\u662f\u6807\u91cf\u7684\u5956\u52b1\u503c\u3002<\/strong>RM\u4e00\u822c\u662f\u5728\u540c\u4e00\u8f93\u5165\u7684\u4e24\u4e2a\u6a21\u578b\u8f93\u51fa\u4e4b\u95f4\u7684\u6bd4\u8f83\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u7684\uff0c\u4f7f\u7528\u4ea4\u53c9\u71b5loss\u548c\u5956\u52b1\u7684\u5dee\u5f02\u4f5c\u4e3alabel\u3002\u5728\u6587\u7ae0\u4e2d\uff0c\u4ec5\u4f7f\u75286B\u7684RM\u6a21\u578b\uff0c\u4e00\u65b9\u9762\u8282\u7701\u4e86\u5927\u91cf\u8ba1\u7b97\u6d88\u8017\uff0c\u53e6\u4e00\u65b9\u9762175B\u7684RM\u8bad\u7ec3\u4e0d\u7a33\u5b9a\uff0c\u4e0d\u592a\u9002\u5408\u4f5c\u4e3a\u5f3a\u5316\u5b66\u4e60\u7684\u503c\u51fd\u6570\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u6bcf\u4e2aprompt\uff0cInstructGPT\u4f1a\u968f\u673a\u751f\u6210K\u4e2a\u8f93\u51fa\uff084 &lt;= K &lt;= 9\uff09\u3002labeler\u5bf9\u8fd9\u4e9b\u8f93\u51fa\u8fdb\u884c\u6392\u5e8f\uff0c\u6bcf\u4e2aprompt\u7684\u8f93\u51fa\u53ef\u4ee5\u4ea7\u751f<span class=\"katex-eq\" data-katex-display=\"false\">C_k^2<\/span>\u4e2apair\u5bf9,\u5f53K = 9\u65f6\uff0c\u4f1a\u4ea7\u751f36\u4e2apair\u5bf9\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cInstructGPT\u5c06\u6bcf\u4e2aprompt\u7684<span class=\"katex-eq\" data-katex-display=\"false\">C_k^2<\/span>\u4e2apair\u5bf9\u4f5c\u4e3a\u4e00\u4e2a\u5355\u72ec\u7684batch\u3002\u8fd9\u79cd\u6309\u7167prompt\u4f5c\u4e3abatch\u7684\u8bad\u7ec3\u65b9\u5f0f\uff0c\u6bd4\u4f20\u7edf\u7684\u6309\u7167\u6837\u672c\u4f5c\u4e3abatch\u7684\u8ba1\u7b97\u65b9\u5f0f\u9ad8\u6548\u5f97\u591a\uff08\u53ea\u9700\u8981\u8ba1\u7b97\u4e00\u6b21RM\u7684\u524d\u5411\u4f20\u9012\uff09\uff0c\u4e14\u4e0d\u5bb9\u6613\u8fc7\u62df\u5408\uff0c\u56e0\u4e3a\u8fd9\u79cd\u65b9\u5f0f\u4e0b\uff0c\u6bcf\u4e2aprompt\u4ec5\u4f1a\u5582\u7ed9\u6a21\u578b\u4e00\u6b21\u3002<\/p>\n<p>\u5956\u52b1\u6a21\u578b\u7684\u635f\u5931\u51fd\u6570\u4f7f\u7528\u7684\u662f\u6392\u5e8f\u4e2d\u5e38\u89c1\u7684pairwise ranking loss\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/loss.jpg\" alt=\"\" \/><\/p>\n<p>\u5176\u4e2d\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">r_{\\theta}(x,y)<\/span>\u8868\u793a\u5728prompt<span class=\"katex-eq\" data-katex-display=\"false\">x<\/span>\u548c\u54cd\u5e94<span class=\"katex-eq\" data-katex-display=\"false\">y<\/span>\u5728\u53c2\u6570\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">\\theta<\/span>\u7684\u5956\u52b1\u6a21\u578b\u4e0b\u7684\u5956\u52b1\u503c\u3002\u5728{<span class=\"katex-eq\" data-katex-display=\"false\">y_w,y_l<\/span>}\u7684pair\u5bf9\u4e0b\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">y_w<\/span>\u8868\u793alabeler\u66f4\u559c\u6b22\u7684\u54cd\u5e94\u7ed3\u679c\u3002\u76ee\u6807\u51fd\u6570\u5c31\u662f\u6700\u5927\u5316\u8fd9\u4e24\u4e2a\u5956\u52b1\u7684\u5dee\u503c\uff0c\u8fd9\u4e2a\u5dee\u503c\u7ecf\u8fc7\u4e00\u4e2aSigmoid\u51fd\u6570\u548clog\u51fd\u6570\u518d\u53d6\u8d1f\u540e\uff0c\u5f97\u5230\u635f\u5931<\/p>\n<h3>Reinforcement learning(RL)<\/h3>\n<p>\u6587\u7ae0\u4f7f\u7528PPO\u7b97\u6cd5\u5bf9SFT\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002InstructGPT\u4e2d\u7684RL\u73af\u5883\u662fbandit environment\uff0c\u63d0\u4f9b\u968f\u673a\u7528\u6237\u7684prompt\u5e76\u671f\u671b\u5bf9\u63d0\u793a\u505a\u51fa\u54cd\u5e94\u3002\u7ed9\u5b9a\u63d0\u793a\u548c\u53cd\u5e94\uff0c\u4ed6\u4f1a\u4ea7\u751f\u7531RM\u51b3\u5b9a\u7684\u5956\u52b1\uff0c\u5e76\u4e14\u7ed3\u675f\u5f53\u524depisode.<br \/>\n\u5728RL\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6700\u5927\u5316\u7ec4\u5408\u76ee\u6807\u51fd\u6570\u5982\u4e0b\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/rl_loss.jpg\" alt=\"\" \/><\/p>\n<p>\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_\\Theta^{RL}<\/span>\u8868\u793a\u5f85\u5b66\u4e60\u7684RL\u7b56\u7565\uff08\u6700\u5f00\u59cb\u7531 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi^{SFT}<\/span>\u521d\u59cb\u5316\uff09\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\pi^{SFT}<\/span>\u662f\u5728\u6807\u6ce8\u597d\u7684\u95ee\u9898\u548c\u7b54\u6848\u57fa\u7840\u4e0a\uff0c\u7528\u6709\u76d1\u7763\u5fae\u8c03\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\uff08step1\uff09\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">D_{pretrain}<\/span>\u8868\u793a\u9884\u8bad\u7ec3\u5206\u5e03\u3002<\/p>\n<p>\u5bf9\u76ee\u6807\u51fd\u6570\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u62c6\u89e3\u548c\u5206\u6790\uff1a<\/p>\n<ul>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">x<\/span>\u8868\u793aPPO\u6570\u636e\u96c6\uff0c\u5305\u542b31000\u4e2aprompt\uff08\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e00\u76f4\u56fa\u5b9a\uff09\u3002\u5bf9\u4e8e\u6bcf\u4e00\u4e2aprompt\uff0c\u672a\u5165\u5f53\u524d\u7684RL\u7b56\u7565\uff0c\u4ea7\u751f<span class=\"katex-eq\" data-katex-display=\"false\">y<\/span>\uff08\u6bcf\u4e00\u6b21\u901a\u8fc7\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u66f4\u65b0\u6a21\u578b\u540e\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">y<\/span>\u603b\u4f1a\u53d1\u751f\u53d8\u5316\uff09\uff0c\u7136\u540e\u628a<span class=\"katex-eq\" data-katex-display=\"false\">\uff08x,y\uff09<\/span>\u4e22\u8fdb\u4e0a\u4e00\u6b65\u8bad\u7ec3\u597d\u7684RM\u6a21\u578b\u4e2d\uff0c\u8ba1\u7b97\u51fa\u5206\u6570\u3002\u6700\u5927\u5316<span class=\"katex-eq\" data-katex-display=\"false\">r_\\theta(x,y)<\/span>\u3002\u76ee\u6807\uff1a\u65b0\u8bad\u7ec3\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\u603b\u662f\u4eba\u7c7b\u89c9\u5f97\u6700\u597d\u7684\u90a3\u4e2a\u7b54\u6848\u3002\u5f3a\u5316\u5b66\u4e60\u5e76\u975e\u7b80\u5355\u7684\u7edf\u8ba1\u5b66\u4e60\uff0c\u5b83\u7684\u6570\u636e\u5206\u5e03\u4f1a\u968f\u7740\u6a21\u578b\u7684\u66f4\u65b0\u800c\u4e0d\u65ad\u53d8\u5316\uff08\u73af\u5883\u53d1\u751f\u53d8\u5316\uff09\u3002<span class=\"katex-eq\" data-katex-display=\"false\">r_\\theta(x,y)<\/span>\u5176\u5b9e\u5c31\u662f\u5b66\u4e60\u4eba\u7684\u6392\u5e8f\uff0c\u4ece\u800c\u7ed9\u6a21\u578b\u5b9e\u65f6\u7684\u53cd\u9988\u3002<\/li>\n<li>\u968f\u7740\u6a21\u578b\u7684\u66f4\u65b0\uff0cRL\u4ea7\u751f\u7684\u8f93\u51fa\u548cSFT\u6a21\u578b\u7684\u6570\u636e\u5206\u5e03\u5dee\u5f02\u4f1a\u8d8a\u6765\u8d8a\u5927\u3002\u901a\u8fc7\u5728loss\u4e2d\u6dfb\u52a0KL\u6563\u5ea6\uff08\u8bc4\u4f30\u4e24\u4e2a\u6982\u7387\u5206\u5e03\u7684\u76f8\u4f3c\u5ea6\uff09\uff0c\u4ee5\u671fRL\uff08PPO\uff09\u4ea7\u751f\u7684\u8f93\u51fa\u4e0eSFT\u6a21\u578b\u8f93\u51fa\u4e0d\u8981\u504f\u79bb\u592a\u8fdc\u3002\u8be5\u9879\u76f8\u5f53\u4e8e\u4e00\u4e2a\u6b63\u5219\u9879<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/kl.jpg\" alt=\"\" \/><\/li>\n<li>\u5982\u679c\u53ea\u4f7f\u7528\u4e0a\u8ff0\u4e24\u9879\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f1a\u5bfc\u81f4\u8be5\u6a21\u578b\u4ec5\u4ec5\u5bf9\u4eba\u7c7b\u7684\u6392\u5e8f\u7ed3\u679c\u8f83\u597d\uff0c\u800c\u5728\u901a\u7528NLP\u4efb\u52a1\u4e0a\uff0c\u6027\u80fd\u53ef\u80fd\u4f1a\u5927\u5e45\u4e0b\u964d\u3002\u6587\u7ae0\u4e2d\u901a\u8fc7\u5728loss\u4e2d\u52a0\u5165GPT-3\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u76ee\u6807\u6765\u89c4\u907f\u8fd9\u4e00\u95ee\u9898\u3002\u52a0\u5165\u8be5\u9879\u540e\uff0cInstructGPT\u6307\u4ee3\u7684\u5c31\u662fPPO-ptx\u6a21\u578b\u3002PPO-ptx\u5c31\u662fPPO\u7684\u76ee\u6807\u51fd\u6570 + \u539f\u59cbGPT-3\u7684\u76ee\u6807\u51fd\u6570\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/gpt3.jpg\" alt=\"\" \/><\/li>\n<\/ul>\n<h2>\u8bc4\u4f30<\/h2>\n<p>\u6587\u7ae0\u4e2d\u7684\u8bc4\u4f30(helpful,honest and harmless)\u5206\u4e3a\u4e24\u4e2a\u90e8\u5206\uff1a<\/p>\n<ul>\n<li>\u8bc4\u4f30API\u5206\u5e03<\/li>\n<li>\u8bc4\u4f30\u516c\u5171NLP\u6570\u636e\u96c6\u3002\u5305\u542b\u4e24\u79cd\u7c7b\u578b\uff0c\u4e00\u662f\u80fd\u591f\u6355\u6349\u8bed\u8a00\u6a21\u578b\u5b89\u5168\u6027\u7684\u6570\u636e\u96c6\uff0c\u6bd4\u5982\u771f\u5b9e\u6027\u3001\u6bd2\u6027\u548c\u504f\u89c1\u7b49\uff1b\u4e8c\u662f\u4f20\u7edfNLP\u6570\u636e\u96c6\u5728zero-shot\u4e0a\u7684\u8868\u73b0\uff0c\u6bd4\u5982\u95ee\u7b54\u3001\u9605\u8bfb\u7406\u89e3\u548c\u6458\u8981\u7b49\u3002<\/li>\n<\/ul>\n<h1>\u8ba8\u8bba<\/h1>\n<h2>\u5bf9\u9f50<\/h2>\n<p>\u4f5c\u8005\u63d0\u51fa\uff0c\u672c\u6587\u4f7f\u7528\u7684\u201c\u5bf9\u9f50\u201d\u6280\u672f\u2014\u2014RLHF\uff0c\u662f\u7528\u4e8e\u5bf9\u9f50\u4eba\u7c7b\u7cfb\u7edf\u7684\u4e00\u4e2a\u91cd\u8981\u65b9\u6cd5\u3002<br \/>\n\u4e0e\u9884\u8bad\u7ec3\u76f8\u6bd4\uff0c\u589e\u52a0\u6a21\u578b\u5bf9\u9f50\u7684\u6210\u672c\u662f\u9002\u4e2d\u7684\u3002\u6536\u96c6\u548c\u8bad\u7ec3\u4e0a\u6587\u63d0\u5230\u7684\u51e0\u4e07\u6761prompt\u6570\u636e\uff0c\u4e0e\u8bad\u7ec3GPT-3\u7684\u82b1\u8d39\u76f8\u6bd4\uff0c\u53ea\u5360\u4e00\u5c0f\u90e8\u5206\u3002\u800c\u4e14\u4e0a\u8ff0\u7ed3\u679c\u4e5f\u8868\u660e\uff0cRLHF\u5728\u4f7f\u8bed\u8a00\u6a21\u578b\u66f4\u52a0helpful\u65b9\u9762\u975e\u5e38\u6709\u6548\uff0c\u751a\u81f3\u6bd4\u6a21\u578b\u589e\u52a0100\u500d\u66f4\u6709\u6548\u3002\u6240\u4ee5\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u9886\u57df\uff0c\u7814\u7a76alignment\u53ef\u80fd\u6bd4\u8bad\u7ec3\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u66f4\u5177\u6027\u4ef7\u6bd4\u3002<\/p>\n<h2>\u5c40\u9650\u6027<\/h2>\n<p>1\u3001\u65b9\u6cd5\u3002InstructGPT\u7684\u8868\u73b0\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\u4ece\u5916\u5305\u4eba\u5458\u90a3\u91cc\u83b7\u5f97\u7684\u53cd\u9988\u3002\u6709\u4e9b\u6807\u6ce8\u4efb\u52a1\u53ef\u80fd\u4f1a\u53d7\u5230\u6807\u6ce8\u8005\u4ef7\u503c\u89c2\uff08\u8eab\u4efd\u3001\u4fe1\u4ef0\u3001\u6587\u5316\u80cc\u666f\u548c\u4e2a\u4eba\u7ecf\u5386\u7b49\uff09\u7684\u5f71\u54cd\u3002\u753140\u4e2a\u4eba\u7ec4\u6210\u7684\u6807\u6ce8\u7fa4\u4f53\uff0c\u663e\u7136\u65e0\u6cd5\u4ee3\u8868\u6a21\u578b\u7684\u6240\u6709\u53d7\u4f17\u3002<\/p>\n<p>2\u3001\u6a21\u578b\u3002InstructGPT\u65e0\u6cd5\u505a\u5230\u5b8c\u5168align\u548c\u5b8c\u5168\u5b89\u5168\u3002\u6a21\u578b\u4ecd\u7136\u4f1a\u5728\u6ca1\u6709\u660e\u786eprompting\u7684\u60c5\u51b5\u4e0b\uff0c\u8f93\u51fa\u4e00\u4e9b\u6709\u6bd2\/\u6709\u504f\u89c1\u7684\u5185\u5bb9\u3001\u7f16\u9020\u4e8b\u5b9e\uff0c\u751a\u81f3\u4ea7\u751f\u6027\u548c\u66b4\u529b\u5185\u5bb9\u3002\u6a21\u578b\u53ef\u80fd\u4e5f\u65e0\u6cd5\u9488\u5bf9\u67d0\u4e9b\u8f93\u5165\u4ea7\u751f\u5408\u7406\u7684\u8f93\u51fa\u3002\u4f5c\u8005\u8ba4\u4e3a\uff0cInstructGPT\u6a21\u578b\u7684\u6700\u5927\u7f3a\u70b9\u662f\uff0c\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0c\u5c3d\u7ba1\u4f1a\u5bf9\u73b0\u5b9e\u4e16\u754c\u9020\u6210\u4f24\u5bb3\uff0c\u5b83\u4ecd\u7136\u4f1a\u9075\u5faa\u7528\u6237\u7684\u6307\u4ee4\u3002\u4f8b\u5982\uff0c\u5f53\u8f93\u5165\u4e00\u4e2a\u6709\u504f\u89c1\u7684prompt\u6307\u4ee4\u65f6\uff0cInstructGPT\u4f1a\u4ea7\u751f\u6bd4GPT-3\u6a21\u578b\u6bd2\u6027\u66f4\u591a\u7684\u8f93\u51fa\u7ed3\u679c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u80cc\u666f ChatGPT\u548cInstructGPT\u662f\u4e00\u5bf9\u5b6a\u751f\u5144\u5f1f\uff0c\u5b83\u4eec\u5728\u6a21\u578b\u7ed3\u6784\u548c\u8bad\u7ec3\u65b9\u5f0f\u4e0a\u90fd\u5b8c\u5168\u4e00\u81f4\uff0c\u6838\u5fc3\u601d\u60f3\u5728\u4e8e\u4f7f\u7528\u6307\u793a\u5b66\u4e60\uff08Ins &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[45],"tags":[],"class_list":["post-562","post","type-post","status-publish","format-standard","hentry","category-llms"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/562","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=562"}],"version-history":[{"count":12,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/562\/revisions"}],"predecessor-version":[{"id":571,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/562\/revisions\/571"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=562"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=562"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=562"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}