{"id":634,"date":"2024-08-30T12:14:26","date_gmt":"2024-08-30T12:14:26","guid":{"rendered":"https:\/\/119.91.218.15\/?p=634"},"modified":"2024-09-02T09:50:45","modified_gmt":"2024-09-02T09:50:45","slug":"%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0-llama-2-open-foundation-and-fine-tuned-chat-models","status":"publish","type":"post","link":"https:\/\/blog.liguanxin.cn\/index.php\/2024\/08\/30\/%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0-llama-2-open-foundation-and-fine-tuned-chat-models\/","title":{"rendered":"\u8bba\u6587\u7b14\u8bb0\u2014\u2014Llama 2: Open Foundation and Fine-Tuned Chat Models"},"content":{"rendered":"<h1>Introduction<\/h1>\n<p>\u5f00\u6e90\u7248\u672c\uff1a<\/p>\n<ul>\n<li><strong>Llama2 7B 13B 70B<\/strong><\/li>\n<li><strong>Llama 2-Chat 7B 13B 70B<\/strong><\/li>\n<\/ul>\n<p>\u8bad\u7ec3\u8fc7\u7a0b\uff1a<\/p>\n<p><img src=\"https:\/\/blog.liguanxin.cn\/wp-content\/uploads\/2024\/08\/\u56fe\u7247.png\" alt=\"\" \/><\/p>\n<p>1\u3001\u901a\u8fc7\u516c\u5f00\u6570\u636e\u96c6\u8bad\u7ec3<strong>Llama2<\/strong>\u3002<br \/>\n2\u3001\u901a\u8fc7\u6709\u76d1\u7763\u5fae\u8c03\u8bad\u7ec3\u51fa<strong>Llama 2-Chat<\/strong>\u7684\u521d\u59cb\u7248\u672c\u3002<br \/>\n3\u3001\u4f7f\u7528\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08<strong>RLHF<\/strong>\uff09\uff0c\u5982\u62d2\u7edd\u91c7\u6837*\u548c\u8fd1\u7aef\u7b56\u7565\u4f18\u5316*\uff08<strong>PPO<\/strong>\uff09\u3002<br \/>\n4\u3001\u5728\u6574\u4e2aRLHF\u9636\u6bb5\uff0c\u4e0e\u6a21\u578b\u589e\u5f3a\u5e76\u884c\u5730\u79ef\u7d2f\u8fed\u4ee3\u5956\u52b1\u5efa\u6a21\u6570\u636e\u5bf9\u4e8e\u786e\u4fdd\u5956\u52b1\u6a21\u578b\u4fdd\u6301\u5728\u5206\u5e03\u8303\u56f4\u5185\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p><strong>\u62d2\u7edd\u91c7\u6837<\/strong>*\uff1a\u4ece\u4e00\u4e2a\u5bb9\u6613\u91c7\u6837\u7684\u8f85\u52a9\u5206\u5e03\u4e2d\u62bd\u53d6\u6837\u672c\uff0c\u5e76\u4f7f\u7528\u67d0\u79cd\u63a5\u53d7\/\u62d2\u7edd\u673a\u5236\u6765\u51b3\u5b9a\u662f\u5426\u4fdd\u7559\u6240\u62bd\u5230\u7684\u6837\u672c\uff0c\u4ece\u800c\u786e\u4fdd\u6700\u7ec8\u5f97\u5230\u7684\u6837\u672c\u9075\u5faa\u76ee\u6807\u5206\u5e03\u3002<\/p>\n<p><strong>\u8fd1\u7aef\u7b56\u7565\u4f18\u5316<\/strong>*\uff1a\u6838\u5fc3\u601d\u60f3\u662f\u5728\u7b56\u7565\u66f4\u65b0\u65f6\u4fdd\u6301\u65b0\u65e7\u7b56\u7565\u4e4b\u95f4\u7684\u4e00\u4e2a\u5e73\u8861\u70b9\uff0c\u5373\u201c\u8fd1\u7aef\u201d\uff08proximal\uff09\uff0c\u8fd9\u6837\u53ef\u4ee5\u9632\u6b62\u65b0\u7b56\u7565\u4e0e\u65e7\u7b56\u7565\u76f8\u5dee\u592a\u8fdc\u800c\u5bfc\u81f4\u5b66\u4e60\u4e0d\u7a33\u5b9a\u3002PPO\u91c7\u7528\u4e86\u4e00\u79cd\u79f0\u4e3a\u526a\u88c1\uff08clipping\uff09\u7684\u6280\u672f\u6765\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u5b83\u9650\u5236\u4e86\u7b56\u7565\u66f4\u65b0\u7684\u65b9\u5411\u548c\u6b65\u957f\uff0c\u4f7f\u5f97\u65b0\u7b56\u7565\u4e0d\u4f1a\u8fc7\u4e8e\u504f\u79bb\u65e7\u7b56\u7565\u3002<\/p>\n<h1>Pretraining<\/h1>\n<ul>\n<li>\n<p>\u9884\u8bad\u7ec3\u6570\u636e\uff1a<strong>2 trillion tokens<\/strong><\/p>\n<\/li>\n<li>\n<p>\u6a21\u578btrick\uff1a<\/p>\n<ul>\n<li><strong>RMSNorm<\/strong><\/li>\n<li><strong>SwiGLU<\/strong><\/li>\n<li><strong>RoPE<\/strong><\/li>\n<li><strong>GQA<\/strong> (grouped-query attention)<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u8d85\u53c2\uff1a<strong>AdamW<\/strong>\u3001\u4f59\u5f26\u8870\u51cf\u5b66\u4e60\u7387\uff08warmup=2000steps\uff0cfinal decay rate=10%\uff09\u3001weight decay*=0.1\u3001gradient clipping=1.0<\/p>\n<\/li>\n<li>\n<p>tokenizer: BPE (bytepair encoding)\uff0c\u5927\u5c0f32k tokens<\/p>\n<\/li>\n<\/ul>\n<p><strong>weight decay<\/strong>*\uff1a\u5728\u635f\u5931\u51fd\u6570\u4e2d\u52a0\u5165\u4e00\u4e2a\u989d\u5916\u7684\u9879<span class=\"katex-eq\" data-katex-display=\"false\">\\lambda \\sum_{i} w_i^2<\/span>\u6765\u60e9\u7f5a\u6743\u91cd\u5411\u91cf\u7684\u5927\u5c0f\uff0c\u4ece\u800c\u9632\u6b62\u6a21\u578b\u8fc7\u62df\u5408\u3002<\/p>\n<h1>\u9884\u8bad\u7ec3\u6a21\u578b\u8bc4\u4f30<\/h1>\n<ul>\n<li><strong>\u4ee3\u7801<\/strong>\uff1a\u8bc4\u4f30\u4e86\u6211\u4eec\u7684\u6a21\u578b\u5728HumanEval\u548cMBPP\u4e0a\u7684\u5e73\u5747pass@1\u5f97\u5206\uff1b<\/li>\n<li><strong>\u5e38\u8bc6\u63a8\u7406<\/strong>\uff1a\u6211\u4eec\u8bc4\u4f30\u4e86PIQA\u3001SIQA\u3001HellaSwag\u3001WinoGrande\u3001ARC\u7b80\u5355\u548c\u6311\u6218\u3001OpenBookQA\u548cCommonsenseQA\u7684\u5e73\u5747\u5f97\u5206\u3002\u5bf9\u4e8eCommonsenseQA\uff0c\u6211\u4eec\u8bc4\u4f30\u4e867-shot\u7ed3\u679c\uff0c\u5bf9\u4e8e\u5176\u4ed6\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\uff0c\u6211\u4eec\u8bc4\u4f30\u4e860-shot\u7ed3\u679c\uff1b<\/li>\n<li><strong>World Knowledge<\/strong>\uff1a\u6211\u4eec\u8bc4\u4f30\u4e86 NaturalQuestions \u548c TriviaQA \u7684 5-shot \u6027\u80fd\uff1b<\/li>\n<li><strong>\u9605\u8bfb\u7406\u89e3<\/strong>\uff1a\u6211\u4eec\u8bc4\u4f30\u4e86\u5728SQuAD\u3001QuAC \u548c BoolQ \u4e0a\u76840-shot\u5e73\u5747\u503c\uff1b<\/li>\n<li><strong>\u6570\u5b66<\/strong>\uff1a\u6211\u4eec\u8bc4\u4f30\u4e86GSM8K\uff088\u4e2a\u6837\u672c\uff09\u548cMATH\uff084\u4e2a\u6837\u672c\uff09\u57fa\u51c6\u6d4b\u8bd5\u7684\u5e73\u5747\u7ed3\u679c<\/li>\n<li><strong>Popular Aggregated Benchmarks<\/strong>\uff1a\u6211\u4eec\u8bc4\u4f30\u4e86MMLU\uff085\u4e2a\u6837\u672c\uff09\u3001Big Bench Hard\uff083\u4e2a\u6837\u672c\uff09\u548cAGI Eval\uff083-5\u4e2a\u6837\u672c\uff09\u7684\u6574\u4f53\u7ed3\u679c\u3002\u5bf9\u4e8eAGI Eval\uff0c\u6211\u4eec\u4ec5\u8bc4\u4f30\u82f1\u8bed\u4efb\u52a1\u5e76\u62a5\u544a\u5e73\u5747\u7ed3\u679c\u3002<br \/>\n<img src=\"https:\/\/blog.liguanxin.cn\/wp-content\/uploads\/2024\/08\/llama2-m1.png\" alt=\"\" \/><\/li>\n<\/ul>\n<p>Llama 2\u6a21\u578b\u76f8\u8f83\u4e8eLlama 1\u6a21\u578b\u6709\u663e\u8457\u63d0\u5347\u3002\u5177\u4f53\u800c\u8a00\uff0cLlama 2 70B\u6a21\u578b\u5728MMLU\u548cBBH\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u5f97\u5206\u5206\u522b\u6bd4Llama 1 65B\u6a21\u578b\u9ad8\u51fa\u7ea65\u548c8\u4e2a\u767e\u5206\u70b9\u3002\u9664\u4e86\u4ee3\u7801\u57fa\u51c6\u6d4b\u8bd5\u5916\uff0cLlama 2\u76847B\u548c30B\u7248\u672c\u5728\u5176\u5bf9\u5e94\u89c4\u6a21\u7684\u6240\u6709\u7c7b\u522b\u4e0a\u5747\u8d85\u8fc7\u4e86MPT\u6a21\u578b\u7684\u8868\u73b0\u3002<\/p>\n<p>\u540c\u65f6\uff0c\u5728\u4e0eFalcon\u7684\u5bf9\u6bd4\u4e2d\uff0cLlama 2\u76847B\u548c34B\u7248\u672c\u5728\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u7c7b\u522b\u4e2d\u4e5f\u8d85\u8d8a\u4e86Falcon 7B\u548c40B\u6a21\u578b\u3002\u6b64\u5916\uff0cLlama 2 70B\u6a21\u578b\u7684\u6027\u80fd\u8d85\u8d8a\u4e86\u6240\u6709\u5df2\u77e5\u7684\u5f00\u6e90\u6a21\u578b\u3002\u4e3a\u4e86\u8fdb\u4e00\u6b65\u8bc4\u4f30\u5176\u6027\u80fd\uff0c\u6211\u4eec\u5c06Llama 2 70B\u6a21\u578b\u4e0e\u95ed\u6e90\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002<\/p>\n<p><img src=\"https:\/\/blog.liguanxin.cn\/wp-content\/uploads\/2024\/08\/llama2-m2.png\" alt=\"\" \/><\/p>\n<p>\u6839\u636e\u88684\u7684\u6570\u636e\uff0cLlama 2 70B\u6a21\u578b\u5728MMLU\u548cGSM8K\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6210\u7ee9\u63a5\u8fd1\u4e8eGPT-3.5\uff08\u7531OpenAI\u57282023\u5e74\u53d1\u5e03\uff09\uff0c\u5c3d\u7ba1\u5728\u7f16\u7801\u57fa\u51c6\u6d4b\u8bd5\u65b9\u9762\u4ecd\u5b58\u5728\u4e00\u5b9a\u5dee\u8ddd\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u51e0\u4e4e\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cLlama 2 70B\u6a21\u578b\u7684\u8868\u73b0\u8981\u4e48\u4f18\u4e8ePaLM\uff08540B\uff09\u6a21\u578b\uff08\u7531Chowdhery\u7b49\u4eba\u57282022\u5e74\u63d0\u51fa\uff09\uff0c\u6216\u8005\u81f3\u5c11\u4e0e\u5176\u76f8\u5f53\u3002\u4e0d\u8fc7\uff0c\u4e0eGPT-4\u4ee5\u53caPaLM-2-L\u76f8\u6bd4\uff0cLlama 2 70B\u6a21\u578b\u4ecd\u7136\u5b58\u5728\u7740\u8f83\u5927\u7684\u6027\u80fd\u5dee\u8ddd\u3002<\/p>\n<h1>Fine-tuning<\/h1>\n<p>Llama 2-Chat\u662f\u7ecf\u8fc7\u6570\u6708\u5bf9\u9f50\u6280\u672f\u7814\u7a76\u548c\u8fed\u4ee3\u5e94\u7528\u7684\u6210\u679c\uff0c\u5176\u4e2d\u5305\u62ec\u6307\u4ee4\u8c03\u4f18\u548c\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u548c\u6807\u6ce8\u8d44\u6e90\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u62a5\u544a\u4f7f\u7528\u76d1\u7763\u5fae\u8c03\u3001\u521d\u59cb\u53ca\u8fed\u4ee3\u5956\u52b1\u5efa\u6a21\u4ee5\u53caRLHF\u7684\u5b9e\u9a8c\u548c\u53d1\u73b0\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5c06\u4ecb\u7ecd\u4e00\u79cd\u65b0\u6280\u672f\u2014\u2014Ghost Attention\uff08GAtt\uff09\uff0c\u6211\u4eec\u53d1\u73b0\u5b83\u6709\u52a9\u4e8e\u63a7\u5236\u591a\u8f6e\u5bf9\u8bdd\u7684\u6d41\u7a0b\u3002<\/p>\n<ul>\n<li>SFT<\/li>\n<\/ul>\n<p><strong>Quality Is All You Need<\/strong>\uff1a\u6570\u636e\u53ef\u4ece\u591a\u79cd\u6e20\u9053\u83b7\u53d6\uff0c\u4f46\u6211\u4eec\u53d1\u73b0\u8bb8\u591a\u8fd9\u7c7b\u6570\u636e\u5728\u591a\u6837\u6027\u548c\u8d28\u91cf\u4e0a\u5b58\u5728\u4e0d\u8db3\uff0c\u5c24\u5176\u662f\u5728\u8ba9\u8bed\u8a00\u6a21\u578b\u9002\u5e94\u5bf9\u8bdd\u5f0f\u6307\u4ee4\u65b9\u9762\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u6536\u96c6\u9ad8\u8d28\u91cf\u7684SFT\u6570\u636e\u793a\u4f8b\uff0c\u5982\u88685\u6240\u793a\u7684\u6570\u636e\u96c6\u3002\u901a\u8fc7\u4ece\u7b2c\u4e09\u65b9\u6570\u636e\u96c6\u4e2d\u7b5b\u9009\u51fa\u6570\u767e\u4e07\u6761\u793a\u4f8b\uff0c\u5e76\u7ed3\u5408\u6211\u4eec\u81ea\u5df1\u4f9b\u5e94\u5546\u56e2\u961f\u751f\u6210\u7684\u5c11\u91cf\u4f46\u9ad8\u8d28\u7684\u6ce8\u91ca\uff0c\u6211\u4eec\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6548\u679c\u63d0\u5347\u3002\u8fd9\u4e9b\u7ed3\u8bba\u4e0eZhou\u7b49\u4eba\u57282023\u5e74\u7684\u7814\u7a76\u76f8\u547c\u5e94\uff0c\u4ed6\u4eec\u540c\u6837\u6307\u51fa\uff0c\u6709\u9650\u89c4\u6a21\u7684\u9ad8\u8d28\u91cf\u6307\u4ee4\u8c03\u4f18\u6570\u636e\u4fbf\u80fd\u4ea7\u751f\u4f18\u79c0\u7684\u6a21\u578b\u6027\u80fd\u3002\u6211\u4eec\u53d1\u73b0\u5927\u7ea6\u6570\u4e07\u7ea7\u522b\u7684SFT\u6ce8\u91ca\u91cf\u5c31\u8db3\u4ee5\u5b9e\u73b0\u9ad8\u8d28\u91cf\u7684\u7ed3\u679c\uff1b\u5728\u7d2f\u79ef\u4e86\u603b\u8ba127,540\u4e2a\u6ce8\u91ca\u4e4b\u540e\uff0c\u6211\u4eec\u7ed3\u675f\u4e86\u8fd9\u4e00\u9636\u6bb5\u7684\u5de5\u4f5c\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\u6211\u4eec\u6ca1\u6709\u4f7f\u7528\u4efb\u4f55\u5143\u7528\u6237\u6570\u636e\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u6ce8\u610f\u5230\u4e0d\u540c\u7684\u6ce8\u91ca\u5e73\u53f0\u548c\u4f9b\u5e94\u5546\u53ef\u80fd\u5bfc\u81f4\u6a21\u578b\u6027\u80fd\u4e0a\u7684\u663e\u8457\u5dee\u5f02\uff0c\u8fd9\u5f3a\u8c03\u4e86\u5373\u4fbf\u662f\u5728\u5229\u7528\u5916\u90e8\u4f9b\u5e94\u5546\u7684\u60c5\u51b5\u4e0b\uff0c\u8fdb\u884c\u6570\u636e\u8d28\u91cf\u5ba1\u6838\u7684\u91cd\u8981\u6027\u3002\u4e3a\u4e86\u786e\u4fdd\u6570\u636e\u7684\u8d28\u91cf\uff0c\u6211\u4eec\u8be6\u7ec6\u5ba1\u67e5\u4e86180\u4e2a\u6848\u4f8b\uff0c\u5e76\u5c06\u4eba\u7c7b\u6ce8\u91ca\u7684\u7ed3\u679c\u4e0e\u6a21\u578b\u901a\u8fc7\u4eba\u5de5\u5ba1\u6838\u751f\u6210\u7684\u6837\u672c\u8fdb\u884c\u4e86\u5bf9\u6bd4\u3002\u6709\u8da3\u7684\u662f\uff0c\u6211\u4eec\u53d1\u73b0\u57fa\u4e8eSFT\u6a21\u578b\u751f\u6210\u7684\u8f93\u51fa\u5e38\u5e38\u80fd\u591f\u5ab2\u7f8e\u751a\u81f3\u8d85\u8d8a\u4eba\u5de5\u7f16\u5199\u7684SFT\u6570\u636e\uff0c\u8fd9\u63d0\u793a\u6211\u4eec\u53ef\u4ee5\u91cd\u65b0\u8c03\u6574\u7b56\u7565\uff0c\u5c06\u66f4\u591a\u7684\u6ce8\u91ca\u5de5\u4f5c\u91cd\u70b9\u653e\u5728RLHF\uff08\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff09\u7684\u504f\u597d\u6ce8\u91ca\u4e0a\u3002<\/p>\n<p><strong>Fine-Tuning Details<\/strong>\uff1a\u5fae\u8c03\u8fc7\u7a0b\u91c7\u7528\u4f59\u5f26\u5b66\u4e60\u7387\uff0c\u521d\u59cb\u5b66\u4e60\u7387\u4e3a2\u00d710-5\u3002\u6743\u91cd\u8870\u51cf\u4e3a0.1\uff0cbatch size is 64, a sequence length of 4096 tokens.<\/p>\n<ul>\n<li>RLHF<\/li>\n<\/ul>\n<p>RLHF\uff08Reinforcement Learning with Human Feedback\uff0c\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff09\u662f\u4e00\u79cd\u6a21\u578b\u8bad\u7ec3\u6d41\u7a0b\uff0c\u5b83\u5728\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u57fa\u7840\u4e0a\u8fdb\u4e00\u6b65\u5b9e\u65bd\u5fae\u8c03\uff0c\u76ee\u7684\u662f\u4f7f\u6a21\u578b\u7684\u884c\u4e3a\u66f4\u52a0\u7b26\u5408\u4eba\u7c7b\u7684\u504f\u597d\u5e76\u51c6\u786e\u9075\u5faa\u7ed9\u5b9a\u7684\u6307\u4ee4\u3002\u5728\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u6240\u6536\u96c6\u7684\u6570\u636e\u53cd\u6620\u4e86\u62bd\u6837\u5f97\u51fa\u7684\u4eba\u7c7b\u504f\u597d\uff0c\u5177\u4f53\u505a\u6cd5\u662f\u7531\u4eba\u7c7b\u6ce8\u91ca\u5458\u5728\u6a21\u578b\u7ed9\u51fa\u7684\u4e24\u4e2a\u8f93\u51fa\u9009\u9879\u4e2d\u6311\u9009\u51fa\u4ed6\u4eec\u8ba4\u4e3a\u66f4\u4f73\u7684\u4e00\u4e2a\u3002\u8fd9\u4e9b\u4eba\u7c7b\u7684\u9009\u62e9\u53cd\u9988\u968f\u540e\u88ab\u7528\u6765\u8bad\u7ec3\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\uff0c\u8be5\u5956\u52b1\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u5e76\u7406\u89e3\u4eba\u7c7b\u6ce8\u91ca\u5458\u7684\u504f\u597d\u89c4\u5f8b\uff0c\u4ece\u800c\u5728\u672a\u6765\u4e0d\u9700\u8981\u76f4\u63a5\u7684\u4eba\u7c7b\u5e72\u9884\u4e0b\uff0c\u4e5f\u80fd\u4f9d\u636e\u6240\u5b66\u7684\u504f\u597d\u6a21\u5f0f\u81ea\u52a8\u505a\u51fa\u51b3\u7b56\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u6548\u5730\u5c06\u4eba\u7c7b\u7684\u4e3b\u89c2\u5224\u65ad\u8f6c\u5316\u4e3a\u6a21\u578b\u7684\u4f18\u5316\u65b9\u5411\uff0c\u8fdb\u800c\u63d0\u5347\u4e86\u6a21\u578b\u751f\u6210\u5185\u5bb9\u7684\u8d28\u91cf\u548c\u76f8\u5173\u6027\u3002<\/p>\n<p><strong> Human Preference Data Collection<\/strong>\uff1a\u63a5\u4e0b\u6765\u6536\u96c6\u7528\u4e8e\u5956\u52b1\u5efa\u6a21\u7684\u4eba\u7c7b\u504f\u597d\u6570\u636e\u3002\u672c\u6a21\u578b\u9009\u62e9\u4e86\u4e8c\u5143\u6bd4\u8f83\u534f\u8bae\u800c\u4e0d\u662f\u5176\u4ed6\u65b9\u6848\uff0c\u4e3b\u8981\u662f\u56e0\u4e3a\u5b83\u4f7f\u6211\u4eec\u80fd\u591f\u6700\u5927\u5316\u6536\u96c6\u5230\u7684\u63d0\u793a\u7684\u591a\u6837\u6027\u3002<\/p>\n<p>\u6211\u4eec\u7684\u6807\u6ce8\u7a0b\u5e8f\u5982\u4e0b\uff1a\u9996\u5148\u8981\u6c42\u6807\u6ce8\u8005\u5199\u4e0b\u4e00\u4e2a\u63d0\u793a\uff0c\u7136\u540e\u6839\u636e\u63d0\u4f9b\u7684\u6807\u51c6\u5728\u8fd9\u4e24\u4e2a\u6a21\u578b\u6837\u672c\u54cd\u5e94\u4e4b\u95f4\u4f5c\u51fa\u9009\u62e9\u3002\u4e3a\u4e86\u6700\u5927\u5316\u591a\u6837\u6027\uff0c\u5bf9\u7ed9\u5b9a\u63d0\u793a\u7684\u4e24\u4e2a\u54cd\u5e94\u662f\u4ece\u4e24\u4e2a\u4e0d\u540c\u7684\u6a21\u578b\u53d8\u4f53\u4e2d\u91c7\u6837\u7684\uff0c\u5e76\u4e14\u901a\u8fc7\u6539\u53d8temperature\u8d85\u53c2\u6570\u6765\u5b9e\u73b0\u3002\u9664\u4e86\u8ba9\u53c2\u4e0e\u8005\u8fdb\u884c\u5f3a\u5236\u9009\u62e9\u5916\uff0c\u6211\u4eec\u8fd8\u8981\u6c42\u6807\u6ce8\u8005\u6807\u8bb0\u4ed6\u4eec\u5bf9\u5176\u6240\u9009\u54cd\u5e94\u7684\u504f\u597d\u7a0b\u5ea6\uff1a\u5373\u4ed6\u4eec\u7684\u9009\u62e9\u662f\u663e\u8457\u66f4\u597d\u3001\u8f83\u597d\u3001\u7565\u597d\u8fd8\u662f\u51e0\u4e4e\u6ca1\u5dee\u522b\/\u4e0d\u786e\u5b9a\u3002<\/p>\n<p>\u5728\u6536\u96c6\u504f\u597d\u6807\u6ce8\u65f6\uff0c\u6211\u4eec\u5173\u6ce8\u7684\u662f\u6709\u7528\u6027\u548c\u5b89\u5168\u6027\u3002\u6709\u7528\u6027\u662f\u6307Llama 2-Chat\u7684\u54cd\u5e94\u5982\u4f55\u6ee1\u8db3\u7528\u6237\u7684\u8bf7\u6c42\u5e76\u63d0\u4f9b\u6240\u9700\u7684\u4fe1\u606f\uff1b\u5b89\u5168\u6027\u662f\u6307Llama 2-Chat\u7684\u54cd\u5e94\u662f\u5426\u5b89\u5168\uff0c\u4f8b\u5982\uff0c\u201c\u63d0\u4f9b\u8be6\u7ec6\u7684\u70b8\u5f39\u5236\u4f5c\u8bf4\u660e\u201d\u53ef\u80fd\u662f\u6709\u7528\u7684\uff0c\u4f46\u6839\u636e\u6211\u4eec\u7684\u5b89\u5168\u6307\u5357\u8fd9\u662f\u4e0d\u5b89\u5168\u7684\u3002\u5c06\u4e8c\u8005\u5206\u5f00\u53ef\u4ee5\u4f7f\u6211\u4eec\u9488\u5bf9\u6bcf\u4e00\u79cd\u60c5\u51b5\u5e94\u7528\u7279\u5b9a\u7684\u6307\u5bfc\u65b9\u9488\uff0c\u66f4\u597d\u5730\u6307\u5bfc\u6807\u6ce8\u8005\uff1b\u4f8b\u5982\uff0c\u6211\u4eec\u7684\u5b89\u5168\u6807\u6ce8\u63d0\u4f9b\u4e86\u96c6\u4e2d\u4e8e\u5bf9\u6297\u6027\u63d0\u793a\u7684\u6307\u793a\u7b49\u5176\u4ed6\u6307\u5bfc\u3002<\/p>\n<p>\u4eba\u7c7b\u6807\u6ce8\u662f\u6309\u5468\u5206\u6279\u6536\u96c6\u7684\u3002\u968f\u7740\u6211\u4eec\u6536\u96c6\u4e86\u66f4\u591a\u7684\u504f\u597d\u6570\u636e\uff0c\u6211\u4eec\u7684\u5956\u52b1\u6a21\u578b\u5f97\u5230\u4e86\u6539\u8fdb\uff0c\u6211\u4eec\u4e5f\u80fd\u591f\u8bad\u7ec3\u51fa\u6e10\u8fdb\u5f0f\u66f4\u597d\u7684Llama 2-Chat\u7248\u672c\u3002Llama 2-Chat\u7684\u6539\u8fdb\u4e5f\u6539\u53d8\u4e86\u6a21\u578b\u7684\u6570\u636e\u5206\u5e03\u3002\u7531\u4e8e\u5982\u679c\u5956\u52b1\u6a21\u578b\u6ca1\u6709\u63a5\u89e6\u5230\u8fd9\u79cd\u65b0\u7684\u6837\u672c\u5206\u5e03\uff0c\u5176\u51c6\u786e\u6027\u4f1a\u8fc5\u901f\u4e0b\u964d\uff0c\u56e0\u6b64\u5728\u65b0\u7684Llama 2-Chat\u8c03\u4f18\u8fed\u4ee3\u4e4b\u524d\u6536\u96c6\u4f7f\u7528\u6700\u65b0Llama 2-Chat\u8fed\u4ee3\u7684\u65b0\u504f\u597d\u6570\u636e\u662f\u5f88\u91cd\u8981\u7684\u3002\u8fd9\u4e00\u6b65\u9aa4\u6709\u52a9\u4e8e\u4fdd\u6301\u5956\u52b1\u6a21\u578b\u4e0e\u5206\u5e03\u4e00\u81f4\uff0c\u5e76\u4e3a\u6700\u65b0\u6a21\u578b\u4fdd\u6301\u51c6\u786e\u7684\u5956\u52b1\u3002<\/p>\n<p>\u672c\u6587\u6536\u96c6\u4e86\u4e00\u4e2a\u5927\u578b\u6570\u636e\u96c6\uff0c\u5305\u542b\u8d85\u8fc7100\u4e07\u4e2a\u57fa\u4e8e\u4eba\u7c7b\u5e94\u7528\u4ed6\u4eec\u6307\u5b9a\u7684\u6307\u5357\u7684\u4e8c\u5143\u6bd4\u8f83\uff0c\u5c06\u5176\u79f0\u4e3aMeta\u5956\u52b1\u5efa\u6a21\u6570\u636e\u3002\u4e0e\u73b0\u6709\u7684\u5f00\u6e90\u6570\u636e\u96c6\u76f8\u6bd4\uff0c\u672c\u6570\u636e\u96c6\u7684\u504f\u597d\u6570\u636e\u5177\u6709\u66f4\u591a\u7684\u5bf9\u8bdd\u8f6e\u6b21\uff0c\u5e76\u4e14\u5e73\u5747\u957f\u5ea6\u66f4\u957f\u3002<\/p>\n<ul>\n<li>\u5956\u52b1\u5efa\u6a21<\/li>\n<\/ul>\n<p>\u5956\u52b1\u6a21\u578b\u63a5\u53d7\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\u548c\u76f8\u5e94\u7684\u63d0\u793a\uff08\u5305\u62ec\u4e4b\u524d\u5bf9\u8bdd\u7684\u4e0a\u4e0b\u6587\uff09\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u8f93\u51fa\u4e00\u4e2a\u6807\u91cf\u5206\u6570\uff0c\u4ee5\u6307\u793a\u6a21\u578b\u751f\u6210\u7684\u8d28\u91cf\uff08\u4f8b\u5982\uff0c\u6709\u7528\u6027\u548c\u5b89\u5168\u6027\uff09\u3002\u901a\u8fc7\u5c06\u8fd9\u4e9b\u56de\u590d\u5206\u6570\u4f5c\u4e3a\u5956\u52b1\uff0c\u6211\u4eec\u53ef\u4ee5\u5728RLHF\u8fc7\u7a0b\u4e2d\u4f18\u5316Llama 2-Chat\uff0c\u4ee5\u5b9e\u73b0\u66f4\u597d\u7684\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u548c\u63d0\u9ad8\u6709\u7528\u6027\u548c\u5b89\u5168\u6027\u3002<\/p>\n<p>\u6709\u7814\u7a76\u53d1\u73b0\uff0c\u6709\u65f6\u5728\u5e2e\u52a9\u6027\u548c\u5b89\u5168\u6027\u4e4b\u95f4\u5b58\u5728\u6743\u8861\uff0c\u8fd9\u4f7f\u5f97\u5355\u4e00\u5956\u52b1\u6a21\u578b\u5728\u4e24\u8005\u4e0a\u8868\u73b0\u826f\u597d\u53d8\u5f97\u5177\u6709\u6311\u6218\u6027\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u8bad\u7ec3\u4e86\u4e24\u4e2a\u72ec\u7acb\u7684\u5956\u52b1\u6a21\u578b\uff0c\u4e00\u4e2a\u9488\u5bf9\u5e2e\u52a9\u6027\uff08\u79f0\u4e3aHelpfulness RM\uff09\uff0c\u53e6\u4e00\u4e2a\u9488\u5bf9\u5b89\u5168\u6027\uff08Safety RM\uff09\u3002\u6211\u4eec\u4ece\u9884\u8bad\u7ec3\u7684\u804a\u5929\u6a21\u578b\u68c0\u67e5\u70b9\u521d\u59cb\u5316\u5956\u52b1\u6a21\u578b\uff0c\u4ee5\u786e\u4fdd\u4e24\u4e2a\u6a21\u578b\u90fd\u80fd\u4ece\u9884\u8bad\u7ec3\u4e2d\u83b7\u5f97\u7684\u77e5\u8bc6\u7684\u597d\u5904\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u5956\u52b1\u6a21\u578b\u201c\u77e5\u9053\u201d\u804a\u5929\u6a21\u578b\u6240\u77e5\u9053\u7684\u5185\u5bb9\u3002\u6a21\u578b\u67b6\u6784\u548c\u8d85\u53c2\u6570\u4e0e\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u76f8\u540c\uff0c\u53ea\u662f\u5c06\u4e0b\u4e00\u4e2a\u6807\u8bb0\u9884\u6d4b\u7684 classification head \u66ff\u6362\u4e3a\u7528\u4e8e\u8f93\u51fa\u6807\u91cf\u5956\u52b1\u7684 regression head\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Introduction \u5f00\u6e90\u7248\u672c\uff1a Llama2 7B 13B 70B Llama 2-Chat 7B 13 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":636,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[36,35],"tags":[37,38],"_links":{"self":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/634"}],"collection":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/comments?post=634"}],"version-history":[{"count":0,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/634\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/media\/636"}],"wp:attachment":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/media?parent=634"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/categories?post=634"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/tags?post=634"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}