

{"id":397,"date":"2025-11-25T09:12:46","date_gmt":"2025-11-25T01:12:46","guid":{"rendered":"https:\/\/high-flyer.in.suopu.cc\/?p=397"},"modified":"2025-11-25T09:12:46","modified_gmt":"2025-11-25T01:12:46","slug":"llama-2-%e6%8a%80%e6%9c%af%e8%af%a6%e8%a7%a3%ef%bc%88%e4%b8%80%ef%bc%89%ef%bc%9a%e6%95%b0%e6%8d%ae%e6%89%93%e6%a0%87","status":"publish","type":"post","link":"https:\/\/high-flyer.in.suopu.cc\/en\/blog\/397\/","title":{"rendered":"LLaMA-2 \u6280\u672f\u8be6\u89e3\uff08\u4e00\uff09\uff1a\u6570\u636e\u6253\u6807"},"content":{"rendered":"<p>LLaMA \u662f\u76ee\u524d\u5907\u53d7\u5173\u6ce8\u7684\u5f00\u6e90\u5927\u8bed\u8a00\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u6700\u8fd1 Meta \u53d1\u5e03\u4e86 LLaMA 2\uff0c\u5b83\u662f LLaMA \u7684\u4e0b\u4e00\u4ee3\u7248\u672c\uff0c\u5177\u6709\u5546\u4e1a\u53cb\u597d\u7684\u8bb8\u53ef\u8bc1\u3002LLaMA 2 \u6709 3 \u79cd\u4e0d\u540c\u7684\u5c3a\u5bf8\uff1a7B\u300113B \u548c 70B\uff0c\u5728\u5bf9\u8bdd\u7528\u4f8b\u4e0b\u5fae\u8c03\u7684\u6a21\u578b\u7248\u672c\u88ab\u79f0\u4e3a LLaMA 2-Chat\u3002Llama-2-Chat \u6a21\u578b\u5728\u5927\u591a\u6570\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u90fd\u4f18\u4e8e\u73b0\u6709\u7684\u5f00\u6e90\u804a\u5929\u6a21\u578b\uff0c\u5728\u5bf9\u6709\u7528\u6027\u548c\u5b89\u5168\u6027\u7684\u4eba\u5de5\u8bc4\u4f30\u4e2d\uff0cLLama-2-Chat \u4e0e ChatGPT\uff0cPaLM \u7b49\u4e00\u4e9b\u6d41\u884c\u7684\u5c01\u95ed\u6e90\u6a21\u578b\u4e0d\u76f8\u4e0a\u4e0b\u3002<\/p>\n<p>LLama-2-Chat \u51e0\u4e4e\u662f\u73b0\u6709\u5f00\u6e90\u6a21\u578b\u4e2d\u552f\u4e00\u505a\u4e86 RLHF \u7684\u6a21\u578b\u3002LLaMA-2 \u7ecf\u8fc7 5 \u8f6e RLHF \u540e\uff0c\u5728 Meta \u81ea\u5df1\u7684 Reward \u6a21\u578b\u4e0e GPT-4 \u7684\u8bc4\u4ef7\u4e0b\uff0c\u90fd\u8868\u73b0\u51fa\u4e86\u8d85\u8fc7 ChatGPT \u6027\u80fd\u3002<\/p>\n<p>\u672c\u6587\u5c06\u4e3a\u5927\u5bb6\u6574\u7406 LLaMA-2 \u8bba\u6587\u4e2d\u63d0\u5230\u7528\u6765\u8bad\u7ec3 Reward Model \u7684\u6570\u636e\u83b7\u53d6\u6d41\u7a0b\u3002<\/p>\n<h2 id=\"\u6807\u6ce8\u6d41\u7a0b\">\u6807\u6ce8\u6d41\u7a0b<\/h2>\n<p>LLaMA-2 \u7684\u6253\u6807\u4e3b\u8981\u8003\u8651\u6709\u7528\u6027(Helpfulness)\u548c\u5b89\u5168\u6027(Safety)<\/p>\n<ul>\n<li>\u6709\u7528\u6027(Helpfulness)\uff1aresponse \u5bf9 prompt \u4e2d\u95ee\u9898\u56de\u7b54\u5f97\u6709\u591a\u597d\uff0c\u662f\u5426\u63d0\u4f9b\u4e86\u9700\u8981\u7684\u4fe1\u606f<\/li>\n<li>\u5b89\u5168\u6027(Safety)\uff1a\u56de\u7b54\u662f\u5426\u662f\u5b89\u5168\u7684\uff0c\u4f8b\u5982\u6559\u7528\u6237\u9020\u70b8\u5f39\uff0c\u662f\u4e0d\u5b89\u5168\u7684<\/li>\n<\/ul>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/29bcbd03f5cecc6c9fe1bcd3d13f2bc6\/302a4\/01.png\" target=\"_blank\" rel=\"noopener\"><img fetchpriority=\"high\" decoding=\"async\" class=\"alignnone size-full wp-image-398\" title=\"96a3be3cf272e017046d89\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d89.png\" alt=\"96a3be3cf272e017046d89\" width=\"650\" height=\"308\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d89.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d89-300x142.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d89-150x71.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u5bf9 LLaMA-2 \u9884\u8bad\u7ec3\u6a21\u578b\u7ecf\u8fc7\u5bf9\u8bdd\u7528\u4f8b\u5fae\u8c03\u4e4b\u540e\uff0c\u7531\u4eba\u53cd\u9988\u53c2\u4e0e\u7684 RLHF \u6301\u7eed\u63d0\u5347\u6a21\u578b\u7684\u80fd\u529b\uff0c\u5176\u6253\u6807\u6d41\u7a0b\u4e3b\u8981\u6709\u5982\u4e0b\u51e0\u4e2a\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li>\u6807\u6ce8\u8005\u5199\u4e00\u4e2a prompt<\/li>\n<li>\u8c03\u7528\u6a21\u578b\uff0c\u968f\u673a\u53c2\u6570\uff0c\u751f\u6210\u4e24\u4e2a\u56de\u7b54\uff1aresponse_1, response_2<\/li>\n<li>\u6807\u6ce8\u8005\u9488\u5bf9\u8fd9\u4e24\u4e2a\u56de\u7b54\u6807\u6ce8\u5982\u4e0b\u4fe1\u606f\uff1a\n<p>a. \u54ea\u4e2a\u56de\u7b54\u66f4\u597d\uff1aresponse_1 \u6216\u8005 response_2<\/p>\n<p>b. \u597d\u7684\u7a0b\u5ea6\uff1a\u975e\u5e38\u663e\u8457\u7684\u597d\uff0c\u597d\u4e00\u70b9\uff0c\u4e0d\u786e\u5b9a\uff08\u5ffd\u7565\u4e0d\u8ba1\u7684\u597d\uff09<\/p>\n<p>c. response_1 \u662f\u5426\u5b89\u5168<\/p>\n<p>d. response_2 \u662f\u5426\u5b89\u5168<\/p>\n<p>e. \u54ea\u4e2a\u56de\u7b54\u66f4\u5b89\u5168\uff1aresponse_1 \u6216\u8005 response_2<\/li>\n<\/ol>\n<p>\u6700\u7ec8\u6bcf\u4e00\u6761\u6807\u8bb0\u6570\u636e\u7ec4\u7ec7\u6210\u5982\u4e0b\uff1a<\/p>\n<div class=\"gatsby-highlight\" data-language=\"python\">\n<pre class=\"language-python\"><code class=\"language-python\"><span class=\"token punctuation\">{<\/span>\r\n<span class=\"token string\">'prompt'<\/span><span class=\"token punctuation\">:<\/span>P<span class=\"token punctuation\">,<\/span> <span class=\"token comment\"># \u5982\u679c\u662f\u591a\u8f6e\u5bf9\u8bdd\u6570\u636e\uff0c\u5219\u9664\u4e86\u6700\u540e\u4e00\u8f6e\u7684\u56de\u7b54\uff0c\u524d\u6587\u7684\u95ee\u9898\u548c\u56de\u7b54\u90fd\u5728prompt\u91cc<\/span>\r\n<span class=\"token string\">'response_1'<\/span><span class=\"token punctuation\">:<\/span>R1<span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'response_2'<\/span><span class=\"token punctuation\">:<\/span>R2<span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'chosen_response'<\/span><span class=\"token punctuation\">:<\/span><span class=\"token number\">1<\/span> <span class=\"token keyword\">or<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'prefer_degree'<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">'significantly better'<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">'better'<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">'slightly better'<\/span> <span class=\"token keyword\">or<\/span> <span class=\"token string\">'negligibly better\/unsure'<\/span><span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'response_1_safe'<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">'yes'<\/span> <span class=\"token keyword\">or<\/span> <span class=\"token string\">'no'<\/span><span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'response_2_safe'<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">'yes'<\/span> <span class=\"token keyword\">or<\/span> <span class=\"token string\">'no'<\/span><span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'safer_response'<\/span><span class=\"token punctuation\">:<\/span><span class=\"token number\">1<\/span> <span class=\"token keyword\">or<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span>\r\n\r\n<span class=\"token string\">'\u6253\u5206\u4eba'<\/span><span class=\"token punctuation\">:<\/span> xxx<span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'response_1_model'<\/span><span class=\"token punctuation\">:<\/span> model_name1<span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'response_2_model'<\/span><span class=\"token punctuation\">:<\/span> model_name2\uff0c\r\n<span class=\"token string\">'\u6807\u6ce8\u65f6\u95f4'<\/span><span class=\"token punctuation\">:<\/span> date<span class=\"token punctuation\">,<\/span>\r\n<span class=\"token string\">'\u6807\u6ce8\u7684\u8d85\u53c2'<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token punctuation\">{<\/span>\r\n    <span class=\"token string\">'\u6700\u957f\u751f\u6210\u957f\u5ea6'<\/span><span class=\"token punctuation\">:<\/span> max_len\r\n    <span class=\"token string\">'temperature'<\/span><span class=\"token punctuation\">:<\/span> temperature\r\n    <span class=\"token string\">'\u591a\u8f6e\u5bf9\u8bdd\u7684\u8d85\u53c2'<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token punctuation\">]<\/span>\r\n    <span class=\"token punctuation\">.<\/span><span class=\"token punctuation\">.<\/span><span class=\"token punctuation\">.<\/span>\r\n    <span class=\"token punctuation\">}<\/span>\r\n<span class=\"token punctuation\">}<\/span><\/code><\/pre>\n<\/div>\n<p>\u8fd9\u91cc\uff0c\u5728\u8c03\u7528\u6a21\u578b\u751f\u6210\u4e24\u4e2a\u56de\u7b54\u65f6\uff0cLLaMA-2 \u4f1a\u5c3d\u53ef\u80fd\u9009\u62e9\u5dee\u5f02\u5927\u7684\u53c2\u6570\uff0c\u4f7f\u5f97\u751f\u6210\u7684\u4e24\u4e2a\u56de\u7b54\u5c3d\u53ef\u80fd\u4e0d\u540c\uff0c\u65b9\u4fbf\u7528\u6237\u8fdb\u884c\u51c6\u786e\u7684\u6253\u6807\u53cd\u9988\u3002<\/p>\n<h2 id=\"\u5199-prompt-\u4ee5\u53ca\u6253\u5206\u7684\u601d\u8def\">\u5199 Prompt \u4ee5\u53ca\u6253\u5206\u7684\u601d\u8def<\/h2>\n<p>\u4e3a\u4e86\u8ba9 RLHF \u8bad\u7ec3\u80fd\u66f4\u6709\u6548\u7684\u63d0\u5347\u6a21\u578b\u7684\u80fd\u529b\uff0cLLaMA-2 \u8bba\u6587\u91cc\u63d0\u5230\u4e86\u82e5\u5e72 Prompt \u6784\u9020\u4ee5\u53ca\u6253\u5206\u7684 Tips\u3002<\/p>\n<h3 id=\"\u5bf9\u6297-prompt-\u7684\u7f16\u5199\">\u5bf9\u6297 Prompt \u7684\u7f16\u5199<\/h3>\n<p>\u9664\u4e86\u6b63\u5e38\u7528\u6237\u4ea4\u4e92\u7684 Prompt\uff0c\u4e3a\u4e86\u4ea7\u751f\u66f4\u591a\u5177\u6709\u4ee3\u8868\u6027\u7684\u6570\u636e\uff0cLLaMA-2 \u4f1a\u8ba9\u6807\u6ce8\u8005\u5199\u4e86\u4e00\u4e9b\u6709\u5bf9\u6297\u6027\u7684 Prompt\uff1a<\/p>\n<ul>\n<li>Risk category\uff1a\u53ef\u80fd\u4f1a\u8ba9 LLM \u4ea7\u751f\u4e0d\u5b89\u5168\u5185\u5bb9\u7684\u8bdd\u9898\uff1a\n<ul>\n<li>\u975e\u6cd5\u548c\u5211\u4e8b\u6d3b\u52a8\uff08\u4f8b\u5982\u6050\u6016\u4e3b\u4e49\u3001\u76d7\u7a83\u3001\u4eba\u53e3\u8d29\u8fd0\uff09<\/li>\n<li>\u4ec7\u6068\u548c\u6709\u5bb3\u6d3b\u52a8\uff08\u4f8b\u5982\uff0c\u8bfd\u8c24\u3001\u81ea\u6b8b\u3001\u996e\u98df\u969c\u788d\u3001\u6b67\u89c6\uff09<\/li>\n<li>\u4e0d\u5408\u683c\u7684\u5efa\u8bae\uff08\u4f8b\u5982\uff0c\u533b\u7597\u5efa\u8bae\u3001\u8d22\u52a1\u5efa\u8bae\u3001\u6cd5\u5f8b\u5efa\u8bae\uff09<\/li>\n<\/ul>\n<\/li>\n<li>Attack vector\uff1a\u53ef\u80fd\u8ba9 LLM \u4ea7\u751f\u9519\u8bef\u884c\u4e3a\u7684\u63d0\u793a\uff1a\n<ul>\n<li>\u5fc3\u7406\u64cd\u4f5c\uff08\u4f8b\u5982\u6743\u5a01\u64cd\u4f5c\uff09<\/li>\n<li>\u903b\u8f91\u64cd\u4f5c\uff08\u4f8b\u5982\u9519\u8bef\u524d\u63d0\uff09<\/li>\n<li>\u53e5\u6cd5\u64cd\u4f5c\uff08\u4f8b\u5982\u62fc\u5199\u9519\u8bef\uff09<\/li>\n<li>\u8bed\u4e49\u64cd\u4f5c\uff08\u4f8b\u5982\u9690\u55bb\uff09<\/li>\n<li>\u89c6\u89d2\u64cd\u4f5c\uff08\u4f8b\u5982\u89d2\u8272\u626e\u6f14\uff09<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3 id=\"\u5bf9\u6297-prompt-\u7684\u6253\u5206\">\u5bf9\u6297 Prompt \u7684\u6253\u5206<\/h3>\n<p>\u5bf9\u4e8e\u4e0a\u9762\u5199\u5f97\u5bf9\u6297\u6027 Prompt\uff0c\u6a21\u578b\u9700\u8981\u6309\u7167\u4e0b\u9762\u7684\u6d41\u7a0b\u56de\u7b54\uff0c\u624d\u80fd\u7b97\u4f5c\u597d\u7684\u56de\u7b54\uff1a<\/p>\n<ul>\n<li>\u9996\u5148\u89e3\u51b3\u5b89\u5168\u95ee\u9898<\/li>\n<li>\u89e3\u91ca\u7528\u6237\u53ef\u80fd\u9762\u4e34\u7684\u6f5c\u5728\u98ce\u9669<\/li>\n<li>\u5728\u53ef\u80fd\u7684\u60c5\u51b5\u4e0b\u5c3d\u53ef\u80fd\u63d0\u4f9b\u5176\u4ed6\u6709\u7528\u4fe1\u606f<\/li>\n<li>\u907f\u514d\u8d1f\u9762\u7684\u7528\u6237\u4f53\u9a8c<\/li>\n<\/ul>\n<p>\u8d1f\u9762\u7684\u7528\u6237\u4f53\u9a8c\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u4e0d\u8981\u9f13\u52b1\u6216\u8005\u652f\u6301\u72af\u7f6a\u6d3b\u52a8<\/li>\n<li>\u4e0d\u8981\u9f13\u52b1\u6216\u652f\u6301\u7528\u6237\u81ea\u5df1\u6216\u8005\u5bf9\u4ed6\u4eba\u8fdb\u884c\u5371\u9669\u884c\u4e3a<\/li>\n<li>\u4e0d\u8981\u5305\u542b\u3001\u9f13\u52b1\u6216\u652f\u6301\u5bf9\u7528\u6237\u6216\u5176\u4ed6\u4eba\u7684\u653b\u51fb\u6027\u548c\u8fb1\u9a82\u884c\u4e3a<\/li>\n<li>\u4e0d\u8981\u5305\u542b\u3001\u9f13\u52b1\u6216\u652f\u6301\u660e\u786e\u7684\u6027\u522b\u76f8\u5173\u5185\u5bb9<\/li>\n<\/ul>\n<h2 id=\"\u5176\u4ed6\u7ec6\u8282\">\u5176\u4ed6\u7ec6\u8282<\/h2>\n<p><strong>1. \u4e24\u4e2a\u56de\u7b54\u4ece\u4ec0\u4e48\u6a21\u578b\u6765\uff1f<\/strong><\/p>\n<p>\u4e24\u4e2a\u56de\u7b54\uff08Response\uff09\u7531\u4e0d\u540c\u7684\u6a21\u578b\u53d8\u4f53\u3001\u4e0d\u540c\u7684\u751f\u6210\u6e29\u5ea6\uff08Temperature\uff09\u751f\u6210\uff0c\u76ee\u7684\u662f\u589e\u52a0\u56de\u7b54\u591a\u6837\u6027\u3002<\/p>\n<p>\u4f46\u662f\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u91cc\u9009\u62e9\u7684\u6a21\u578b\uff0c\u5e94\u8be5\u662f\u6b63\u5728\u8bad\u7ec3\u7684\u6700\u65b0\u6a21\u578b\u548c\u4e0a\u4e00\u7248\u672c\u7684\u6a21\u578b\uff0c\u4ee5\u4fdd\u8bc1 Reward Model \u662f\u5728\u6700\u65b0\u7684\u6570\u636e\u5206\u5e03\u8fdb\u884c\u6253\u5206\uff0c\u7ed9\u5f53\u524d\u6a21\u578b\u8f93\u51fa\u7684\u6253\u5206\u8db3\u591f\u51c6\u786e\u3002<\/p>\n<p>\u8bba\u6587\u91cc\u628a\u8fd9\u4e2a\u8fc7\u7a0b\u79f0\u4e3a Iterative reward modeling data\uff0c\u662f\u4fdd\u8bc1 Reward \u6a21\u578b\u5728\u6700\u65b0\u6570\u636e\u5206\u5e03\u4e0a\u4f7f\u7528\u5f53\u524d\u7248\u672c\u6a21\u578b\u548c\u4e0a\u4e00\u4e2a\u8fed\u4ee3\u7248\u672c\u7684\u6a21\u578b\u8fdb\u884c\u91c7\u6837\u7684\u5173\u952e\uff0c\u540c\u65f6\u8fd8\u53ef\u4ee5\u8fdb\u884c\u4e24\u4e2a\u7248\u672c\u6a21\u578b\u7684\u8bc4\u4f30\u6bd4\u8f83\u3002<\/p>\n<p><strong>2. \u4e3a\u4ec0\u4e48\u53ea\u751f\u6210\u4e24\u4e2a\u56de\u7b54\uff1f<\/strong><\/p>\n<p>InstructGPT \u662f\u7ed9\u5b9a\u4e00\u4e2a Prompt \u751f\u6210 K \u4e2a\u56de\u7b54\uff08K \u901a\u5e38\u4e3a6\u5de6\u53f3\uff09\uff0c\u8ba9\u6807\u6ce8\u8005\u6392\u5e8f\u3002\u4e0e\u4e4b\u5bf9\u6bd4\uff0cLLaMA-2 \u7684\u6807\u6ce8\u65b9\u5f0f\u6709\u5982\u4e0b\u597d\u5904\uff1a<\/p>\n<ul>\n<li>\u76f8\u540c\u6807\u6ce8\u91cf\u4e0b\uff0cPrompt \u591a\u6837\u6027\u66f4\u9ad8\uff1b<\/li>\n<li>\u540c\u4e00\u4e2a Prompt\uff0c\u6807\u6ce8\u5458\u7684\u6807\u6ce8\u96be\u5ea6\u66f4\u4f4e<\/li>\n<\/ul>\n<p><strong>3. LLaMA-2 \u6570\u636e\u91cf\u6709\u591a\u5927\uff1f<\/strong><\/p>\n<p>LLaMA-2 \u6bcf\u5468\u66f4\u65b0\u4e00\u6b21\u6807\u6ce8\u6570\u636e\u3002\u5982\u4e0b\u8868\u6240\u793a\uff0c\u6bcf\u4e00\u884c\u662f\u4e00\u5468\u65b0\u6807\u6ce8\u7684\u6570\u636e\uff0cLLaMA-2 \u603b\u5171\u7ea6\u6709\u767e\u4e07\u6761\u81ea\u5df1\u6807\u6ce8\u7684\u6570\u636e\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/adbb7a59002bb39f3e26eebfa59fcc9d\/302a4\/02.png\" target=\"_blank\" rel=\"noopener\"><img decoding=\"async\" class=\"alignnone size-full wp-image-401\" title=\"a2ef406e2c2351e0b9e819\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e819.png\" alt=\"a2ef406e2c2351e0b9e819\" width=\"650\" height=\"358\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e819.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e819-300x165.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e819-150x83.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u8868\u683c\u542b\u4e49\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>batch\uff1a\u7b2c\u51e0\u5468<\/li>\n<li>Num comparisons\uff1a\u6709\u591a\u5c11\u6761\u6570\u636e\uff0c\u6bcf\u4e00\u6761\u6570\u636e\u5982\u4e0a\u6587\u6240\u8ff0\uff0c\u5305\u62ec\u4e86\u4e00\u4e2a\u95ee\u9898\uff0c\u4e24\u4e2a\u56de\u7b54\uff0c\u4ee5\u53ca\u6807\u6ce8<\/li>\n<li>Avg Turns\uff1a\u5e73\u5747\u5bf9\u8bdd\u8f6e\u6570<\/li>\n<li>Avg Tokens\uff1a\u6bcf\u6761\u6837\u672c\u4e2d\u7684\u5e73\u5747 Token \u6570\u91cf<\/li>\n<li>Avg Tokens in prompt\uff1aPrompt \u91cc\u7684\u5e73\u5747 Token \u6570\u91cf<\/li>\n<li>Avg Tokens in response\uff1aResponse \u91cc\u7684\u5e73\u5747 Toekn \u6570\u91cf<\/li>\n<\/ul>\n<p>\u603b\u5171 141w \u6761\u6570\u636e\uff0c\u4f46\u662f\u6bcf\u5468\u90fd\u66f4\u6362\u4e86\u751f\u6210\u56de\u7b54\u7684\u6a21\u578b\uff0c\u8fd9\u6837\u6765\u4fdd\u8bc1 Reward Model \u5b66\u4e60\u7684\u662f\u5fae\u8c03\u540e\u6700\u65b0\u7684\u751f\u6210\u7ed3\u679c\u3002<\/p>\n<p><strong>4. \u6570\u636e\u968f\u65f6\u95f4\u7684\u53d8\u5316<\/strong><\/p>\n<p>\u7531\u4e0a\u8868\u6240\u793a\uff0c\u6bcf\u5468\u6807\u6ce8\u7684\u6570\u636e\u8d8a\u6765\u8d8a\u591a\uff0c\u8fd9\u662f\u56e0\u4e3a\u6807\u6ce8\u8005\u5728\u4e0d\u65ad\u7684\u589e\u52a0\uff0c\u4ed6\u4eec\u4e5f\u8d8a\u6765\u8d8a\u719f\u7ec3\u3002\u540c\u65f6\u5bf9\u8bdd\u8f6e\u6570\u8d8a\u6765\u8d8a\u591a\uff0c\u589e\u52a0\u4e86 Reward Model \u5b66\u4e60\u7684\u96be\u5ea6\u3002<\/p>\n<p>Prompt \u8d8a\u6765\u8d8a\u56f0\u96be\u3002LLaMA-2 \u8bba\u6587\u91cc\u63d0\u5230\uff0c\u521a\u5f00\u59cb\u8981\u6c42\u6807\u6ce8\u8005\u95ee\u4e00\u4e9b\u7b80\u5355\u7684\u95ee\u9898\uff0c\u540e\u9762\u8981\u6c42\u6807\u6ce8\u8005\u95ee\u66f4\u96be\u7684\u95ee\u9898\u3002\u4e0b\u56fe\u53ef\u4ee5\u770b\u5230\uff0c\u8d8a\u5f80\u540e\u7684\u65f6\u95f4\uff0cResponse \u5f97\u5230\u7684\u5206\u6570\u8d8a\u4f4e\uff0c\u56e0\u4e3a\u95ee\u9898\u66f4\u96be\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/f958988883b58725a89394aba64fff37\/23296\/03.png\" target=\"_blank\" rel=\"noopener\"><img decoding=\"async\" class=\"alignnone size-full wp-image-400\" title=\"e45ee7ce7e88149af8dd52\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd52.png\" alt=\"e45ee7ce7e88149af8dd52\" width=\"650\" height=\"508\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd52.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd52-300x234.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd52-150x117.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u800c significantly better \u7684\u6807\u7b7e\u91cf\u8d8a\u6765\u8d8a\u5c11\uff0c\u800c unsure \u7684\u6807\u7b7e\u91cf\u8d8a\u6765\u8d8a\u591a\uff0c\u8fd9\u662f\u56e0\u4e3a\u6a21\u578b\u7684\u6027\u80fd\u5728\u4e0d\u65ad\u7684\u63d0\u5347\uff0c\u6240\u4ee5\u91c7\u6837\u5230\u7684\u4e24\u4e2a\u56de\u7b54\u90fd\u662f\u9ad8\u8d28\u91cf\u7684\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u65f6\u95f4\u8d8a\u5f80\u540e\uff0c\u663e\u8457\u597d\u7684\u6807\u7b7e\u91cf\u8d8a\u6765\u8d8a\u5c11\uff0c\u597d\u7684\u5fae\u4e4e\u5176\u5fae\u7684\u6807\u7b7e\u91cf\u8d8a\u6765\u8d8a\u591a\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/9e79221e41a57a6fac489fd9c29f38eb\/9a1cf\/04.png\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-399\" title=\"7d0665438e81d8eceb9845\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9845.png\" alt=\"7d0665438e81d8eceb9845\" width=\"650\" height=\"407\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9845.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9845-300x188.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9845-150x94.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p><strong>5. \u591a\u8f6e\u6570\u636e\u600e\u4e48\u6807\u6ce8\uff1f<\/strong><\/p>\n<p>\u53ef\u4ee5\u786e\u5b9a\u7684\u662f\uff0c\u591a\u8f6e\u5bf9\u8bdd\u7684\u6570\u636e\u4f53\u73b0\u5728 Prompt \u4e2d\uff0c\u9700\u8981\u6807\u6ce8\u7684\u662f\u6700\u540e\u4e00\u8f6e\u7684\u5bf9\u8bdd\u3002\u4f46\u662f\u6784\u5efa\u591a\u8f6e\u7684 Prompt \u8bba\u6587\u6ca1\u6709\u8be6\u7ec6\u63d0\u53ca\uff0c\u6211\u4eec\u601d\u8003\u53ef\u4ee5\u91c7\u7528\u5982\u4e0b\u65b9\u6cd5\uff1a<\/p>\n<ol>\n<li>\u7528\u6237\u5199\u4e00\u4e2a Prompt\uff1aP1\n<p>a. \u5355\u8f6e Prompt\uff1a\u7528\u6237\u76f4\u63a5\u5199<\/p>\n<p>b. \u591a\u8f6e Prompt\uff1a\u7528\u6237\u5148\u548c\u975e\u6807\u6ce8\u6a21\u578b\u5bf9\u8bdd\u591a\u8f6e\uff0c\u6700\u540e\u4e00\u8f6e\u518d\u8fdb\u884c\u6807\u6ce8<\/li>\n<li>\u751f\u6210\uff1a\u5185\u90e8\u8c03\u7528\u4e24\u4e2a\u6a21\u578b\uff0c\u751f\u6210 response_1\uff0cresponse_2<\/li>\n<li>\u6807\u6ce8\uff1a\u7528\u6237\u6807\u6ce8\u54ea\u4e2a\u56de\u7b54\u662f\u66f4\u4f18\u7684\uff0c\u4ee5\u53ca\u66f4\u4f18\u7684\u7a0b\u5ea6\uff0c\u88ab\u9009\u62e9\u7684\u56de\u7b54\u6807\u8bb0\u4e3a R1<\/li>\n<\/ol>\n<p>\u8fd9\u6837\u6807\u6ce8\u6570\u636e\u7684\u591a\u6837\u6027\u5f97\u4ee5\u4fdd\u8bc1\uff0c\u4f46\u662f\u9700\u8981\u4ea4\u4e92\u754c\u9762\u505a\u4e00\u5b9a\u7684\u8c03\u6574\uff0c\u8ba9\u7528\u6237\u9009\u62e9\u6807\u6ce8\u6216\u8005\u5bf9\u8bdd\u6a21\u5f0f\u3002<\/p>","protected":false},"excerpt":{"rendered":"<p>LLaMA \u662f\u76ee\u524d\u5907\u53d7\u5173\u6ce8\u7684\u5f00\u6e90\u5927\u8bed\u8a00\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u6700\u8fd1 Meta \u53d1\u5e03\u4e86 LLaMA 2\uff0c\u5b83\u662f LLaMA \u7684 [&hellip;]<\/p>","protected":false},"author":1,"featured_media":402,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-397","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-basic-research"],"acf":[],"_links":{"self":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/397","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/comments?post=397"}],"version-history":[{"count":1,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/397\/revisions"}],"predecessor-version":[{"id":403,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/397\/revisions\/403"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/media\/402"}],"wp:attachment":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/media?parent=397"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/categories?post=397"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/tags?post=397"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}