{"id":6416,"date":"2024-08-02T10:26:06","date_gmt":"2024-08-02T02:26:06","guid":{"rendered":"https:\/\/t.n-years.com\/?p=6416"},"modified":"2024-08-02T10:26:06","modified_gmt":"2024-08-02T02:26:06","slug":"%e6%a8%a1%e5%9e%8b%e8%ae%ad%e7%bb%83-dpo-%e5%92%8c-sft-%e6%9c%89%e4%bb%80%e4%b9%88%e4%b8%8d%e5%90%8c","status":"publish","type":"post","link":"https:\/\/t.n-years.com\/?p=6416","title":{"rendered":"\u6a21\u578b\u8bad\u7ec3 DPO \u548c SFT \u6709\u4ec0\u4e48\u4e0d\u540c"},"content":{"rendered":"<blockquote>\n<p>\u975e\u539f\u521b\u3001ChatGPT \u95ee\u7b54\u751f\u6210<\/p>\n<\/blockquote>\n<p>DPO\uff08Direct Preference Optimization\uff09\u548c SFT\uff08Supervised Fine-Tuning\uff09\u662f\u673a\u5668\u5b66\u4e60\u548c\u6df1\u5ea6\u5b66\u4e60\u4e2d\u4e24\u79cd\u4e0d\u540c\u7684\u4f18\u5316\u65b9\u6cd5\uff0c\u5b83\u4eec\u4e3b\u8981\u7528\u4e8e\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u4f18\u5316\uff0c\u4f46\u6709\u4e0d\u540c\u7684\u76ee\u6807\u548c\u5e94\u7528\u65b9\u5f0f\u3002\u4e0b\u9762\u662f\u5b83\u4eec\u7684\u5b9a\u4e49\u548c\u533a\u522b\uff1a<\/p>\n<h3>DPO\uff08Direct Preference Optimization\uff09<\/h3>\n<ul>\n<li><strong>\u5b9a\u4e49<\/strong>\uff1aDPO \u662f\u4e00\u79cd\u76f4\u63a5\u4f18\u5316\u6a21\u578b\u8f93\u51fa\u7684\u504f\u597d\uff08\u6216\u6392\u5e8f\uff09\u7684\u65b9\u6cd5\u3002\u901a\u5e38\u5728\u5904\u7406\u6392\u5e8f\u95ee\u9898\u6216\u9700\u8981\u5bf9\u591a\u4e2a\u9009\u9879\u8fdb\u884c\u6bd4\u8f83\u65f6\u4f7f\u7528\u3002<\/li>\n<li><strong>\u76ee\u6807<\/strong>\uff1aDPO \u7684\u76ee\u6807\u662f\u76f4\u63a5\u4f18\u5316\u6a21\u578b\u7684\u8f93\u51fa\u504f\u597d\uff0c\u4f7f\u5176\u66f4\u597d\u5730\u7b26\u5408\u4eba\u7c7b\u7684\u504f\u597d\u6216\u6392\u5e8f\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b83\u4f1a\u4f18\u5316\u6a21\u578b\u751f\u6210\u7684\u7ed3\u679c\uff0c\u4f7f\u5f97\u7528\u6237\u7684\u504f\u597d\u6392\u5e8f\u5f97\u5230\u66f4\u597d\u7684\u4f53\u73b0\u3002<\/li>\n<li><strong>\u5e94\u7528\u573a\u666f<\/strong>\uff1aDPO \u5e38\u7528\u4e8e\u63a8\u8350\u7cfb\u7edf\u3001\u641c\u7d22\u5f15\u64ce\u6392\u5e8f\u7b49\u573a\u666f\uff0c\u5176\u4e2d\u9700\u8981\u6839\u636e\u7528\u6237\u7684\u53cd\u9988\u5bf9\u6a21\u578b\u8fdb\u884c\u4f18\u5316\u3002\u901a\u8fc7\u4f18\u5316\u6a21\u578b\u7684\u8f93\u51fa\uff0c\u4f7f\u5176\u66f4\u7b26\u5408\u7528\u6237\u7684\u5b9e\u9645\u9700\u6c42\u548c\u504f\u597d\u3002<\/li>\n<\/ul>\n<h3>SFT\uff08Supervised Fine-Tuning\uff09<\/h3>\n<ul>\n<li><strong>\u5b9a\u4e49<\/strong>\uff1aSFT \u662f\u4e00\u79cd\u76d1\u7763\u5b66\u4e60\u7684\u65b9\u6cd5\uff0c\u7528\u4e8e\u5728\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u57fa\u7840\u4e0a\u8fdb\u884c\u5fae\u8c03\u3002\u5b83\u901a\u8fc7\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u4f7f\u7528\u6807\u6ce8\u7684\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u4ee5\u63d0\u5347\u6a21\u578b\u5728\u8be5\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002<\/li>\n<li><strong>\u76ee\u6807<\/strong>\uff1aSFT \u7684\u76ee\u6807\u662f\u901a\u8fc7\u76d1\u7763\u4fe1\u53f7\uff08\u6807\u6ce8\u7684\u6570\u636e\uff09\u8c03\u6574\u6a21\u578b\u7684\u53c2\u6570\uff0c\u4f7f\u5176\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u66f4\u597d\u3002\u5b83\u901a\u5e38\u7528\u4e8e\u5c06\u9884\u8bad\u7ec3\u6a21\u578b\u9002\u5e94\u4e8e\u7279\u5b9a\u5e94\u7528\u9886\u57df\u6216\u4efb\u52a1\u3002<\/li>\n<li><strong>\u5e94\u7528\u573a\u666f<\/strong>\uff1aSFT \u5e7f\u6cdb\u5e94\u7528\u4e8e\u5404\u79cd\u4efb\u52a1\uff0c\u5982\u6587\u672c\u5206\u7c7b\u3001\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\u3001\u673a\u5668\u7ffb\u8bd1\u7b49\u3002\u901a\u8fc7\u5728\u8fd9\u4e9b\u4efb\u52a1\u4e0a\u8fdb\u884c\u5fae\u8c03\uff0c\u6a21\u578b\u53ef\u4ee5\u66f4\u597d\u5730\u7406\u89e3\u548c\u5904\u7406\u7279\u5b9a\u9886\u57df\u7684\u6570\u636e\u3002<\/li>\n<\/ul>\n<h3>\u533a\u522b<\/h3>\n<ol>\n<li>\n<p><strong>\u4f18\u5316\u76ee\u6807<\/strong>\uff1a<\/p>\n<ul>\n<li><strong>DPO<\/strong>\uff1a\u76f4\u63a5\u4f18\u5316\u6a21\u578b\u8f93\u51fa\u7684\u504f\u597d\u6216\u6392\u5e8f\uff0c\u5173\u6ce8\u7684\u662f\u5982\u4f55\u4f7f\u6a21\u578b\u7684\u8f93\u51fa\u66f4\u7b26\u5408\u7528\u6237\u7684\u671f\u671b\u3002<\/li>\n<li><strong>SFT<\/strong>\uff1a\u901a\u8fc7\u76d1\u7763\u5b66\u4e60\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u5fae\u8c03\u6a21\u578b\uff0c\u5173\u6ce8\u7684\u662f\u5982\u4f55\u63d0\u9ad8\u6a21\u578b\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\n<p><strong>\u5e94\u7528\u573a\u666f<\/strong>\uff1a<\/p>\n<ul>\n<li><strong>DPO<\/strong>\uff1a\u5e38\u7528\u4e8e\u9700\u8981\u6392\u5e8f\u6216\u63a8\u8350\u7684\u573a\u666f\uff0c\u5982\u641c\u7d22\u5f15\u64ce\u7ed3\u679c\u6392\u5e8f\u3001\u4e2a\u6027\u5316\u63a8\u8350\u7b49\u3002<\/li>\n<li><strong>SFT<\/strong>\uff1a\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5404\u79cd\u76d1\u7763\u5b66\u4e60\u4efb\u52a1\uff0c\u5982\u5206\u7c7b\u3001\u56de\u5f52\u3001\u5e8f\u5217\u6807\u6ce8\u7b49\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\n<p><strong>\u8bad\u7ec3\u6570\u636e<\/strong>\uff1a<\/p>\n<ul>\n<li><strong>DPO<\/strong>\uff1a\u901a\u5e38\u4f9d\u8d56\u4e8e\u7528\u6237\u7684\u53cd\u9988\u6216\u504f\u597d\u6570\u636e\u3002<\/li>\n<li><strong>SFT<\/strong>\uff1a\u4f9d\u8d56\u4e8e\u6807\u6ce8\u7684\u6570\u636e\u96c6\uff0c\u8fd9\u4e9b\u6570\u636e\u96c6\u4e3a\u6a21\u578b\u63d0\u4f9b\u4e86\u5177\u4f53\u7684\u76ee\u6807\u8f93\u51fa\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p>\u603b\u7684\u6765\u8bf4\uff0cDPO \u548c SFT \u662f\u5728\u4e0d\u540c\u7684\u573a\u666f\u548c\u4efb\u52a1\u4e2d\u4f7f\u7528\u7684\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5404\u6709\u5176\u72ec\u7279\u7684\u76ee\u6807\u548c\u5e94\u7528\u65b9\u5f0f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u975e\u539f\u521b\u3001ChatGPT \u95ee\u7b54\u751f\u6210 DPO\uff08Direct Preference Optimization\uff09\u548c S&#8230; <\/p>\n<div class=\"read-more navbutton\"><a href=\"https:\/\/t.n-years.com\/?p=6416\">\u9605\u8bfb\u66f4\u591a<i class=\"fa fa-angle-double-right\"><\/i><\/a><\/div>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[137],"tags":[138,139],"class_list":["post-6416","post","type-post","status-publish","format-standard","hentry","category-llm","tag-dpo","tag-sft"],"_links":{"self":[{"href":"https:\/\/t.n-years.com\/index.php?rest_route=\/wp\/v2\/posts\/6416","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/t.n-years.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/t.n-years.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/t.n-years.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/t.n-years.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6416"}],"version-history":[{"count":1,"href":"https:\/\/t.n-years.com\/index.php?rest_route=\/wp\/v2\/posts\/6416\/revisions"}],"predecessor-version":[{"id":6417,"href":"https:\/\/t.n-years.com\/index.php?rest_route=\/wp\/v2\/posts\/6416\/revisions\/6417"}],"wp:attachment":[{"href":"https:\/\/t.n-years.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6416"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/t.n-years.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6416"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/t.n-years.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6416"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}