{"id":2672,"date":"2026-09-03T14:26:14","date_gmt":"2026-09-03T06:26:14","guid":{"rendered":"http:\/\/blog.xtaa.cn\/?p=2672"},"modified":"2026-09-03T14:40:42","modified_gmt":"2026-09-03T06:40:42","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%be%ae%e8%b0%83","status":"publish","type":"post","link":"http:\/\/blog.xtaa.cn\/index.php\/2026\/09\/03\/%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%be%ae%e8%b0%83\/","title":{"rendered":"\u5927\u6a21\u578b\u5fae\u8c03\uff0c\u5bf9\u9f50"},"content":{"rendered":"\n<p>\u5927\u6a21\u578b\u901a\u8fc7<strong>\u9884\u8bad\u7ec3<\/strong>\u4ece\u6d77\u91cf\u6587\u672c\u4e2d\u5b66\u4e60<strong>\u8bed\u8a00\u7684\u4e00\u822c\u89c4\u5f8b\u548c\u4e16\u754c\u77e5\u8bc6<\/strong>\uff0c\u518d\u901a\u8fc7<strong>\u5fae\u8c03\uff08Fine-tuning\uff09<\/strong>\u5728\u9884\u8bad\u7ec3\u57fa\u7840\u4e0a\u7528\u7279\u5b9a\u9886\u57df\u6570\u636e\u8fdb\u4e00\u6b65\u9002\u914d\u3002<br><br>\u5168\u91cf\u5fae\u8c03\u3001LoRA\u3001Prefix Tuning\u3001P-Tuning<br>\u8fd9\u51e0\u79cd\u90fd\u662f\u5927\u6a21\u578b\u5fae\u8c03\u7684\u4e3b\u6d41\u65b9\u6cd5\uff0c\u6838\u5fc3\u533a\u522b\u5728\u4e8e\u200c<strong>\u66f4\u65b0\u591a\u5c11\u53c2\u6570<\/strong>\u200c\u4ee5\u53ca\u200c<strong>\u5982\u4f55\u66f4\u65b0<\/strong>\u200c\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u200c<strong>\u5168\u91cf\u5fae\u8c03\u662f\u201c\u5927\u52a8\u5e72\u6208\u201d\uff0c\u540e\u9762\u4e09\u79cd\u90fd\u5c5e\u4e8e\u201c\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u201d\uff08PEFT\uff09\uff0c\u53ea\u8bad\u7ec3\u5c11\u91cf\u53c2\u6570\uff0c\u7701\u8d44\u6e90\u3001\u901f\u5ea6\u5feb<\/strong>\u200c\u3002\u200c\u200c<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u5168\u91cf\u5fae\u8c03\uff08Full Fine-tuning\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u539f\u7406<\/strong>\u200c\uff1a\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u200c<strong>\u6240\u6709\u53c2\u6570<\/strong>\u200c\u8fdb\u884c\u66f4\u65b0\uff0c\u8ba9\u6574\u4e2a\u6a21\u578b\u9002\u5e94\u4e0b\u6e38\u4efb\u52a1\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u4f18\u70b9<\/strong>\u200c\uff1a\u7406\u8bba\u4e0a\u6027\u80fd\u4e0a\u9650\u6700\u9ad8\uff0c\u80fd\u5145\u5206\u6316\u6398\u6a21\u578b\u6f5c\u529b\uff0c\u9002\u5408\u6570\u636e\u5145\u8db3\u3001\u8ffd\u6c42\u6781\u81f4\u6548\u679c\u7684\u573a\u666f\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u7f3a\u70b9<\/strong>\u200c\uff1a\u663e\u5b58\u548c\u7b97\u529b\u5f00\u9500\u5de8\u5927\uff08\u5982GPT-3 175B\u5168\u91cf\u5fae\u8c03\u9700\u7ea62.3TB\u663e\u5b58\uff09\uff0c\u8bad\u7ec3\u65f6\u95f4\u957f\uff0c\u5c0f\u6570\u636e\u4e0b\u5bb9\u6613\u8fc7\u62df\u5408\uff0c\u6bcf\u4e2a\u4efb\u52a1\u90fd\u8981\u4fdd\u5b58\u4e00\u4efd\u5b8c\u6574\u6a21\u578b\u526f\u672c\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u9002\u7528\u573a\u666f<\/strong>\u200c\uff1a\u6570\u636e\u91cf\u5927\uff0810\u4e07+\u6837\u672c\uff09\u3001\u7b97\u529b\u5145\u8db3\u3001\u4efb\u52a1\u590d\u6742\u4e14\u5bf9\u7cbe\u5ea6\u8981\u6c42\u6781\u9ad8\u7684\u573a\u666f\uff08\u5982\u533b\u7597\u8bca\u65ad\u3001\u91d1\u878d\u98ce\u63a7\uff09\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">LoRA\uff08Low-Rank Adaptation\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u539f\u7406<\/strong>\u200c\uff1a\u51bb\u7ed3\u539f\u6a21\u578b\u6743\u91cd\uff0c\u5728\u6743\u91cd\u77e9\u9635\u65c1\u6dfb\u52a0\u4e24\u4e2a\u4f4e\u79e9\u5c0f\u77e9\u9635\uff08\u5982A\u548cB\uff09\uff0c\u200c<strong>\u53ea\u8bad\u7ec3\u8fd9\u4e24\u4e2a\u5c0f\u77e9\u9635<\/strong>\u200c\uff0c\u8bad\u7ec3\u5b8c\u628a\u7ed3\u679c\u5408\u5e76\u56de\u539f\u6743\u91cd\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u4f18\u70b9<\/strong>\u200c\uff1a\u53ef\u8bad\u7ec3\u53c2\u6570\u4ece\u51e0\u5341\u4e07\u7ea7\u964d\u5230\u51e0\u5343\u51e0\u4e07\u7ea7\uff0c\u663e\u5b58\u5360\u7528\u6781\u4f4e\uff0c\u5355\u5361\u53ef\u8bad\u7ec370B\u751a\u81f3175B\u6a21\u578b\uff0c\u4e14\u5373\u63d2\u5373\u7528\u3001\u65b9\u4fbf\u5207\u6362\u4efb\u52a1\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u7f3a\u70b9<\/strong>\u200c\uff1a\u4f4e\u79e9\u8fd1\u4f3c\u4f1a\u5f15\u5165\u4e00\u5b9a\u8bef\u5dee\uff0c\u5728\u9700\u8981\u7ec6\u7c92\u5ea6\u7406\u89e3\u7684\u590d\u6742\u4efb\u52a1\u4e0a\u6027\u80fd\u53ef\u80fd\u4e0d\u53ca\u5168\u91cf\u5fae\u8c03\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u9002\u7528\u573a\u666f<\/strong>\u200c\uff1a\u8d44\u6e90\u6709\u9650\u3001\u9700\u8981\u5feb\u901f\u8fed\u4ee3\u6216\u591a\u4efb\u52a1\u9002\u914d\u7684\u573a\u666f\uff0c\u662f\u76ee\u524d\u5de5\u4e1a\u754c\u6700\u5e38\u7528\u7684\u65b9\u6cd5\u4e4b\u4e00\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">Prefix Tuning\uff08\u524d\u7f00\u5fae\u8c03\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u539f\u7406<\/strong>\u200c\uff1a\u5728Transformer\u200c<strong>\u6bcf\u4e00\u5c42<\/strong>\u200c\u7684\u8f93\u5165\u524d\u6dfb\u52a0\u4e00\u7ec4\u53ef\u5b66\u4e60\u7684\u201c\u865a\u62df\u524d\u7f00\u201d\u5411\u91cf\uff08Prefix\uff09\uff0c\u53c2\u4e0e\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u5f15\u5bfc\u6a21\u578b\u9002\u914d\u4efb\u52a1\u3002\u53ea\u8bad\u7ec3\u524d\u7f00\u53c2\u6570\uff0c\u6a21\u578b\u4e3b\u4f53\u51bb\u7ed3\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u4f18\u70b9<\/strong>\u200c\uff1a\u65e0\u9700\u4fee\u6539\u6a21\u578b\u5185\u90e8\u7ed3\u6784\uff0c\u4ec5\u9700\u8bad\u7ec3\u5c11\u91cf\u524d\u7f00\u5411\u91cf\uff0c\u9002\u5408\u591a\u4efb\u52a1\u5b66\u4e60\uff08\u4e0d\u540c\u4efb\u52a1\u53ef\u4fdd\u5b58\u4e0d\u540c\u524d\u7f00\uff09\uff0c\u5728\u751f\u6210\u4efb\u52a1\u4e0a\u6027\u80fd\u63a5\u8fd1\u5168\u91cf\u5fae\u8c03\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u7f3a\u70b9<\/strong>\u200c\uff1a\u4f1a\u200c<strong>\u589e\u52a0\u8f93\u5165\u5e8f\u5217\u957f\u5ea6<\/strong>\u200c\uff0c\u5e26\u6765\u989d\u5916\u8ba1\u7b97\u548c\u63a8\u7406\u5f00\u9500\uff1b\u524d\u7f00\u957f\u5ea6\u9009\u62e9\u5bf9\u6027\u80fd\u5f71\u54cd\u8f83\u5927\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u9002\u7528\u573a\u666f<\/strong>\u200c\uff1a\u200c<strong>\u6587\u672c\u751f\u6210\u7c7b\u4efb\u52a1<\/strong>\u200c\uff08\u5982\u5bf9\u8bdd\u3001\u6458\u8981\u3001\u673a\u5668\u7ffb\u8bd1\uff09\uff0c\u9002\u914dGPT\u3001BART\u3001T5\u7b49\u81ea\u56de\u5f52\u6216\u7f16\u7801\u5668-\u89e3\u7801\u5668\u6a21\u578b\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">P-Tuning\uff08\u63d0\u793a\u5fae\u8c03\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u539f\u7406<\/strong>\u200c\uff1a\u5728\u200c<strong>\u8f93\u5165\u5d4c\u5165\u5c42<\/strong>\u200c\uff08P-Tuning v1\uff09\u6216\u200c<strong>\u6bcf\u4e00\u5c42<\/strong>\u200c\uff08P-Tuning v2\uff09\u63d2\u5165\u53ef\u8bad\u7ec3\u7684\u201c\u8f6f\u63d0\u793a\u201d\uff08Soft Prompt\uff09\u5411\u91cf\uff0c\u901a\u8fc7LSTM\u6216MLP\u7f16\u7801\u5668\u751f\u6210\u63d0\u793a\u5d4c\u5165\uff0c\u5f15\u5bfc\u6a21\u578b\u7406\u89e3\u4efb\u52a1\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u4f18\u70b9<\/strong>\u200c\uff1a\u76f8\u6bd4Prefix Tuning\u66f4\u7075\u6d3b\uff0c\u63d0\u793a\u53ef\u63d2\u5165\u8f93\u5165\u4efb\u610f\u4f4d\u7f6e\uff1bP-Tuning v2 \u5728\u6bcf\u5c42\u52a0\u5165\u63d0\u793a\uff0c\u6027\u80fd\u53ef\u63a5\u8fd1\u5168\u91cf\u5fae\u8c03\uff0c\u4e14\u200c<strong>\u89e3\u51b3\u4e86Prompt Tuning\u5728\u5c0f\u6a21\u578b\u4e0a\u6548\u679c\u5dee\u7684\u95ee\u9898<\/strong>\u200c\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u7f3a\u70b9<\/strong>\u200c\uff1aP-Tuning v1\u4ec5\u4f18\u5316\u8f93\u5165\u5c42\uff0c\u6548\u679c\u6709\u9650\uff1bv2\u5b9e\u73b0\u590d\u6742\u5ea6\u66f4\u9ad8\uff0c\u5bf9\u63d0\u793a\u957f\u5ea6\u7b49\u8d85\u53c2\u6570\u654f\u611f\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u9002\u7528\u573a\u666f<\/strong>\u200c\uff1a\u200c<strong>\u81ea\u7136\u8bed\u8a00\u7406\u89e3\uff08NLU\uff09\u4efb\u52a1<\/strong>\u200c\uff08\u5982\u6587\u672c\u5206\u7c7b\u3001\u60c5\u611f\u5206\u6790\u3001\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\uff09\uff0cP-Tuning v2 \u5728\u591a\u7c7bNLU\u57fa\u51c6\u4e0a\u8868\u73b0\u4f18\u5f02\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u603b\u7ed3\u4e0e\u9009\u62e9\u5efa\u8bae<\/h3>\n\n\n\n<p>\u8868\u683c<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u65b9\u6cd5<\/th><th>\u66f4\u65b0\u53c2\u6570\u4f4d\u7f6e<\/th><th>\u53c2\u6570\u91cf<\/th><th>\u5178\u578b\u4efb\u52a1<\/th><th>\u8d44\u6e90\u9700\u6c42<\/th><\/tr><\/thead><tbody><tr><td>\u200c<strong>\u5168\u91cf\u5fae\u8c03<\/strong>\u200c<\/td><td>\u6240\u6709\u5c42\u6240\u6709\u53c2\u6570<\/td><td>100%<\/td><td>\u590d\u6742\u4efb\u52a1\u3001\u8ffd\u6c42\u6781\u81f4\u6027\u80fd<\/td><td>\u6781\u9ad8<\/td><\/tr><tr><td>\u200c<strong>LoRA<\/strong>\u200c<\/td><td>\u6743\u91cd\u77e9\u9635\u65c1\u7684\u4f4e\u79e9\u77e9\u9635<\/td><td>\u7ea60.01%-1%<\/td><td>\u901a\u7528\uff0c\u5de5\u4e1a\u754c\u9996\u9009<\/td><td>\u4f4e<\/td><\/tr><tr><td>\u200c<strong>Prefix Tuning<\/strong>\u200c<\/td><td>\u6bcf\u5c42\u8f93\u5165\u524d\u7f00<\/td><td>\u7ea60.1%-3%<\/td><td>\u6587\u672c\u751f\u6210\uff08NLG\uff09<\/td><td>\u4f4e<\/td><\/tr><tr><td>\u200c<strong>P-Tuning<\/strong>\u200c<\/td><td>\u8f93\u5165\u5c42\u6216\u6bcf\u5c42\u8f6f\u63d0\u793a<\/td><td>\u7ea60.1%-3%<\/td><td>\u6587\u672c\u7406\u89e3\uff08NLU\uff09<\/td><td>\u4f4e<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u200c<strong>\u9009\u578b\u53c2\u8003<\/strong>\u200c\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7b97\u529b\u548c\u6570\u636e\u5145\u8db3\u3001\u8ffd\u6c42\u6700\u9ad8\u6027\u80fd \u2192 \u200c<strong>\u5168\u91cf\u5fae\u8c03<\/strong>\u200c\u3002<\/li>\n\n\n\n<li>\u8d44\u6e90\u6709\u9650\u3001\u9700\u8981\u5feb\u901f\u8fed\u4ee3\u6216\u591a\u4efb\u52a1\u9002\u914d \u2192 \u200c<strong>LoRA<\/strong>\u200c\u3002<\/li>\n\n\n\n<li>\u751f\u6210\u7c7b\u4efb\u52a1\uff08\u5bf9\u8bdd\u3001\u6458\u8981\uff09 \u2192 \u200c<strong>Prefix Tuning<\/strong>\u200c \u6216 \u200c<strong>P-Tuning v2<\/strong>\u200c\u3002<\/li>\n\n\n\n<li>\u7406\u89e3\u7c7b\u4efb\u52a1\uff08\u5206\u7c7b\u3001\u5b9e\u4f53\u8bc6\u522b\uff09 \u2192 \u200c<strong>P-Tuning v2<\/strong>\u200c\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<p><\/p>\n\n\n\n<p><\/p>\n\n\n\n<p>\u6a21\u578b\u5fae\u8c03\u4e4b\u540e\u8fd8\u6709\u4e00\u6b65&#8221;<strong>\u5bf9\u9f50<\/strong>&#8220;\uff0c\u8ba9\u6a21\u578b\u8f93\u51fa\u7b26\u5408\u4eba\u7c7b\u671f\u671b\u3002<strong>RLHF<\/strong>\uff08Reinforcement Learning from Human Feedback\uff0c\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff09\u901a\u8fc7<strong>\u4eba\u7c7b\u53cd\u9988\u6765\u4f18\u5316\u6a21\u578b\u884c\u4e3a<\/strong>\u2014\u2014\u5224\u65ad&#8221;RLHF \u4e0d\u9700\u8981\u4eba\u7c7b\u53c2\u4e0e&#8221;\u4e3a<strong>\u9519\u8bef<\/strong>\uff0c\u5b83\u6070\u6070\u4ee5\u4eba\u7c7b\u53cd\u9988\u4e3a\u6838\u5fc3\u3002<strong>DPO<\/strong>\uff08Direct Preference Optimization\uff0c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff09\u5219\u662f\u66f4\u8f7b\u91cf\u7684\u5bf9\u9f50\u65b9\u5f0f\u3002<br>\u5fae\u8c03\u548c\u5bf9\u9f50\u662f\u200c<strong>\u5927\u6a21\u578b\u8bad\u7ec3\u4e2d\u4e24\u4e2a\u4e0d\u540c\u9636\u6bb5<\/strong>\u200c\uff1a\u5fae\u8c03\u8ba9\u6a21\u578b\u201c\u5b66\u4f1a\u505a\u4e8b\u201d\uff0c\u5bf9\u9f50\u8ba9\u6a21\u578b\u201c\u505a\u5f97\u7b26\u5408\u4eba\u7c7b\u671f\u671b\u548c\u4ef7\u503c\u89c2\u201d\u3002\u53ef\u4ee5\u7b80\u5355\u7406\u89e3\u4e3a\uff1a\u200c<strong>\u5fae\u8c03\u662f\u6559\u6280\u80fd\uff0c\u5bf9\u9f50\u662f\u6559\u505a\u4eba<\/strong>\u200c\u3002\u200c\u200c<\/p>\n\n\n\n<p>\u4e24\u8005\u5728\u76ee\u6807\u3001\u6570\u636e\u3001\u65b9\u6cd5\u4e0a\u90fd\u6709\u660e\u663e\u5dee\u5f02\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83c\udfaf \u76ee\u6807\u4e0d\u540c<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u5fae\u8c03<\/strong>\u200c\uff1a\u8ba9\u6a21\u578b\u5b66\u4f1a\u7279\u5b9a\u4efb\u52a1\u3001\u9886\u57df\u6216\u4ea4\u4e92\u65b9\u5f0f\uff0c\u6bd4\u5982\u5199\u6cd5\u5f8b\u5408\u540c\u3001\u505a\u5ba2\u670d\u95ee\u7b54\u3001\u6309\u6307\u5b9a\u683c\u5f0f\u8f93\u51fa\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u5bf9\u9f50<\/strong>\u200c\uff1a\u8ba9\u6a21\u578b\u8f93\u51fa\u7b26\u5408\u4eba\u7c7b\u504f\u597d\u548c\u4ef7\u503c\u89c2\uff0c\u51cf\u5c11\u6709\u5bb3\u3001\u504f\u89c1\u3001\u5e7b\u89c9\u5185\u5bb9\uff0c\u505a\u5230\u6709\u7528\uff08Helpful\uff09\u3001\u8bda\u5b9e\uff08Honest\uff09\u3001\u65e0\u5bb3\uff08Harmless\uff09\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udcca \u6570\u636e\u4e0d\u540c<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u5fae\u8c03<\/strong>\u200c\uff1a\u4f7f\u7528\u9ad8\u8d28\u91cf\u7684\u201c\u95ee\u9898-\u7b54\u6848\u201d\u5bf9\uff08Instruction-Response\uff09\uff0c\u6570\u636e\u91cf\u4ece\u6570\u5343\u5230\u6570\u767e\u4e07\u6761\u4e0d\u7b49\uff0c\u9488\u5bf9\u6027\u8d8a\u5f3a\u8d8a\u597d\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u5bf9\u9f50<\/strong>\u200c\uff1a\u4f7f\u7528\u4eba\u7c7b\u504f\u597d\u6570\u636e\uff0c\u6bd4\u5982\u201c\u56de\u7b54A\u6bd4\u56de\u7b54B\u597d\u201d\u7684\u6210\u5bf9\u6bd4\u8f83\u6570\u636e\uff0c\u901a\u5e38\u9700\u8981\u6570\u4e07\u5230\u6570\u5341\u4e07\u6761\u504f\u597d\u5bf9\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udd27 \u65b9\u6cd5\u4e0d\u540c<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u5fae\u8c03<\/strong>\u200c\uff1a\u5305\u62ec\u5168\u53c2\u6570\u5fae\u8c03\uff08\u66f4\u65b0\u6240\u6709\u53c2\u6570\uff09\u548c\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff08PEFT\uff09\uff0c\u5e38\u7528\u65b9\u6cd5\u6709 LoRA\u3001Adapter\u3001Prefix-Tuning \u7b49\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u5bf9\u9f50<\/strong>\u200c\uff1a\u7ecf\u5178\u65b9\u6cd5\u6709 RLHF\uff08\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff09\u3001DPO\uff08\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff09\u3001KTO\u3001Constitutional AI \u7b49\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u26a0\ufe0f \u9700\u8981\u6f84\u6e05\u7684\u51e0\u70b9<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u200c<strong>\u4e24\u8005\u6709\u91cd\u53e0<\/strong>\u200c\uff1a\u5f88\u591a\u8d44\u6599\u4f1a\u628a\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u4e5f\u5f52\u5165\u5bf9\u9f50\u7684\u5e7f\u4e49\u8303\u7574\uff0c\u5de5\u7a0b\u4e0a\u5e38\u6309\u201c\u9884\u8bad\u7ec3\u2192\u5fae\u8c03\u2192\u5bf9\u9f50\u201d\u4e09\u9636\u6bb5\u5212\u5206\uff0c\u65b9\u4fbf\u7406\u89e3\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u662f\u5426\u5fc5\u987b\u5bf9\u9f50\u770b\u573a\u666f<\/strong>\u200c\uff1a\u5982\u679c\u6a21\u578b\u5728\u7279\u5b9a\u5c01\u95ed\u4efb\u52a1\u4e0a\u6548\u679c\u8fbe\u6807\u3001\u6ca1\u6709\u660e\u663e\u4e0d\u826f\u884c\u4e3a\uff0c\u53ef\u4ee5\u4e0d\u8fdb\u884c\u5bf9\u9f50\uff1b\u4f46\u901a\u7528\u5bf9\u8bdd\u6a21\u578b\u51e0\u4e4e\u4e00\u5b9a\u4f1a\u505a\u5bf9\u9f50\uff0c\u56e0\u4e3a\u5f00\u653e\u573a\u666f\u4e0d\u53ef\u63a7\u56e0\u7d20\u592a\u591a\u3002<\/li>\n\n\n\n<li>\u200c<strong>\u7b97\u529b\u5f00\u9500\u5dee\u5f02\u5927<\/strong>\u200c\uff1a\u5fae\u8c03\u901a\u5e38\u5360\u5927\u6a21\u578b\u603b\u8bad\u7ec3\u7b97\u529b\u7684 1%-5%\uff0c\u5bf9\u9f50\u4e0d\u5230 1%\uff0c\u4f46\u5de5\u7a0b\u7a33\u5b9a\u6027\u8981\u6c42\u6781\u9ad8\u3002\u200c\u200c<\/li>\n<\/ul>\n\n\n\n<p>\u7b80\u5355\u8bf4\uff0c\u5fae\u8c03\u51b3\u5b9a\u6a21\u578b\u201c\u4f1a\u4e0d\u4f1a\u505a\u201d\uff0c\u5bf9\u9f50\u51b3\u5b9a\u6a21\u578b\u201c\u80fd\u4e0d\u80fd\u597d\u597d\u505a\u201d\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5927\u6a21\u578b\u901a\u8fc7\u9884\u8bad\u7ec3\u4ece\u6d77\u91cf\u6587\u672c\u4e2d\u5b66\u4e60\u8bed\u8a00\u7684\u4e00\u822c\u89c4\u5f8b\u548c\u4e16\u754c\u77e5\u8bc6\uff0c\u518d\u901a\u8fc7\u5fae\u8c03\uff08Fine-tuning\uff09\u5728\u9884\u8bad\u7ec3\u57fa\u7840\u4e0a\u7528 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[30],"tags":[],"class_list":["post-2672","post","type-post","status-publish","format-standard","hentry","category-30"],"_links":{"self":[{"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/posts\/2672","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/comments?post=2672"}],"version-history":[{"count":4,"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/posts\/2672\/revisions"}],"predecessor-version":[{"id":2682,"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/posts\/2672\/revisions\/2682"}],"wp:attachment":[{"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/media?parent=2672"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/categories?post=2672"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/blog.xtaa.cn\/index.php\/wp-json\/wp\/v2\/tags?post=2672"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}