{"id":675,"date":"2024-03-15T10:58:13","date_gmt":"2024-03-15T02:58:13","guid":{"rendered":"http:\/\/lingbo.online\/?p=675"},"modified":"2024-03-15T10:58:13","modified_gmt":"2024-03-15T02:58:13","slug":"%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0rolellm-benchmarking-eliciting-and-enhancing-role-playing-abilities-of-large-language-models","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/uncategorized\/%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0rolellm-benchmarking-eliciting-and-enhancing-role-playing-abilities-of-large-language-models\/","title":{"rendered":"[\u8bba\u6587\u7b14\u8bb0]ROLELLM: BENCHMARKING, ELICITING, AND ENHANCING ROLE-PLAYING ABILITIES OF LARGE LANGUAGE MODELS"},"content":{"rendered":"<h1>MetaData<\/h1>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4\uff1a2023.10.01<\/li>\n<li>\u4f5c\u8005\uff1aZekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu,Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni,Man Zhang, Zhaoxiang Zhang, Wanli Ouyang, Ke Xu, Wenhu Chen, Jie Fu, Junran Peng<\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/InteractiveNLP-Team\/RoleLLM-public\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/InteractiveNLP-Team\/RoleLLM-public<\/a><\/li>\n<\/ul>\n<h1>\u6458\u8981<\/h1>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86RoleLLM\uff0c\u4e00\u4e2a\u5305\u62ecbenchmark\u3001elicit\u548c\u589e\u5f3a\u5927\u6a21\u578b\u89d2\u8272\u626e\u6f14\u80fd\u529b\u7684\u6846\u67b6\u3002RoleLLM\u5305\u62ec\u56db\u4e2a\u9636\u6bb5\uff1a\uff081\uff09100\u4e2a\u89d2\u8272\u7684\u89d2\u8272\u914d\u7f6e\u6784\u5efa\uff1b<br \/>\n\uff082\uff09\u9762\u5411\u89d2\u8272\u77e5\u8bc6\u62bd\u53d6\u7684\u57fa\u4e8e\u4e0a\u4e0b\u6587\u7684\u6307\u4ee4\u751f\u6210\uff08Context-Instruct\uff09\uff1b(3)RoleGPT\uff1a\u901a\u8fc7\u57fa\u4e8e\u63d0\u793a\u5de5\u7a0b\u7684\u89d2\u8272\u63d0\u793a\u8bed\u4ee5\u53ca\u7cfb\u7edf\u6307\u4ee4\u548c\u68c0\u7d22\u589e\u5f3a\uff0c\u5229\u7528GPT(Role Prompting using GPT,RoleGPT)\u7684\u89d2\u8272\u63d0\u793a\u8fdb\u884c\u8bf4\u8bdd\u98ce\u683c\u6a21\u4eff\uff1b\uff084\uff09\u89d2\u8272\u6761\u4ef6\u6307\u4ee4\u8c03\u4f18(Role-Conditioned Instruction Tuning, RoCIT)\uff0c\u7528\u4e8e\u5fae\u8c03\u5f00\u6e90\u6a21\u578b\u4ee5\u53ca\u89d2\u8272\u5b9a\u5236\u3002\u901a\u8fc7\u57fa\u4e8e\u4e0a\u4e0b\u6587\u7684\u6307\u4ee4\u751f\u6210\u548c\u57fa\u4e8e\u89d2\u8272\u63d0\u793a\u7684GPT\uff0c\u6211\u4eec\u521b\u5efa\u4e86RoleBench\uff0c\u8fd9\u662f\u7b2c\u4e00\u4e2a\u7cfb\u7edf\u7684\u3001\u7ec6\u7c92\u5ea6\u7684\u89d2\u8272\u626e\u6f14\u57fa\u51c6\u6570\u636e\u96c6\uff0c\u5305\u542b168,093\u4e2a\u6837\u672c\u3002\u6b64\u5916\uff0cRoCIT\u5728RoleBench\u4e0a\u751f\u6210\u4e86RoleLLaMA (English)\u548cRoleGLM (Chinese)\uff0c\u663e\u8457\u63d0\u9ad8\u4e86\u89d2\u8272\u626e\u6f14\u80fd\u529b\uff0c\u751a\u81f3\u8fbe\u5230\u4e0eRoleGPT\uff08\u4f7f\u7528GPT-4\uff09\u76f8\u5f53\u7684\u7ed3\u679c\u3002<\/p>\n<h1>\u5f15\u8a00<\/h1>\n<p>\u89d2\u8272\u626e\u6f14\u65e8\u5728\u8ba9LLM\u80fd\u591f\u6a21\u62df\u5177\u6709\u4e0d\u540c\u5c5e\u6027\u548c\u4f1a\u8bdd\u98ce\u683c\u7684\u5404\u79cd\u89d2\u8272\u6216\u4eba\u7269\uff0c\u4e3a\u7528\u6237\u63d0\u4f9b\u66f4\u7ec6\u81f4\u5165\u5fae\u7684\u4ea4\u4e92\u4f53\u9a8c\uff0c\u5e76\u4f7fllm\u66f4\u52a0\u4eb2\u8fd1\uff08familiar\uff09\uff0c\u53cb\u597d\uff08companionable\uff09\u548c\u6c89\u6d78\uff08immersive\uff09\u3002<br \/>\n\u7136\u800c\uff0c\u73b0\u6709\u7684\u5f00\u6e90\u6a21\u578b\u4e3b\u8981\u662f\u5728\u901a\u7528\u9886\u57df\u8fdb\u884c\u8bad\u7ec3\uff0c\u7f3a\u4e4f\u5bf9\u89d2\u8272\u626e\u6f14\u7684\u7279\u5b9a\u4f18\u5316\u3002\u6b64\u5916\uff0c\u5c3d\u7ba1GPT-4\u7b49\u6700\u73b0\u91d1\u7684llm\u5c55\u73b0\u51fa\u4f18\u79c0\u7684\u89d2\u8272\u626e\u6f14\u529f\u80fd\uff0c\u4f46\u4ed6\u4eec\u95ed\u6e90\u7684\u7279\u6027\u9020\u6210\u4e86\u4e00\u4e9b\u9650\u5236\uff0c\u5305\u62ec\u9ad8\u6602\u7684API\u6210\u672c\u3001\u65e0\u6cd5\u8fdb\u884c\u5fae\u8c03\u4ee5\u53ca\u6709\u9650\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u5927\u5c0f\u3002<br \/>\n\u4e3a\u4e86\u7f13\u89e3\u8fd9\u4e9b\u95ee\u9898\uff0c\u4e4b\u524d\u5df2\u7ecf\u9488\u5bf9\u95ed\u6e90\u548c\u5f00\u6e90\u6a21\u578b\u63d0\u51fa\u4e86\u82e5\u5e72\u65b9\u6cd5\uff0c\u7136\u800c\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u90fd\u6709\u5c40\u9650\u6027\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u7c92\u5ea6\u6709\u9650\uff08limited granularity\uff09\uff1a<\/strong>\u4e3b\u8981\u5173\u6ce8\u7c97\u7c92\u5ea6\u7684\u4eba\u683c\u7279\u5f81\u3001\u804c\u4e1a\uff08\u5982\u7a0b\u5e8f\u5458\u3001\u4f5c\u5bb6\uff09\uff0c\u5ffd\u7565\u4e86\u89d2\u8272\u5c42\u9762\u4e0a\u66f4\u52a0\u590d\u6742\u3001\u7cbe\u7ec6\u7684\u89d2\u8272\u626e\u6f14\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u7f3a\u5c11\u6570\u636e\u548cbenchmark\uff1a<\/strong>\u7f3a\u4e4f\u9ad8\u8d28\u91cf\u3001\u591a\u6837\u5316\u548c\u5e7f\u6cdb\u7684\u5f00\u6e90\u6570\u636e\u96c6\uff0c\u4e5f\u7f3a\u4e4f\u8bc4\u4f30\u57fa\u51c6;<\/p>\n<\/li>\n<li>\n<p><strong>API\u548c\u4e0a\u4e0b\u6587\u6210\u672c\uff1a<\/strong>ChatGPT\u548cGPT-4\u7b49\u4f9d\u8d56\u95ed\u6e90\u6a21\u578b\u7684\u65b9\u6cd5\u65e0\u6cd5\u81ea\u7531\u5fae\u8c03\uff0c\u56e0\u6b64\u9700\u8981\u5728\u63d0\u793a\u8bcd\u4e2d\u5305\u542b\u6240\u6709\u8865\u5145\u4fe1\u606f\uff0c\u4e0d\u5fc5\u8981\u5730\u5360\u7528\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u6b64\u5916\uff0cAPI\u6210\u672c\u9ad8\u5f97\u4ee4\u4eba\u671b\u800c\u5374\u6b65\u3002\u56e0\u6b64\uff0c\u63a2\u7d22\u6700\u5c0f\u5316\u4e0a\u4e0b\u6587\u7a97\u53e3\u5229\u7528\u7387\u5e76\u57fa\u4e8e\u5fae\u8c03\u5f00\u6e90\u6a21\u578b\u7684\u89e3\u51b3\u65b9\u6848\u662f\u503c\u5f97\u7814\u7a76\u7684\u3002<br \/>\n\u5728\u672c\u6587\u4e2d\uff0c\u4f5c\u8005\u4ecb\u7ecd\u4e86RoleLLM\uff0c\u4e00\u4e2a\u7528\u4e8e\u6570\u636e\u6784\u5efa\u3001\u8bc4\u4f30\u548c\u89e3\u51b3\u65b9\u6848\u7684\u89d2\u8272\u626e\u6f14\u6846\u67b6\uff0c\u9002\u7528\u4e8e\u95ed\u6e90\u548c\u5f00\u6e90\u6a21\u578b\u3002\u5728\u4e0b\u56fe\u4e2d\uff0cRoleLLM\u5305\u542b\u56db\u4e2a\u5173\u952e\u9636\u6bb5\uff1a<\/p>\n<\/li>\n<li>\n<p><strong>\u89d2\u8272\u914d\u7f6e\u6587\u4ef6\u6784\u5efa\uff08Role Profile Construction\uff09<\/strong>\uff1a\u6211\u4eec\u4ece916\u4e2a\u82f1\u6587\u548c24\u4e2a\u4e2d\u6587\u516c\u5f00\u53f0\u672c\u4e2d\u7cbe\u5fc3\u6311\u9009\u4e8695\u4e2a\u82f1\u6587\u548c5\u4e2a\u4e2d\u6587\u89d2\u8272\uff0c\u5728\u7ec6\u7c92\u5ea6\u7684\u89d2\u8272\u7ea7\u522b\u4e0a\u6784\u5efa\u4e86\u4e0d\u540c\u4e2a\u6027\u7684\u89d2\u8272\u914d\u7f6e\u6587\u4ef6;<\/p>\n<\/li>\n<li>\n<p><strong>\u57fa\u4e8e\u4e0a\u4e0b\u6587\u7684\u6307\u4ee4\u751f\u6210\uff08Context-Based Instruction Generation (Context-Instruct):<\/strong>\u4f7f\u7528GPT\u4ece\u5206\u5272\u7684\u914d\u7f6e\u6587\u4ef6\u4e2d\u751f\u6210\u9ad8\u8d28\u91cf\u7684QA\u5bf9\uff0c\u4ee5\u63d0\u53d6\u89d2\u8272\u7279\u5b9a\u7684\u77e5\u8bc6\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u57fa\u4e8e\u89d2\u8272\u626e\u6f14\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b\uff08Role Prompting using GPT (RoleGPT)\uff09:<\/strong>\u6211\u4eec\u901a\u8fc7\u57fa\u4e8e\u5bf9\u8bdd\u5de5\u7a0b\u7684\u89d2\u8272\u63d0\u793a\uff0c\u5229\u7528\u7cfb\u7edf\u6307\u4ee4\u548c\u68c0\u7d22\u589e\u5f3a\uff0c\u5728GPT\u4e2d\u57f9\u517b\u89d2\u8272\u626e\u6f14\u80fd\u529b\uff0c\u4ee5\u4ea7\u751f\u5b9a\u5236\u7684\u53e3\u8bed\u6a21\u4eff\u53cd\u5e94;<\/p>\n<\/li>\n<li>\n<p><strong>\u89d2\u8272\u6761\u4ef6\u6307\u4ee4\u8c03\u4f18\uff08Role-Conditioned Instruction Tuning (RoCIT)\uff09:<\/strong>\u901a\u8fc7\u5fae\u8c03\u5f00\u6e90LLaMA\uff0c\u83b7\u5f97\u4e86RoleLLaMA\u548cRoleGLM\u3002RoleBench\u662f\u7b2c\u4e00\u4e2a\u7528\u4e8e\u7ec6\u7c92\u5ea6\u89d2\u8272\u626e\u6f14\u7684\u7cfb\u7edf\u6307\u4ee4\u8c03\u4f18\u6570\u636e\u96c6\u548c\u57fa\u51c6\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg1.jpg\" alt=\"\" \/><br \/>\n\u5728\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u4e09\u4e2aRouge-L-based\u6307\u6807\u6765\u8bc4\u4f30\u6a21\u578b\u7684\u8bf4\u8bdd\u98ce\u683c\u6a21\u4eff\uff08speaking style imitation\uff09\u3001\u56de\u7b54\u51c6\u786e\u6027\uff08answering accuracy\uff09\u548c\u89d2\u8272\u7279\u5b9a\u77e5\u8bc6\u6355\u83b7\uff08role-specific knowledge capturing\uff09\uff0c\u4e0eAlpacaEval\u4e00\u81f4\u7684GPT\u8bc4\u4f30\u4e5f\u88ab\u4f7f\u7528\u3002\u4f5c\u8005\u7684\u4e3b\u8981\u7ed3\u8bba\u5305\u62ec\uff1a<\/p>\n<\/li>\n<li>\n<p>RoleGPT\u6bd4\u63d0\u793a\u5de5\u7a0b\u7684\u8bc4\u4f30\u7ed3\u679c\u66f4\u597d<\/p>\n<\/li>\n<li>\n<p>RoleBench\u663e\u8457\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u89d2\u8272\u626e\u6f14\u80fd\u529b\uff0c\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u751a\u81f3\u8fbe\u5230\u4e86\u4e0eRoleGPT\u7ade\u4e89\u7684\u6548\u679c\u3002<\/p>\n<\/li>\n<li>\n<p>RoleLLaMA\u5728\u8bf4\u8bdd\u53d1\u98ce\u683c\u6a21\u4eff\u548c\u5bf9\u4f4d\u7f6e\u89d2\u8272\u7684\u51c6\u786e\u6027\u65b9\u9762\u8868\u73b0\u51fa\u5f3a\u5927\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u53ea\u9700\u8981\u89d2\u8272\u63cf\u8ff0\u548c\u53e3\u5934\u8bed\u5373\u53ef\u6709\u6548\u9002\u5e94\uff0c\u7528\u6237\u53ef\u4ee5\u65e0\u7f1d\u5b9a\u5236\u65b0\u7684\u89d2\u8272\u3002<\/p>\n<\/li>\n<li>\n<p>\u57fa\u4e8e\u7cfb\u7edf\u6307\u4ee4\u7684\u65b9\u6cd5\u5728\u89d2\u8272\u5b9a\u5236\u6709\u6548\u6027\u548c\u4e0a\u4e0b\u6587\u6548\u7387\u65b9\u9762\u4f18\u4e8e\u68c0\u7d22\u589e\u5f3a\u65b9\u6cd5\u3002<\/p>\n<\/li>\n<li>\n<p>Context-Instruct\u663e\u8457\u589e\u5f3a\u4e86\u6a21\u578b\u5bf9\u89d2\u8272\u7684\u989d\u7684\u8ba4\u8bc6\uff0c\u5728\u4f7f\u7528\u6709\u566a\u58f0\u7684\u89d2\u8272\u914d\u7f6e\u65f6\u4f18\u4e8e\u57fa\u4e8e\u68c0\u7d22\u589e\u5f3a\u7684\u65b9\u6cd5\u3002<\/p>\n<\/li>\n<\/ul>\n<h1>\u65b9\u6cd5<\/h1>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u63cf\u8ff0\u6211\u4eec\u7684\u89d2\u8272\u626e\u6f14\u65b9\u6cd5\u3002\u6211\u4eec\u9996\u5148\u4ecb\u7ecd\u6211\u4eec\u7684\u89e3\u51b3\u65b9\u6848\u7684\u8bbe\u8ba1\u539f\u5219\u3002\u7136\u540e\uff0c\u6211\u4eec\u8bf4\u660e\u4e86\u4e24\u79cd\u89d2\u8272\u626e\u6f14\u6570\u636e\u589e\u5f3a\u673a\u5236:RoleGPT\u548cContext-Instruct\u3002\u6700\u540e\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e0e\u57fa\u4e8e\u7cfb\u7edf\u6307\u4ee4\u7684\u89d2\u8272\u5b9a\u5236\u8fc7\u7a0b\u76f8\u5173\u7684\u89d2\u8272\u6761\u4ef6\u6307\u4ee4\u8c03\u4f18(RoCIT)\u3002<\/p>\n<h2>\u8bbe\u8ba1\u539f\u5219<\/h2>\n<p><strong>\u6a21\u4eff\u8bf4\u8bdd\u98ce\u683c\u3002<\/strong>\u4e3a\u4e86\u6a21\u4eff\u7279\u5b9a\u89d2\u8272\u7684\u8bf4\u8bdd\u98ce\u683c\uff0c\u6a21\u578b\u5bf9\u6307\u4ee4\u7684\u56de\u5e94\u9700\u8981\u7b26\u5408\u4e24\u4e2a\u6807\u51c6\uff1a\uff081\uff09<strong>\u8bcd\u6c47\u4e00\u81f4\u6027\uff08Lexical Consistency\uff09\uff1a<\/strong>\u6a21\u578b\u7684\u56de\u7b54\u5e94\u8be5\u5305\u542b\u89d2\u8272\u5e38\u7528\u7684\u53e3\u5934\u7985\u6216\u4e60\u60ef\u7528\u8bed\uff0c\u4ee5\u786e\u4fdd\u8bcd\u6c47\u4e0e\u89d2\u8272\u72ec\u7279\u7684\u8bed\u8a00\u98ce\u683c\u4fdd\u6301\u4e00\u81f4\u3002\uff082\uff09<strong>\u5bf9\u8bdd\u4fdd\u771f\u5ea6\uff08Dialogic Fidelity\uff09\uff1a<\/strong>\u6a21\u578b\u7684\u7b54\u6848\u4e0d\u4ec5\u5e94\u8be5\u5728\u8bed\u5883\u4e0a\u5408\u9002\uff0c\u800c\u4e14\u5728\u98ce\u683c\u4e0a\u4e0e\u89d2\u8272\u7684\u793a\u4f8b\u5bf9\u8bdd\u76f8\u4f3c\u3002\u4f8b\u5982\uff0c\u6d77\u76d7\u89d2\u8272\u7684\u8bcd\u6c47\u4e00\u81f4\u6027\u5305\u62ec\u9891\u7e41\u4f7f\u7528\u822a\u6d77\u672f\u8bed\uff08aweigh\uff09\u548c\u53e3\u5934\u8bed\u201cmatey\u201d\u6216\u201cahoy\u201d\u3002\u6b64\u5916\uff0c\u5bf9\u8bdd\u7684\u4fdd\u771f\u5ea6\u5e94\u8be5\u6355\u6349\u89d2\u8272\u72ec\u7279\u7684\u8bed\u6cd5\u548c\u8bed\u8c03\uff08\u4f8b\u5982\uff1a\u53e3\u8bed\u8868\u8fbe\u3001\u7c97\u9c81\u7684\u8bf4\u8bdd\u65b9\u5f0f\u548c\u65e0\u6cd5\u65e0\u5929\u7684\u8bed\u6c14\uff09<br \/>\n<strong>\u89d2\u8272\u7279\u5b9a\u77e5\u8bc6\u548c\u8bb0\u5fc6\u7684\u6ce8\u5165\u3002<\/strong>\u89d2\u8272\u626e\u6f14\u7684\u53e6\u4e00\u4e2a\u91cd\u8981\u65b9\u9762\u662f\u704c\u8f93\u7279\u5b9a\u89d2\u8272\u7684\u77e5\u8bc6\u548c\u60c5\u666f\u8bb0\u5fc6\u3002\u8003\u8651\u4e24\u79cd\u5427\u5189\u7684\u77e5\u8bc6\u7c7b\u522b\uff1a\uff081\uff09\u57fa\u4e8e\u53f0\u672c\u7684\u77e5\u8bc6\uff0c\u6d89\u53ca\u8bb0\u5f55\u5728\u811a\u672c\u4e2d\u7684\u660e\u786e\u7ec6\u8282\uff0c\u4f8b\u5982\u8be6\u7ec6\u7684\u89d2\u8272\u80cc\u666f\u3001\u60c5\u666f\u8bb0\u5fc6\u548c\u89d2\u8272\u7ecf\u5386\u7684\u7279\u5b9a\u4e8b\u4ef6\u3002\uff082\uff09\u5267\u672c\u4e2d\u6ca1\u6709\u660e\u786e\u7684\u77e5\u8bc6\uff0c\u5305\u62ec\u89d2\u8272\u53ef\u80fd\u62e5\u6709\u7684\u4e00\u822c\u77e5\u8bc6\u6216\u4e13\u4e1a\u77e5\u8bc6\u3002\u4f8b\u5982\uff0c\u5f53\u626e\u6f14\u94a2\u94c1\u4fa0\u65f6\uff0cllm\u5e94\u8be5\u5305\u542b\u57fa\u4e8e\u5267\u672c\u7684\u77e5\u8bc6\uff08\u4f8b\u5982\u4f55\u65f6\u521b\u5efa\u4e86\u7b2c\u4e00\u4e2a\u94a2\u94c1\u4fa0\u5957\u88c5\uff09\u548c\u4f5c\u4e3a\u4f01\u4e1a\u5bb6\u4e0e\u5267\u672c\u65e0\u5173\u7684\u77e5\u8bc6\uff08\u4f8b\u5982\u5546\u4e1a\u5934\u8111\u3001\u9886\u5bfc\u7d20\u8d28\u548c\u6280\u672f\u4e13\u957f\u7b49\uff09\u3002<\/p>\n<h2>RoleGPT\uff1a\u901a\u8fc7prompt\u6fc0\u53d1\u6a21\u578b\u7684\u89d2\u8272\u626e\u6f14\u80fd\u529b<\/h2>\n<p>\u8003\u8651\u5230\u5fae\u8c03\u7684\u9650\u5236\uff0c\u4e3a\u89d2\u8272\u626e\u6f14GPT\u901a\u5e38\u4f7f\u7528\u63d0\u793a\u5de5\u7a0b\uff0c\u5982zero-shot\u548cfew-shot\uff08in context learning\uff09\u3002\u7136\u800c\uff0c\u5bf9\u4e8eChatGPT\u548cGPT4\u6765\u8bf4\uff0c\u4ed6\u4eec\u5df2\u7ecf\u653e\u5f03\u4e86\u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd\u529b\u6765\u8fdb\u884c\u5bf9\u8bdd\u5386\u53f2\u5efa\u6a21\uff0c\u4f20\u7edf\u7684zero-shot\u63d0\u793a\u5e76\u4e0d\u8db3\u4ee5\u5145\u5206\u6fc0\u53d1\u89d2\u8272\u626e\u6f14\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u4f5c\u8005\u4fee\u6539\u4e86\u63d0\u5bf9\u8bdd\u5de5\u7a0b\u4e2d\u7684few-shot\u63d0\u793a\u8bcd\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg2.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg3.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg4.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg5.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg6.jpg\" alt=\"\" \/><\/p>\n<p>\u5177\u4f53\u7684\uff0c\u662f\u7531GPT-4\u751f\u6210\u89d2\u8272\u63cf\u8ff0\u548c\u77ed\u8bed\u4f5c\u4e3a\u7cfb\u7edf\u6307\u4ee4\u7684\u6838\u5fc3\uff0c\u7136\u540e\uff0c\u6211\u4eec\u521b\u5efa\u4e86\u4e00\u4e2a\u6574\u4f53\u7684\u89d2\u8272\u626e\u6f14\u4efb\u52a1\u6307\u4ee4\uff0c\u5982\u201c\u8bf7\u5411{role_name}\u90a3\u6837\u8bf4\u8bdd\u201d\u3002\u5e76\u4f7f\u7528BM25\u5728\u89d2\u8272\u7b80\u4ecb\u4e2d\u68c0\u7d22\u524d\u4e94\u4e2a\u76f8\u5173\u5bf9\u8bdd\u4f5c\u4e3afew-shot\u7684\u793a\u8303\u3002\u901a\u8fc7\u8fd9\u6837\u505a\uff0cRoleGPT\u7684\u56de\u7b54\u53ef\u4ee5\u6355\u6349\u89d2\u8272\u7684\u8bf4\u8bdd\u98ce\u683c\uff0c\u5e76\u5305\u542b\u4e00\u4e9b\u89d2\u8272\u7279\u5b9a\u7684\u77e5\u8bc6\u3002\u7136\u800c\uff0c\u8f6e\u5ed3\u7684\u7a00\u758f\u6027\u548c\u566a\u58f0\u9650\u5236\u4e86\u68c0\u7d22\u589e\u5f3a\u77e5\u8bc6\u53d1\u73b0\u7684\u6709\u6548\u6027\u3002<\/p>\n<h2>Context-Instruct:\u57fa\u4e8e\u4e0a\u4e0b\u6587\u7684\u6307\u4ee4\u751f\u6210<\/h2>\n<p>\u4e3a\u4e86\u63d0\u9ad8\u5408\u6210\u6307\u4ee4\u6570\u636e\u96c6\u4e2d\u89d2\u8272\u7279\u5b9a\u77e5\u8bc6\u7684\u5bc6\u5ea6\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e0a\u4e0b\u6587\u6307\u4ee4\u6765\u8fdb\u884c\u957f\u6587\u672c\u77e5\u8bc6\u63d0\u53d6\u548c\u6307\u4ee4\u6570\u636e\u751f\u6210\u3002\u9488\u5bf9\u89d2\u8272\u7684\u6307\u4ee4\u6570\u636e\u751f\u6210\u5305\u62ec\u4e09\u4e2a\u6b65\u9aa4:(1)\u5206\u5272\u89d2\u8272\u6982\u8981;(2)\u751f\u6210\u95ee\u9898-\u7f6e\u4fe1\u5ea6-\u7b54\u6848\u4e09\u5143\u7ec4\u5019\u9009\u4eba;(3)\u4f4e\u8d28\u91cf\u6570\u636e\u7684\u6ee4\u6ce2\u548c\u540e\u5904\u7406\u3002\u6211\u4eec\u5c06\u7b80\u8981\u6982\u8ff0\u5982\u4e0b\u3002\u8be6\u60c5\u8bf7\u53c2\u9605\u9644\u5f55C\u3002<\/p>\n<ul>\n<li><strong> \u89d2\u8272\u914d\u7f6e\u6587\u4ef6\u5206\u5272\uff1a<\/strong>\u8003\u8651\u5230GPT\u4e0a\u4e0b\u6587\u5927\u5c0f\u6709\u9650\uff0c\u5c06\u89d2\u8272\u914d\u7f6e\u6587\u4ef6\u5212\u5206\u4e3a\u66f4\u6613\u7ba1\u7406\u7684\u51e0\u4e2a\u90e8\u5206\uff0c\u5177\u4f53\u5305\u62ec\uff1a\uff081\uff09\u89d2\u8272\u63cf\u8ff0\u548c\u53e3\u5934\u8bed\uff1b\uff082\uff09\u7ed3\u6784\u5316\u5bf9\u8bdd\u3002\u7b2c\u4e00\u5206\u90e8\u7528\u6765\u83b7\u5f97\u4e0e\u53f0\u672c\u65e0\u5173\u7684\u6307\u4ee4\uff0c\u800c\u7b2c\u4e8c\u90e8\u5206\u7528\u6765\u6784\u5efa\u57fa\u4e8e\u53f0\u672c\u7684\u6307\u4ee4\u3002<\/li>\n<li><strong>\u6307\u4ee4\u548c\u56de\u590d\u751f\u6210\uff1a<\/strong>\u5982\u56fe1\u6240\u793a\uff0c\u4e3a\u7279\u5b9a\u89d2\u8272\u7684\u6307\u4ee4\u6570\u636e\u751f\u6210\u5019\u9009\u6570\u636e\u7684\u8fc7\u7a0b\u4e2d\uff0c\u9700\u8981\u8003\u8651\u4e09\u4e2a\u5143\u7d20\uff1a\u95ee\u9898<span class=\"katex-eq\" data-katex-display=\"false\">Q<\/span>\uff0c\u76f8\u5e94\u7684\u7b54\u6848<span class=\"katex-eq\" data-katex-display=\"false\">A<\/span>\u548c\u7f6e\u4fe1\u5ea6<span class=\"katex-eq\" data-katex-display=\"false\">C<\/span>\u3002\u5229\u7528\u5927\u6a21\u578b\u4e3a\u6bcf\u4e2a\u89d2\u8272\u548c\u60c5\u8282\u751f\u6210\u8fd9\u4e9b\u4e09\u5143\u7ec4\u3002\u521d\u6b65\u5b9e\u9a8c\u8868\u660e\uff0c\u5728\u6ca1\u6709\u7f6e\u4fe1\u5ea6\u8bc4\u5206\u7684\u60c5\u51b5\u4e0b\u751f\u6210QA\u5bf9\u4f1a\u5bfc\u81f4\u8d28\u91cf\u95ee\u9898\u3002\u5bf9\u4e8e\u57fa\u4e8e\u811a\u672c\u7684\u6307\u4ee4\uff0c\u7531\u4e8e\u5047\u8bbe\u6709\u5148\u9a8c\u77e5\u8bc6\uff0c\u901a\u5e38\u4f1a\u51fa\u73b0\u4e0d\u5b8c\u6574\uff1b\u5bf9\u4e8e\u811a\u672c\u4e0d\u53ef\u77e5\u7684\u6307\u4ee4\uff0c\u7531\u4e8e\u7f3a\u4e4f\u4e0a\u4e0b\u6587\u800c\u5305\u542b\u5e7b\u89c9\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6a21\u578b\u88ab\u63d0\u793a\u751f\u6210\u8bc4\u4ef7\u95ee\u9898\u5b8c\u6574\u6027\u6216\u771f\u5b9e\u6027\u7684\u7f6e\u4fe1\u5ea6\u8bc4\u5206\u3002\u63d0\u793a\u6a21\u677f\u5305\u62ec\u5bf9\u4e8e\u89d2\u8272\u7684\u63cf\u8ff0\u3001\u53e3\u5934\u7985\u3001few-shot\u793a\u4f8b\u548c\u6a21\u4eff\u89d2\u8272\u8bf4\u8bdd\u98ce\u683c\u7684\u4e09\u5143\u7ec4\u6307\u4ee4\u751f\u6210\u3002\u751f\u6210\u8fc7\u7a0b\u4e2d\u4e3a\u6bcf\u4e2a\u4eba\u7269\u8dd1\u81f3\u5c11400\u4e2a\u5bf9\u8bdd\u3002<\/li>\n<li>\u6570\u636e\u8fc7\u6ee4\u548c\u540e\u5904\u7406\uff1a\u8fc7\u6ee4\u8fc7\u7a0b\u5305\u62ec\u57fa\u4e8e\u7f6e\u4fe1\u5ea6\u5206\u6570\u7684\u8fc7\u6ee4\u548c\u91cd\u590d\u6570\u636e\u7684\u5220\u9664\uff0c\u4ee5\u4fdd\u8bc1\u6570\u636e\u8d28\u91cf\u548c\u591a\u6837\u6027\uff0c\u6709\u5173\u8fc7\u6ee4\u548c\u540e\u671f\u5904\u7406\u7684\u8be6\u60c5\uff0c\u8bf7\u53c2\u8003\u9644\u5f55C<\/li>\n<\/ul>\n<h2>RoCIT\uff1aRole-Conditioned Instruction Tuning<\/h2>\n<p>\u6709\u4e24\u79cd\u7c7b\u578b\u7684\u589e\u5f3a\u6570\u636e\uff1a\u4e00\u79cd\u662f\u7528\u4e8eRoleGPT\u751f\u6210\u7684\u901a\u7528\u9886\u57df\u6307\u4ee4\uff0c\u53e6\u4e00\u79cd\u7528\u4e8e\u901a\u8fc7Context-Instruct\u751f\u6210\u7684\u89d2\u8272\u7279\u5b9a\u6307\u4ee4\u3002\u5bf9\u8fd9\u4e9b\u6570\u636e\u8fdb\u884c\u5fae\u8c03\u4e0d\u4ec5\u53ef\u4ee5\u6539\u5584\u6a21\u578b\u7684\u8bf4\u8bdd\u89c4\u5219\uff0c\u8fd8\u53ef\u4ee5\u5c06\u89d2\u8272\u7684\u7279\u5b9a\u77e5\u8bc6\u5d4c\u5165\u5230\u6a21\u578b\u6743\u91cd\u4e2d\u3002\u5c06\u5176\u5e94\u7528\u4e8e\u82f1\u8bed\u7684LLaMA\u548c\u4e2d\u6587\u7684ChatGLM2\uff0c\u5f97\u5230\u4e86RoleLLaMA\u548cRoleGLM\u3002\u4e0e\u666e\u901aSFT\u4e0d\u540c\uff0c\u91c7\u7528\u89d2\u8272\u6761\u4ef6\u5fae\u8c03\uff0c\u7ee7\u627f\u4e86\u89d2\u8272\u5b9a\u5236\u7684\u7279\u5b9a\u7b56\u7565\uff0c\u5305\u62ec\u7cfb\u7edf\u6307\u4ee4\u548c\u68c0\u7d22\u589e\u5f3a\u3002<br \/>\n\u7cfb\u7edf\u6307\u4ee4\u6307\u5b9a\uff1a\u5728RoCIT\u4e2d\uff0c\u50cf\u5728RoleGPT\u4e2d\u4e00\u6837\uff0c\u5728\u8f93\u5165\u524d\u6dfb\u52a0\u4e00\u4e2a\u7cfb\u7edf\u6307\u4ee4\uff0c\u5176\u4e2d\u5305\u542b\u89d2\u8272\u540d\u79f0\u3001\u63cf\u8ff0\u3001\u53e3\u5934\u8bed\u548c\u89d2\u8272\u626e\u6f14\u6307\u4ee4\u4efb\u52a1\u3002RoleLLaMA\u7684\u804a\u5929\u6807\u8bb0\u8bed\u8a00\u4e3a:<\/p>\n<pre><code>### Instrcution:\n{system instruction}&lt;\/s&gt;\n\n### Input:\n{user input}&lt;\/s&gt;\n\n### Response:\n{model response}&lt;\/s&gt;<\/code><\/pre>\n<p>\u503c\u76d1\u7763\u56de\u7b54\u7684\u7279\u6b8a\u4ee4\u724c\u3002\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u7528\u6237\u53ef\u4ee5\u5f88\u5bb9\u6613\u7684\u901a\u8fc7\u7cfb\u7edf\u6307\u4ee4\u4fee\u6539LLM\u7684\u89d2\u8272\uff0c\u4e0e\u68c0\u7d22\u589e\u5f3a\u76f8\u6bd4\uff0c\u6700\u5c0f\u5316\u4e86\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u6d88\u8017\u3002<\/p>\n<h1>RoleBench\uff1a\u89d2\u8272\u626e\u6f14\u80fd\u529b\u6d4b\u8bd5\u57fa\u51c6<\/h1>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u4ecb\u7ecd\u4e86\u6784\u5efaRoleBench\u5e76\u5206\u6790\u3002RoleBench\u53ef\u4ee5\u7528\u6765\u8bc4\u4f30\u89d2\u8272\u626e\u6f14\u80fd\u529b\uff0c\u5e76\u63d0\u9ad8\u8fd9\u79cd\u80fd\u529b\u3002<\/p>\n<h2>\u6570\u636e\u6784\u5efa<\/h2>\n<p>RoleBench\u6570\u636e\u96c6\u7684\u6784\u5efa\u5305\u62ec\u4e94\u4e2a\u6b65\u9aa4\uff1a\uff081\uff09\u89d2\u8272\u9009\u62e9\uff1b\uff082\uff09\u89d2\u8272\u914d\u7f6e\u6784\u5efa\uff1b\uff083\uff09\u901a\u7528\u6307\u4ee4\u62bd\u6837\uff1b\uff084\uff09\u751f\u6210\u539f\u59cbRoleBench\u6570\u636e\uff1b\uff085\uff09\u6570\u636e\u6e05\u6d17\u3002\u9996\u5148\uff0c\u5728GPT-4\u7684\u5e2e\u52a9\u4e0b\uff0c\u4ece\u5305\u62ecNLP\u7535\u5f71\u811a\u672c\u548c\u4eba\u5de5\u7b5b\u9009\u7684\u4e2d\u6587\u811a\u672c\u4e2d\u6311\u9009\u4e86100\u4e2a\u5177\u6709\u4ee3\u8868\u6027\u548c\u72ec\u7279\u6027\u7684\u89d2\u8272\u3002\u5176\u6b21\uff0c\u89d2\u8272\u914d\u7f6e\u7531GPT-4\u751f\u6210\u7684\u89d2\u8272\u63cf\u8ff0\u548c\u53e3\u5934\u7985\u4ee5\u53ca\u4ece\u811a\u672c\u89e3\u6790\u7684\u7ed3\u6784\u5316\u5bf9\u8bdd\u7ec4\u6210\uff0c\u5e76\u7ecf\u8fc7\u4eba\u5de5\u9a8c\u8bc1\u3002\u7b2c\u4e09\uff0c\u4ece\u591a\u4e2a\u6570\u636e\u96c6\u4e2d\u968f\u673a\u62bd\u53d61500\u6761\u82f1\u8bed\u901a\u7528\u6307\u4ee4\uff0c\u5305\u62ecSuper-NaturalInstruct\u3001UltraChat\u3001 Alpaca\u2019s\u3002\u8fd8\u5bf9COIG\u548cBELLE\u8fdb\u884c\u62bd\u6837\uff0c\u83b7\u5f971479\u7bc7\u4e2d\u6587\u901a\u7528\u8bf4\u660e\u3002\u6240\u6709\u91c7\u6837\u6307\u4ee4\u5305\u542b\u4e0d\u8d85\u8fc7100\u4e2a\u5355\u8bcd\u3002\u6839\u636eBM25\u76f8\u4f3c\u6027\u8fdb\u884c\u53bb\u91cd\u3002\u7b2c\u56db\uff0c\u4f7f\u7528RoleGPT\u83b7\u5f97\u6bcf\u4e2a\u901a\u7528\u6307\u4ee4\u7684\u591a\u4e2a\u54cd\u5e94(RoleBench-general)\uff0c\u5e76\u4f7f\u7528\u4e0a\u4e0b\u6587-\u6307\u4ee4\u751f\u6210\u7279\u5b9a\u89d2\u8272\u7684\u95ee\u7b54\u5bf9(RoleBench-specific)\u3002\u6700\u540e\uff0c\u5bf9\u83b7\u5f97\u7684\u539f\u59cb\u6570\u636e\u96c6\u8fdb\u884c\u5f7b\u5e95\u6e05\u7406\uff0c\u4ee5\u786e\u4fdd\u56de\u7b54\u7684\u5b8c\u6574\u6027\u3001\u4eba\u5de5\u667a\u80fd\u548c\u89d2\u8272\u8eab\u4efd\u7684\u9690\u853d\u6027\u548c\u975e\u62d2\u7edd\u6027\u3002\u8be6\u60c5\u5185\u5bb9\u89c1\u9644\u5f55D\u3002<\/p>\n<h2>\u6570\u636e\u5206\u6790<\/h2>\n<p>\u4f7f\u7528GPT4\u5f97\u5230RoleBench-general-en\uff08\u82f1\u6587\uff09\u548cRoleBench-general-zh\uff08\u4e2d\u6587\uff09\u3002Context-Instruct\u7684GPT-3.5\u83b7\u5f97RoleBench-specific-en\u548cRoleBench-specific-zh\u3002\u89d2\u8272\u9009\u62e9\u3001\u63cf\u8ff0\u548c\u53e3\u5934\u7985\u751f\u6210\u90fd\u662f\u7528GPT4\u751f\u6210\u7684\u3002<\/p>\n<ul>\n<li>\u7edf\u8ba1\u548c\u8d28\u91cf\uff1a\u8868\u4e2d\u63d0\u4f9b\u4e86RoleBench\u7684\u57fa\u672c\u7edf\u8ba1\u4fe1\u606f\u3002\u4ece\u901a\u7528\u548c\u89d2\u8272\u7279\u5b9a\u81ea\u5df1\u79cd\u968f\u673a\u62bd\u53d6100\u4e2a\u5b9e\u4f8b\u8bc4\u4f30RoleBench\u7684\u8d28\u91cf\uff0c\u7136\u540e\u5176\u8bf7\u4e00\u540d\u4e13\u5bb6\u6807\u6ce8\u4ece\u4e09\u4e2a\u65b9\u9762\u8bc4\u4f30\u8d28\u91cf\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u5927\u90e8\u5206\u662f\u9ad8\u8d28\u91cf\u7684\u3002\u5173\u4e8e\u6709\u6548\u548c\u65e0\u6548\u7684\u793a\u4f8b\u53c2\u8003\u4e0b\u56fe\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg10.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg11.jpg\" alt=\"\" \/><\/li>\n<li>\u591a\u6837\u6027\uff1a\u5206\u6790\u4e86RoleBench\u7684\u5168\u9762\u6027\u548c\u591a\u6837\u6027\u3002\u56fe4\u663e\u793a\u4e86RoleBench-en\u6307\u4ee4\u7684\u52a8\u8bcd-\u540d\u8bcd\u7ed3\u6784\uff0c\u5176\u4e2d\u63cf\u7ed8\u4e86\u524d10\u4e2a\u52a8\u8bcd(\u5185\u5708)\u548c\u524d4\u4e2a\u76f4\u63a5\u540d\u8bcd\u5bbe\u8bed(\u5916\u5708)\uff0c\u5360\u6307\u4ee4\u76845.6%\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5c06RoleBench\u89d2\u8272\u5206\u7c7b\u4e3a\u591a\u4e2a\u7c7b\uff0c\u5e76\u6784\u5efa\u4e00\u4e2a\u8bcd\u4e91\u6765\u663e\u793a\u5b83\u4eec\u7684\u591a\u6837\u6027\uff0c\u5982\u56fe3\u6240\u793a\u3002\u6307\u4ee4\u548c\u54cd\u5e94\u7684\u957f\u5ea6\u5206\u5e03\u5982\u56fe5\u6240\u793a\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg12.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg13.jpg\" alt=\"\" \/><\/li>\n<\/ul>\n<h1>\u5b9e\u9a8c<\/h1>\n<h2>\u5b9e\u9a8c\u8bbe\u7f6e<\/h2>\n<p>RoleLLaMA&amp;RoleGLM\uff1a\u4f7f\u7528RoleBench-general-en\u548cRoleBench-specific-en\u5bf9LLaMA-7B model\u8fdb\u884clora\u5fae\u8c03\u3002\u4f7f\u7528 RoleBench-general-zh\u548cRoleBench-specific-zh\u5bf9ChatGLM2-6B\u8fdb\u884clora\u5fae\u8c03\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cLLaMA-7B\u53ea\u662f\u4e00\u79cd\u9884\u8bad\u7ec3\u6a21\u578b\uff0cChatGLM2-6B\u7ecf\u8fc7\u540e\u8bad\u7ec3\uff0c\u5177\u6709\u589e\u5f3a\u7684\u6307\u4ee4\u8ddf\u968f\u548c\u5bf9\u8bdd\u80fd\u529b\u3002<\/p>\n<h2>Baseline<\/h2>\n<p>RoleGPT\u662f\u4e2d\u3001\u82f1\u89d2\u8272\u7684\u57fa\u7ebf\u3002\u5bf9\u4e8e\u82f1\u6587\u89d2\u8272\uff0cLLaMA-7B\u548c\u5b83\u7684\u6307\u4ee4\u5fae\u8c03\u53d8\u4f53\u3001Vicuna-13B\u548cAlpaca-7B\u4f5c\u4e3a\u57fa\u7ebf\u3002\u5bf9\u4e8e\u4e2d\u6587\u89d2\u8272\uff0cChatGLM2-6B\u4f5c\u4e3a\u57fa\u7ebf\u3002\u6211\u4eec\u8fd8\u7ed3\u5408\u4e86\u5728\u591a\u56de\u5408\u5bf9\u8bdd\u6a21\u5f0f\u4e0b\u5bf9\u811a\u672c\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u7684\u6a21\u578b\u4f5c\u4e3a\u989d\u5916\u7684\u57fa\u7ebf\u3002<\/p>\n<h2>\u8bc4\u4f30\u65b9\u6cd5<\/h2>\n<p>\u4f7f\u7528Rouge-L\u6765\u8bc4\u4ef7\u6a21\u578b\u9884\u6d4b\u548c\u57fa\u7840\u4e8b\u5b9e\u95f4\u7684\u91cd\u53e0\u3002\u57fa\u672c\u4e8b\u5b9e\u5305\u542b\u4e09\u7c7b\uff1a<br \/>\n\uff081\uff09\u65e0\u89d2\u8272\u626e\u6f14\u7684\u901a\u7528\u6307\u4ee4\u7684ground truth(RAW)\uff1a\u8bc4\u4f30\u56de\u590d\u7684\u51c6\u786e\u6027<br \/>\n\uff082\uff09\u5e26\u6709\u89d2\u8272\u626e\u6f14\u7684\u5b9a\u5236\u5316\u7684\u901a\u7528\u6307\u4ee4\u56de\u590d(CUS)\uff1a\u8bc4\u4f30\u6a21\u62df\u8bf4\u8bdd\u98ce\u683c\u7684\u80fd\u529b<br \/>\n\uff083\uff09\u89d2\u8272\u7279\u5b9a\u7684\u6307\u4ee4\u56de\u590d\uff08SPE\uff09\uff1a\u8bc4\u4f30\u89d2\u8272\u7279\u5b9a\u7684\u77e5\u8bc6\u548c\u8bb0\u5fc6\u80fd\u529b<br \/>\n\u6b64\u5916\uff0c\u8fd8\u4f7f\u7528GPT\u8fdb\u884c\u8bc4\u4f30\uff0c\u63d0\u793a\u8bcd\u5982\u4e0b\u3002\u8fd9\u4e9b\u63d0\u793a\u6709\u52a9\u4e8e\u6837\u672c\u6bd4\u8f83\u548c\u6392\u540d\uff0c\u4ece\u800c\u83b7\u5f97\u80dc\u7387\u6216\u5e73\u5747\u6392\u540d\u3002\u8bad\u7ec3\u96c6\u5206\u5272\u7b56\u7565\u4e3b\u8981\u96c6\u4e2d\u5728\u4e24\u4e2a\u65b9\u9762\uff1a\uff081\uff09\u57fa\u4e8e\u6307\u4ee4\u7684\u5206\u5272\u2014\u2014\u7528\u4e8e\u8bc4\u4f30\u6307\u4ee4\u6cdb\u5316\uff08instruction generalization\uff09\uff082\uff09\u57fa\u4e8e\u89d2\u8272\u7684\u5206\u5272\u2014\u2014\u7528\u4e8e\u8bc4\u4f30\u89d2\u8272\u6cdb\u5316\uff08 role generalization\uff09\uff0c\u4ec5\u9650\u82f1\u8bed\u3002\u8be6\u7ec6\u6b65\u9aa4\u89c1\u9644\u5f55G<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg15.jpg\" alt=\"\" \/><\/p>\n<h2>\u4e3b\u8981\u5b9e\u9a8c<\/h2>\n<p><strong>\u6307\u4ee4\u6cdb\u5316\uff1a<\/strong>\u88683a\u548c\u88683b\u5206\u522b\u662fRouge-L\u548cGPT\u7528\u4e8e\u8bc4\u4f30\u6307\u4ee4\u6cdb\u5316\u80fd\u529b\u7684\u5206\u6570\u3002\u5bf9\u4e8eGPT-3.5\u8bc4\u4f30\uff0c\u5c06\u6bcf\u4e2a\u6a21\u578b\u4e0eRoleGPT\u6bd4\u8f83\uff0c\u786e\u5b9a\u80dc\u7387\uff0c\u5373GPT-3.5\u5bf9\u6a21\u578b\u7531\u4e8eRoleGPT\u7684\u9891\u7387<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg16.jpg\" alt=\"\" \/><\/p>\n<h1>\u9644\u5f55C<\/h1>\n<h2>\u89d2\u8272\u914d\u7f6e\u5206\u5272<\/h2>\n<p>\u89d2\u8272\u914d\u7f6e\u88ab\u5206\u5272\u4e3a\u4e24\u4e2a\u90e8\u5206\uff081\uff09\u89d2\u8272\u63cf\u8ff0\u548c\u53e3\u5934\u7985\uff1b\uff082\uff09\u4ece\u53f0\u672c\u4e2d\u89e3\u6790\u51fa\u7684\u771f\u5b9e\u4e14\u7ed3\u6784\u5316\u7684\u5bf9\u8bdd\uff0c\u5176\u4e2d\u76ee\u6807\u89d2\u8272\u5728\u6bcf\u4e2a\u5bf9\u8bdd\u56de\u5408\u4e2d\u90fd\u6709\u8f6c\u6298\u3002\u8003\u8651\u5230\u8fd9\u4e9b\u914d\u7f6e\u7684\u6587\u4ef6\u53ef\u80fd\u5f88\u957f\uff0c\u800c\u6211\u4eec\u7684Context-Instruct\u6846\u67b6\u4f9d\u8d56\u4e8e\u4e0a\u4e0b\u6587\u5927\u5c0f\u6709\u9650\u7684llm\uff0c\u5728\u5c06\u8fd9\u4e9b\u957f\u6982\u8981\u6587\u4ef6\u8f93\u5165\u6a21\u578b\u4e4b\u524d\uff0c\u5c06\u5b83\u4eec\u5206\u89e3\u6210\u66f4\u77ed\u7684\u90e8\u5206\u662f\u5fc5\u8981\u7684\u3002\u6211\u4eec\u8ba4\u4e3aGPT\u751f\u6210\u7684\u89d2\u8272\u63cf\u8ff0\u662f\u4e00\u4e2a\u7279\u6b8a\u7684\u90e8\u5206\u3002\u8fd9\u7528\u4e8e\u751f\u6210\u4e0e\u811a\u672c\u65e0\u5173\u7684\u95ee\u9898-\u7f6e\u4fe1-\u56de\u7b54\u4e09\u5143\u7ec4\u3002\u5bf9\u4e8e\u5269\u4f59\u7684\u57fa\u4e8e\u53f0\u672c\u7684\u5185\u5bb9\uff0c\u4f7f\u7528\u7279\u5b9a\u7684\u5206\u5272\u89c4\u5219\uff1a\uff081\uff09\u5728\u4e00\u4e2a\u7247\u6bb5\u4e2d\u6ca1\u6709\u4e0d\u5b8c\u6574\u7684\u5bf9\u8bdd\u56de\u5408\uff1b\uff082\uff09\u6bcf\u4e2a\u7247\u6bb5\u4e0d\u5c11\u4e8e500\u4e2a\u82f1\u6587\u5355\u8bcd\uff0c\u4e2d\u6587\u6982\u8981\u4e0d\u5c11\u4e8e1000\u4e2a\u6c49\u5b57\uff0c\u4fdd\u8bc1\u4e0d\u5c11\u4e8e4\u4e2a\u5bf9\u8bdd\u56de\u5408\uff1b(3)\u6bcf\u4e2a\u5bf9\u8bdd\u56de\u5408\u4e0d\u8d85\u8fc7500\u4e2a\u82f1\u6587\u5355\u8bcd(\u4e2d\u6587\u6982\u8981\u4e0d\u8d85\u8fc7500\u4e2a\u5b57\u7b26)\uff0c\u8d85\u8fc7\u5219\u653e\u5f03\uff1b(4)\u5c06\u6bcf\u4e2a\u5256\u9762\u7684\u6bb5\u6570\u9650\u5236\u4e3a100;\u5982\u679c\u8d85\u8fc7\uff0c\u6211\u4eec\u968f\u673a\u62bd\u6837;(5)\u5982\u679c\u7247\u6bb5\u8d85\u8fc72000\u4e2a\u5355\u8bcd\uff0c\u6211\u4eec\u5c06\u914d\u7f6e\u622a\u65ad\u4e3a2000\u4e2a\u5355\u8bcd\uff0c\u4ee5\u786e\u4fdd\u8fdb\u5165GPT\u7684\u4ee4\u724c\u6570\u91cf\u4e0d\u8d85\u8fc74096\u4e2a\u3002(6)\u6211\u4eec\u4e0d\u4f1a\u57fa\u4e8e\u811a\u672c\u5212\u5206(\u5982\u884c\u4e3a\u6216\u60c5\u8282)\u8fdb\u884c\u5206\u5272\uff0c\u56e0\u4e3a\u8fd9\u6837\u4f1a\u4ea7\u751f\u592a\u5c11\u7684\u7247\u6bb5\uff0c\u5e76\u963b\u788d\u540e\u7eed\u6307\u4ee4\u6570\u636e\u751f\u6210\u7684\u591a\u6837\u6027\u3002\u8bf7\u6ce8\u610f\uff0c\u867d\u7136\u67d0\u4e9b\u7279\u5b9a\u90e8\u5206\u7684\u5bf9\u8bdd\u53ef\u80fd\u4e0d\u4f1a\u4ee5\u611f\u5174\u8da3\u7684\u89d2\u8272\u7ed3\u675f\uff0c\u4f46\u5927\u591a\u6570\u5bf9\u8bdd\u56de\u5408\u7684\u8bbe\u8ba1\u90fd\u662f\u7531\u89d2\u8272\u201c\u89c2\u5bdf\u201d\u5230\u7684\u3002\u8fd9\u662f\u56e0\u4e3a\u6211\u4eec\u7684\u8f6e\u5ed3\u7ed3\u6784\u786e\u4fdd\u6bcf\u4e2a\u5bf9\u8bdd\u56de\u5408\u90fd\u4ee5\u611f\u5174\u8da3\u7684\u89d2\u8272\u7ed3\u675f\uff0c\u4ece\u800c\u4f7f\u5927\u591a\u6570\u56de\u5408\u90fd\u843d\u5728\u89d2\u8272\u7684\u60c5\u666f\u8bb0\u5fc6\u4e2d\u3002\u7136\u540e\u5c06\u8fd9\u4e9b\u6bb5\u7528\u4e8e\u57fa\u4e8e\u811a\u672c\u7684\u6307\u4ee4\u6570\u636e\u751f\u6210\u3002<\/p>\n<h2>\u6307\u4ee4\u548c\u56de\u7b54\u751f\u6210<\/h2>\n<p>\u6211\u4eec\u4e3a\u7279\u5b9a\u89d2\u8272\u7684\u6307\u4ee4\u6570\u636e\u751f\u6210\u7684\u5019\u9009\u6570\u636e\u5305\u62ec\u4e09\u4e2a\u5143\u7d20\uff1a\uff081\uff09\u4e0e\u7ed9\u5b9a\u7247\u6bb5\uff08\u5373\u4e0a\u4e0b\u6587\uff09\u76f8\u5173\u7684\u95ee\u9898\uff0c\u8bb0\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">Q<\/span>\uff1b\uff082\uff09\u95ee\u9898\u5bf9\u5e94\u7684\u7b54\u6848\uff0c\u8bb0\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">A<\/span>\uff1b\uff083\uff09\u751f\u6210\u7684<span class=\"katex-eq\" data-katex-display=\"false\">QA<\/span>\u5bf9\u7684\u6267\u884c\u5ea6\u5f97\u5206\u4ee5\u53ca\u5f97\u5206\u4f9d\u636e\uff0c\u8868\u793a\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">C<\/span>\u3002\u8bbe<span class=\"katex-eq\" data-katex-display=\"false\">R<\/span>\u4e3a\u89d2\u8272\u7684\u4e2a\u6570\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">N_r<\/span>\u662f\u89d2\u8272<span class=\"katex-eq\" data-katex-display=\"false\">r<\/span>\u7684\u60c5\u8282\u4e2a\u6570\u3002\u4f7f\u7528LLM\u751f\u6210\u4e09\u5143\u7ec4\u4e3a\u6bcf\u4e2a\u89d2\u8272\u548c\u6bcf\u4e2a\u60c5\u8282<span class=\"katex-eq\" data-katex-display=\"false\">s<\/span>\u751f\u6210\u4e09\u5143\u7ec4<span class=\"katex-eq\" data-katex-display=\"false\">[Q;C;A]<\/span>\u3002\u5728\u521d\u6b65\u5b9e\u9a8c\u4e2d\uff0c\u53d1\u73b0\u5728\u6ca1\u6709\u7f6e\u4fe1\u5ea6\u8bc4\u5206\u7684\u60c5\u51b5\u4e0b\u751f\u6210QA\u5bf9\u7684\u4f20\u7edf\u65b9\u6cd5\u5bfc\u81f4\u4e86\u4f4e\u8d28\u91cf\u7684\u95ee\u9898\u3002\u5728\u57fa\u4e8e\u811a\u672c\u7684QA\u5bf9\u7684\u60c5\u51b5\u4e0b\uff0c\u95ee\u9898\u5f80\u5f80\u663e\u5f97\u4e0d\u5b8c\u6574\uff0c\u56e0\u4e3a\u5b83\u4eec\u5047\u8bbe\u4e86\u7ed9\u5b9a\u90e8\u5206\u7684\u5148\u9a8c\u77e5\u8bc6\u3002\u5bf9\u4e8e\u811a\u672c\u4e0d\u53ef\u77e5\u7684<span class=\"katex-eq\" data-katex-display=\"false\">[Q;A]<\/span>\uff0c\u7531\u4e8e\u7f3a\u4e4f\u4e0a\u4e0b\u6587\uff0c\u8fd9\u4e9b\u95ee\u9898\u7ecf\u5e38\u5305\u542b\u5e7b\u89c9\u95ee\u9898\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5728\u751f\u6210<span class=\"katex-eq\" data-katex-display=\"false\">[Q;A]<\/span>\u7684\u540c\u65f6\uff0c\u4e5f\u751f\u6210\u4e86\u8bc4\u4f30\u201c\u5b8c\u6574\u6027\u201d\u6216\u201c\u771f\u5b9e\u6027\u201d\u7684\u7f6e\u4fe1\u5ea6\u5206\u6570\u3002\u5206\u6570\u4ece\u5dee\u5230\u597d\u5206\u4e24\u4e2a\u7ea7\u522b\u3002\u63a5\u4e0b\u6765\uff0c\u5bf9\u4e8e\u57fa\u4e8e\u811a\u672c\u7684<span class=\"katex-eq\" data-katex-display=\"false\">[Q;C;A]<\/span>\u4e09\u5143\u7ec4\uff0c\u6bcf\u4e2a\u7247\u6bb5\u63d0\u793aGPT3.5\u751f\u62103\u4e2a\u5019\u9009\u56de\u7b54\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u89d2\u8272\u6700\u591a\u4ea7\u751f3*100=300\u4e2a\u5019\u9009\u7b54\u6848\uff0c\u4e14\u5e73\u5747\u5019\u9009\u56de\u7b54\u4e3a150\u4e2a\u3002\u5bf9\u4e8e\u811a\u672c\u65e0\u5173\u7684<span class=\"katex-eq\" data-katex-display=\"false\">[Q;C;A]<\/span>\u4e09\u5143\u7ec4\uff0c\u89d2\u8272\u63cf\u8ff0\u63d0\u793a\u8bcd\u5728\u4e00\u8f6e\u4e2d\u900910\u4e2a\u5019\u9009\u4eba\uff0c\u91cd\u590d20\u6b21\uff0c\u4e00\u5171200\u4e2a\u5019\u9009\u4eba\u3002\u5982\u679c\u57fa\u4e8e\u811a\u672c\u548c\u811a\u672c\u65e0\u5173\u7684\u7ec4\u5408\u6280\u672f\u5c11\u4e8e400\u4e2a\u6ca1\u7ee7\u7eed\u751f\u6210\u4e0e\u811a\u672c\u65e0\u5173\u7684\u95ee\u7b54\uff0c\u76f4\u5230\u8fbe\u5230\u9608\u503c\u3002<\/p>\n<h2>\u6570\u636e\u8fc7\u6ee4\u548c\u540e\u671f\u5904\u7406<\/h2>\n<p>\u4e3a\u4e86\u63d0\u9ad8\u6570\u636e\u8d28\u91cf\uff0c\u53ea\u4fdd\u7559\u7f6e\u4fe1\u5ea6\u6700\u9ad8\u7684\u5019\u9009\u6570\u636e\uff0c\u5254\u966430%\u7684\u6570\u636e\u3002\u5bf9\u4e8e\u591a\u6837\u6027\uff0c\u4f7f\u7528BM25\u76f8\u4f3c\u6027\u6765\u6d88\u9664\u7279\u5b9a\u89d2\u8272\u7684\u63d0\u95ee\uff0c\u5220\u9664\u4e86\u5269\u4e0b66%\u7684\u6570\u636e\u3002\u8003\u8651\u5230\u6211\u4eec\u7684\u53bb\u91cd\u4f7f\u5f97\u4e0d\u540c\u6837\u672c\u7684\u77e5\u8bc6\u65e0\u5173\uff0c\u4f20\u7edf\u7684\u8bad\u7ec3-\u6d4b\u8bd5\u5206\u5272\u662f\u4e0d\u53ef\u884c\u7684\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u4f7f\u7528\u4e0a\u8ff0\u6570\u636e\u8bad\u7ec3\uff0c\u5e76\u5c06\u8fc7\u6ee4\u6389\u7684\u6570\u636e\u89c6\u4e3a\u6d4b\u8bd5\u96c6\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u6709\u8d85\u8fc750\u4e2a\u8fc7\u6ee4\u95ee\u9898\u7684\u89d2\u8272\uff0c\u6839\u636eBM25\u9009\u62e950\u4e2a\u6700\u4e0d\u76f8\u4f3c\u7684\u95ee\u9898\uff0c\u5426\u5219\uff0c\u5c06\u4fdd\u7559\u6240\u6709\u8fc7\u6ee4\u6389\u7684\u95ee\u9898\u3002<\/p>\n<h1>\u9644\u5f55D<\/h1>\n<h2>\u89d2\u8272\u9009\u62e9<\/h2>\n<p>\u4ece<a href=\"https:\/\/github.com\/PedroUria\/NLP-Movie_Scripts\" title=\"NLP\u7535\u5f71\u6570\u636e\u5e93\" target=\"_blank\"  rel=\"nofollow\" >NLP\u7535\u5f71\u6570\u636e\u5e93<\/a>\u4e2d\u6536\u96c6\u4e86661\u4e2a\u5267\u672c\uff0c\u4eceSummScreen\u4e2d\u624b\u673a\u4e86255\u4e2a\u82f1\u6587\u5267\u672c\uff0c\u5e76\u624b\u52a8\u521b\u5efa\u4e86\u53e6\u591624\u4e2a\u4e2d\u6587\u5267\u672c\u3002\u4eba\u5de5\u9009\u62e9\u4e865\u4e2a\u77e5\u540d\u7684\u4e2d\u6587\u5267\u672c\uff0c\u4ece\u4e2d\u8bc6\u522b\u51fa5\u4e2a\u7ecf\u5178\u7684\u89d2\u8272\uff0c\u5982\u5b59\u609f\u7a7a\uff0c\u5e76\u4e14\u7cbe\u5fc3\u5236\u4f5c\u4eba\u7269\u63cf\u8ff0\u548c\u53e3\u5934\u7985\u3002\u5bf9\u4e8e\u82f1\u6587\u5267\u672c\uff0c\u9996\u76f8\u5c06\u5267\u672c\u540d\u5b57\u5448\u73b0\u7ed9GPT-4\uff0c\u5e76\u63d0\u793a\u6a21\u578b\u751f\u6210\u4e3b\u8981\u4efb\u52a1\u7684\u540d\u5b57\uff0c\u8bc4\u4f30\u4ed6\u4eec\u8bf4\u8bdd\u98ce\u683c\u7684\u72ec\u7279\u6027\uff0c\u5e76\u5229\u7528\u4e0b\u56fe\u4e2d\u5217\u51fa\u7684\u63d0\u793a\u8bcd\u6a21\u677f\u63d0\u4f9b\u8fd9\u4e9b\u8bc4\u4f30\u7684\u57fa\u672c\u903b\u8f91\u3002\u63a5\u4e0b\u6765\uff0c\u795e\u6076\u5316\u98ce\u683c\u72ec\u7279\u6027\u5f97\u5206\u6700\u9ad8\u7684\u4e24\u4e2a\u89d2\u8272\uff0c\u9a8c\u8bc1\u4ed6\u4eec\u7684\u5267\u672c\u6765\u6e90\u3001\u8bf4\u8bdd\u98ce\u683c\u7684\u51c6\u786e\u6027\u548c\u8bc4\u5206\u4f9d\u636e\u3002\u540c\u65f6\uff0c\u4ece\u811a\u672c\u5b50\u96c6\u4e2d\u4eba\u5de5\u63d0\u53d6\u4e3b\u8981\u89d2\u8272\uff0c\u8fd9\u4e9b\u89d2\u8272\u4e0eGPT-4\u7684\u504f\u597d\u8868\u73b0\u51fa\u5f3a\u70c8\u7684\u4e00\u81f4\u6027\u3002\u4e00\u5171\u4ea7\u751f107\u4e2a\u89d2\u8272\u3002\u968f\u540e\uff0c\u5c06\u53f0\u672c\u540d\u79f0\u548c\u89d2\u8272\u540d\u79f0\u4f5c\u4e3a\u8f93\u5165\uff0c\u4f7f\u7528\u4e0b\u56fe\u4e2d\u7684\u63d0\u793a\u6a21\u677f\u63d0\u793aGPT-4\u91cd\u65b0\u9009\u62e9\u89d2\u8272\u3002\u901a\u8fc7\u89c2\u5bdf\uff0c\u8fd9\u4e00\u6b65\u4e2d\u8f93\u5165\u7684\u6240\u6709\u89d2\u8272\u5728\u8bf4\u8bdd\u98ce\u683c\u72ec\u7279\u6027\u65b9\u9762\u83b7\u5f97\u5f02\u5e38\u9ad8\u7684\u5206\u6570\u3002\u6839\u636e\u89d2\u8272\u72ec\u7279\u6027\u6765\u9009\u62e9\u89d2\u8272\uff0c\u7136\u540e\u6392\u9664\u5c0f\u4e8e25\u6b21\u5bf9\u8bdd\u7684\u89d2\u8272\u3002\u6700\u7ec8\u4ea7\u751f100\u4e2a\u89d2\u8272\uff0c\u5305\u62ec\u4e94\u4e2a\u4e2d\u6587\u548c95\u4e2a\u82f1\u6587\u89d2\u8272\u3002\u6bcf\u4e2a\u89d2\u8272\u90fd\u6709\u4e0d\u540c\u7684\u4e2a\u6027\u7279\u5f81\u548c\u8bf4\u8bdd\u98ce\u683c\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg7.jpg\" alt=\"\" \/><\/p>\n<h2>\u6784\u5efa\u89d2\u8272\u7b80\u4ecb<\/h2>\n<p>\u6bcf\u4e2a\u89d2\u8272\u7684\u7b80\u4ecb\u7531\u4e09\u90e8\u5206\u7ec4\u6210\uff1a\uff08a\uff09\u63cf\u8ff0;(b)\u53e3\u5934\u7985\uff1b\uff08c\uff09\u4ece\u811a\u672c\u4e2d\u89e3\u6790\u7684\u7ed3\u6784\u5316\u5bf9\u8bdd\u3002\u5bf9\u4e8e\u7ec4\u4ef6\uff08a\uff09\uff0c\u4f7f\u7528GPT-4\u751f\u6210\u89d2\u8272\u63cf\u8ff0\uff0c\u63d0\u793a\u8bcd\u6a21\u677f\u5982\u4e0b\u56fe\u6240\u793a\u3002\u540c\u6837\uff0c\u5bf9\u4e8e\u7ec4\u4ef6\uff08b\uff09\uff0c\u53e3\u5934\u8bed\u662f\u4f7f\u7528\u4e0b\u56fe\u63d0\u4f9b\u7684\u6a21\u677f\u751f\u6210\u7684\u3002\u6240\u6709\u7684\u4eba\u7269\u63cf\u8ff0\u548c\u53e3\u5934\u7985\u968f\u540e\u7531\u4f5c\u8005\u9a8c\u8bc1\u3002\u7531\u4e8e\u811a\u672c\u7684\u7ed3\u6784\u5316\u683c\u5f0f\uff0c\u5bf9\u8bdd\u4f7f\u7528\u6b63\u5219\u8868\u8fbe\u5f0f\u548c\u57fa\u4e8e\u89c4\u5219\u7684\u65b9\u6cd5\u8fdb\u884c\u89e3\u6790\uff0c\u4e0d\u4ec5\u5305\u542b\u4e0d\u540c\u7684\u89d2\u8272\uff0c\u8fd8\u5305\u542b\u63d0\u4f9b\u80cc\u666f\u548c\u4e8b\u4ef6\u7b49\u4e0a\u4e0b\u6587\u7684\u53d9\u8ff0\u3002\u89d2\u8272\u6240\u91c7\u53d6\u7684\u884c\u52a8\u88ab\u6574\u5408\u5230\u4ed6\u4eec\u7684\u5bf9\u8bdd\u5185\u5bb9\u4e2d\uff0c\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u4e00\u8f6e\u5bf9\u8bdd\uff08round\uff09\u5305\u542b\u591a\u4e2a\u5bf9\u8bdd\u56de\u5408\uff08turns\uff09\u3002\u8fd9\u5c31\u4ea7\u751f\u4e86\u5bf9\u8bdd\u56de\u5408\u7684\u7b49\u7ea7\u5217\u8868\uff0c\u4ece\u5355\u4e2a\u56de\u5408\u5230\u5bf9\u8bdd\u8f6e\u6b21\uff0c\u518d\u5230\u52a8\u4f5c\u6216\u60c5\u8282\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/RoleLLM_fg8.jpg\" alt=\"\" \/><\/p>\n<h2>\u901a\u7528\u6307\u4ee4\u62bd\u53d6<\/h2>\n<p>Super-NaturalInstruct\u3001UltraChat\u548cAlpaca\u6307\u4ee4\u96c6\u4e2d\u4ece\u968f\u673a\u62bd\u53d61500\u4e2a\u82f1\u8bed\u6307\u4ee4\uff0c\u6bcf\u4e2a\u6307\u4ee4\u5305\u62ec\u8fc7\u4f1a\u6253\u5c0f\u4e8e100\u4e2a\u5355\u8bcd\u3002\u540c\u65f6\uff0c\u4eceCOIG\u548cBELLE\u4e2d\u968f\u673a\u62bd\u53d61479\u6761\u4e2d\u6587\u6307\u4ee4\u3002\u6240\u6709\u7684\u91c7\u6837\u6307\u4ee4\u90fd\u57fa\u4e8e\u4ed6\u4eec\u7684BM24\u76f8\u4f3c\u6027\u8fdb\u884c\u53bb\u91cd<\/p>\n<h2>\u751f\u6210\u539f\u59cbRoleBench<\/h2>\n<p>\u7ed9\u51fa\u901a\u7528\u6307\u4ee4\uff0c\u4f7f\u7528RoleGPT\u751f\u6210\u5305\u542b\u89d2\u8272\u8bf4\u8bdd\u98ce\u683c\u548c\u76f8\u5173\u77e5\u8bc6\u7684\u7b54\u6848\u3002\u5bf9\u4e8erole-specific\u7684\u6307\u4ee4\uff0c\u6240\u6709\u7684\u6307\u4ee4\u548c\u56de\u7b54\u90fd\u662f\u4f7f\u7528Context-Instruct\u751f\u6210\u7684\u3002\u5728\u901a\u7528\u6307\u4ee4\u7684\u60c5\u51b5\u4e0b\uff0c\u4e3a\u6bcf\u4e2a\u89d2\u8272-\u6307\u4ee4\u5bf9\u751f\u6210\u516d\u4e2a\u56de\u7b54\uff0c\u5176\u4e2d\u4e94\u4e2a\u4f5c\u4e3a\u57fa\u7840\u4e8b\u5b9e\u5019\u9009\uff0c\u5269\u4e0b\u7684\u4e00\u4e2a\u5ea7\u4f4d\u6027\u80fd\u6bd4\u8f83\u7684ROleGPT\u57fa\u7ebf\u3002\u5bf9\u4e8erole-specific\u7684\u6307\u4ee4\uff0c\u6bcf\u4e2a\u95ee\u9898\u53ea\u751f\u6210\u4e00\u4e2a\u7b54\u6848\u3002<\/p>\n<h2>\u6570\u636e\u6e05\u6d17<\/h2>\n<p>\u6e05\u6d17\u539f\u59cbRoleBench\uff0c\u57fa\u4e8e\u56db\u4e2a\u539f\u5219\u8fdb\u884c\u6e05\u6d17\uff1a<\/p>\n<ul>\n<li>\u56de\u590d\u5b8c\u6574\u6027\uff1a\u4e0d\u5b8c\u6574\u56de\u590d\u7684\u6837\u672c\uff0c\u6ca1\u6709\u4ee5\u53e5\u53f7\u6807\u70b9\u7ed3\u5c3e\uff0c\u5220\u9664<\/li>\n<li>AI\u8eab\u4efd\u9690\u85cf\uff1a\u6392\u9664\u56de\u7b54\u63ed\u793aAI\u8eab\u4efd\u7684\u6837\u672c\uff0c\u4f8b\u5982\uff1a\u4f5c\u4e3a\u8bed\u8a00\u6a21\u578b\u5f00\u5934<\/li>\n<li>\u89d2\u8272\u8eab\u4efd\u4fe1\u606f\u9690\u85cf\uff1a\u6392\u9664\u4ee5\u76ee\u6807\u89d2\u8272\u540d\u79f0\u5f00\u5934\u7684\u6837\u672c\uff0c\u4f8b\u5982\u4ee5\uff1a\u201cjigsaw\uff1a\u201d\u5f00\u5934<\/li>\n<li>\u975e\u62d2\u7edd\uff1a\u6392\u9664\u6a21\u578b\u62d2\u7edd\u56de\u7b54\u6216\u63d0\u4f9b\u4fe1\u606f\u7684\u6837\u672c<\/li>\n<\/ul>\n<h1>\u9644\u5f55G<\/h1>\n<p>\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u8bad\u7ec3-\u6d4b\u8bd5\u5206\u5272\u7b56\u7565\u4fa7\u91cd\u4e8e\u4e24\u4e2a\u7ef4\u5ea6\uff1a\uff081\uff09\u7528\u4e8e\u8bc4\u4f30\u901a\u7528\u6307\u4ee4\u7684Intruction-based\u5206\u5272\uff0c\u5728Role-general\u4e2d\u6bd4\u4f8b\u4e3a4:1\uff08\u8bad\u7ec3\uff1a\u6d4b\u8bd5=1200:300\u6307\u4ee4\uff09\u3002\u5728RoleBench-specific\u6d4b\u8bd5\u5206\u5272\u6bcf\u4e2a\u89d2\u8272\u6700\u591a50\u4e2a\u6837\u672c\u3002\uff082\uff09\u57fa\u4e8e\u89d2\u8272\u7684\u5206\u5272\u7528\u4e8e\u8bc4\u4f30\u89d2\u8272\u6cdb\u5316\uff0c\u4ec5\u9650\u82f1\u8bed\uff0c\u6bd4\u4f8b\u4e3a8:1\uff08\u8bad\u7ec3\uff1a\u6d4b\u8bd580:10 roles\uff09<\/p>\n","protected":false},"excerpt":{"rendered":"<p>MetaData \u53d1\u8868\u65f6\u95f4\uff1a2023.10.01 \u4f5c\u8005\uff1aZekun Moore Wang, Zhongyuan Peng, Hao &#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":[1],"tags":[],"class_list":["post-675","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/675","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=675"}],"version-history":[{"count":7,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/675\/revisions"}],"predecessor-version":[{"id":696,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/675\/revisions\/696"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=675"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=675"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=675"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}