{"id":1479,"date":"2025-07-02T21:13:00","date_gmt":"2025-07-02T13:13:00","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=1479"},"modified":"2025-09-17T21:14:28","modified_gmt":"2025-09-17T13:14:28","slug":"agi-eval-%e7%8b%ac%e5%ae%b6%e6%89%98%e7%ae%a1-rm-bench%ef%bc%9a%e7%bb%99%e5%a4%a7%e6%a8%a1%e5%9e%8b%e7%9a%84%e6%89%93%e5%88%86%e8%83%bd%e5%8a%9b%e6%89%93%e4%b8%aa%e5%88%86","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/07\/02\/agi-eval-%e7%8b%ac%e5%ae%b6%e6%89%98%e7%ae%a1-rm-bench%ef%bc%9a%e7%bb%99%e5%a4%a7%e6%a8%a1%e5%9e%8b%e7%9a%84%e6%89%93%e5%88%86%e8%83%bd%e5%8a%9b%e6%89%93%e4%b8%aa%e5%88%86\/","title":{"rendered":"AGI-Eval \u72ec\u5bb6\u6258\u7ba1 RM-Bench\uff1a\u7ed9\u5927\u6a21\u578b\u7684\u201c\u6253\u5206\u80fd\u529b\u201d\u6253\u4e2a\u5206"},"content":{"rendered":"\n

\u524d\u8a00\uff1a\u5f53\u6a21\u578b\u7528\u534e\u4e3d\u8f9e\u85fb\u5305\u88c5\u9519\u8bef\u7b54\u6848\u3001\u4e00\u672c\u6b63\u7ecf\u80e1\u8bf4\u516b\u9053\u65f6\uff0c\u73b0\u6709\u7684\u8bc4\u6d4b\u4f53\u7cfb\u5374\u6d51\u7136\u4e0d\u89c9\uff0c\u75af\u72c2\u6253Call\uff01\u6e05\u534e\u5927\u5b66\u3001\u590d\u65e6\u5927\u5b66\u3001\u9999\u6e2f\u79d1\u6280\u5927\u5b66\u8054\u5408\u53d1\u5e03\u57fa\u51c6\u6d4b\u8bd5 RM-BENCH\uff0c\u5bf9\u5927\u8bed\u8a00\u6a21\u578b\u7684”\u8ba4\u77e5\u654f\u9510\u5ea6”\u53d1\u8d77\u6311\u6218\uff0c\u9996\u6b21\u7cfb\u7edf\u6027\u5730\u6784\u5efa\u4e86\u9488\u5bf9\u5956\u52b1\u6a21\u578b\u7684\u8bc4\u6d4b\u57fa\u51c6\uff0c\u76f4\u51fb\u5176\u201c\u5f62\u5f0f\u5927\u4e8e\u5185\u5bb9\u201d\u7684\u8bc4\u4f30\u56f0\u5883\u3002AGI-Eval\u8bc4\u6d4b\u793e\u533a<\/strong> \u6b63\u5f0f\u4e0e\u8bba\u6587\u4f5c\u8005\u8fbe\u6210\u5408\u4f5c\uff0c\u72ec\u5bb6\u6258\u7ba1 RM-Bench<\/strong>\u8bc4\u6d4b\u96c6<\/strong>\uff0c\u53ef\u81f3\u793e\u533a\u67e5\u770b RM-Bench \u7684\u6240\u6709\u5b50\u96c6\uff01<\/p>\n\n\n\n

\u201c\u662f\u65f6\u5019\uff0c\u7ed9\u5927\u6a21\u578b\u7684\u2018\u6253\u5206\u80fd\u529b\u2019\u6253\u4e2a\u5206\u4e86\uff01\u201d<\/strong><\/p>\n\n\n\n

\u76ee\u5f55\uff1a<\/p>\n\n\n\n

01.RM-Bench \u57fa\u51c6\u7684\u8bde\u751f<\/p>\n\n\n\n

02.RM-Bench \u6784\u5efa<\/p>\n\n\n\n

    2.1\u804a\u5929<\/p>\n\n\n\n

    2.2\u4ee3\u7801\u4e0e\u6570\u5b66<\/p>\n\n\n\n

    2.3\u5b89\u5168<\/p>\n\n\n\n

03.\u5982\u4f55\u7834\u89e3\u201c\u5f62\u5f0f\u5927\u4e8e\u5185\u5bb9\u201d\uff1f<\/p>\n\n\n\n

04.\u8bc4\u4f30\u7ed3\u679c<\/p>\n\n\n\n

    4.1\u6574\u4f53\u6027\u80fd<\/p>\n\n\n\n

    4.2 DPO \u6a21\u578b\u4e0e\u5e8f\u5217\u5206\u7c7b\u5668<\/p>\n\n\n\n

05.\u4e0e\u7b56\u7565\u6a21\u578b\u7684\u76f8\u5173\u6027<\/p>\n\n\n\n

    5.1\u98ce\u683c\u63a7\u5236\u76f8\u5173\u6027<\/p>\n\n\n\n

    5.2\u4e0b\u6e38\u4efb\u52a1\u76f8\u5173\u6027<\/p>\n\n\n\n

06.\u5408\u4f5c\u9080\u7ea6<\/p>\n\n\n\n

\u8bba\u6587\u4f5c\u8005\u4e3a\u6765\u81ea\u4e8e Fudan University \u7684 Yantao Liu\u3001Yixin Cao\uff0cTsinghua University \u7684 Zijun Yao\u3001Lei Hou\u3001Juanzi Li \u4ee5\u53ca Hong Kong University of Science and Technology \u7684 Rui Min\u3002<\/p>\n\n\n\n

AGI-Eval \u5927\u6a21\u578b\u8bc4\u6d4b\u793e\u533a\u5df2\u63a5\u53d7 RM-Bench \u7814\u7a76\u56e2\u961f\u8bc4\u6d4b\u96c6\u6258\u7ba1\u59d4\u6258\uff0c\u8bc4\u6d4b\u7ed3\u679c\u5c06\u968f\u6a21\u578b\u53d1\u5e03\u81ea\u52a8\u66f4\u65b0\uff0c\u540c\u65f6 RM-Bench \u7684\u6240\u6709\u6570\u636e\u96c6\u90fd\u5df2\u6388\u6743\u6258\u7ba1\u81f3 AGI-Eval \u5e73\u53f0\uff0c\u6587\u672b\u53ef\u70b9\u51fb github \u94fe\u63a5\u67e5\u770b\u8be6\u7ec6\u5185\u5bb9\uff0c\u540c\u65f6\u70b9\u51fb\u201c\u9605\u8bfb\u539f\u6587<\/strong>\u201d\u4e5f\u53ef\u5728 AGI-Eval \u5e73\u53f0\u67e5\u770b\u6a21\u578b\u6392\u540d\u699c\u5355\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

01. RM-Bench \u57fa\u51c6\u7684\u8bde\u751f<\/strong><\/p>\n\n\n\n

ChatGPT\u3001Claude \u548c OpenAI o1\u8fd9\u6837\u7684\u5927\u8bed\u8a00\u6a21\u578b\u53d6\u5f97\u7684\u663e\u8457\u6210\u5c31\uff0c\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4f9d\u8d56\u4e8e\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u548c\u63a8\u7406\u7f29\u653e\u5b9a\u5f8b\u3002\u5956\u52b1\u6a21\u578b\u5728\u8fd9\u4e24\u79cd\u6280\u672f\u4e2d\u90fd\u8d77\u7740\u5173\u952e\u4f5c\u7528\u3002\u5728 RLHF \u4e2d\uff0c\u5956\u52b1\u6a21\u578b\u5145\u5f53\u4eba\u7c7b\u4ef7\u503c\u89c2\u7684\u201c\u4ee3\u8a00\u4eba\u201d\uff0c\u7ed9\u8bed\u8a00\u6a21\u578b\uff08\u7b56\u7565\u6a21\u578b\uff09\u7684\u8bad\u7ec3\u63d0\u4f9b\u53cd\u9988\uff0c\u6709\u52a9\u4e8e\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5bf9\u9f50\u8bed\u8a00\u6a21\u578b\uff1b\u5728\u63a8\u7406\u7f29\u653e\u5b9a\u5f8b\u4e2d\uff0c\u5956\u52b1\u6a21\u578b\u7528\u4e8e\u6839\u636e\u9884\u6d4b\u5956\u52b1\u4ece\u4e00\u7ec4\u5019\u9009\u54cd\u5e94\u4e2d\u9009\u62e9\u6700\u4f73\u7b54\u6848\u3002<\/p>\n\n\n\n

\u4e0e\u5bf9\u9f50\u8bed\u8a00\u6a21\u578b\uff08\u5373\u7b56\u7565\u6a21\u578b\uff09\u8bc4\u4f30\u7684\u5feb\u901f\u53d1\u5c55\u76f8\u6bd4\uff0c\u73b0\u6709\u5956\u52b1\u6a21\u578b\u7684\u57fa\u51c6\u6d4b\u8bd5\u4ecd\u672a\u5f97\u5230\u5145\u5206\u63a2\u7d22\u3002\u4ee5\u5f80\u7684\u65b9\u6cd5\u5e38\u4f7f\u7528\u66f4\u5f3a\u7684\u8bed\u8a00\u6a21\u578b\u751f\u6210\u66f4\u597d\u7684\u54cd\u5e94\uff0c\u4f7f\u7528\u8f83\u5f31\u7684\u8bed\u8a00\u6a21\u578b\u751f\u6210\u8f83\u5dee\u7684\u54cd\u5e94\uff0c\u4f46\u8fd9\u6837\u6839\u672c\u6ca1\u6cd5\u8bc4\u4f30\u6a21\u578b\u5bf9\u5185\u5bb9\u7ec6\u5fae\u53d8\u5316\u7684\u654f\u611f\u5ea6\uff0c\u4e5f\u6d4b\u4e0d\u51fa\u5bf9\u98ce\u683c\u504f\u5dee\u7684\u5904\u7406\u80fd\u529b\uff0c\u548c\u7b56\u7565\u6a21\u578b\u6027\u80fd\u7684\u76f8\u5173\u6027\u4e5f\u5f88\u4f4e\u3002<\/p>\n\n\n\n

\u5c31\u50cf\u8ba9\u4e00\u4e2a\u88c1\u5224\u53bb\u8bc4\u5224\u6bd4\u8d5b\uff0c\u53ef\u8fd9\u4e2a\u88c1\u5224\u5374\u5206\u4e0d\u6e05\u9009\u624b\u4eec\u5728\u4e00\u4e9b\u5173\u952e\u7ec6\u8282\u4e0a\u7684\u5dee\u5f02\uff0c\u8fd8\u5bb9\u6613\u88ab\u9009\u624b\u7684\u5916\u5728\u8868\u73b0\u5f71\u54cd\uff0c\u8fd9\u6837\u7684\u8bc4\u5224\u7ed3\u679c\u80af\u5b9a\u4e0d\u9760\u8c31\u3002\u5956\u52b1\u6a21\u578b\u4e0d\u662f”\u9009\u7f8e\u5927\u8d5b”\uff0c\u9700\u5efa\u7acb”\u5185\u5bb9\u88c1\u5224”\u65b0\u6807\u51c6\u3002<\/p>\n\n\n\n

\u5728\u4e0b\u8868\u4e2d\uff0c\u54cd\u5e94 1 \u548c\u54cd\u5e94 2 \u4ec5\u76f8\u5dee\u4e00\u4e2a\u5355\u8bcd\uff0c\u4f46\u8868\u8fbe\u7684\u542b\u4e49\u5b8c\u5168\u4e0d\u540c\uff0c\u8fd9\u5c31\u8981\u6c42\u5956\u52b1\u6a21\u578b\u5173\u6ce8\u5185\u5bb9\u8d28\u91cf\u3002\u54cd\u5e94 3 \u5728\u4e8b\u5b9e\u4e0a\u662f\u9519\u8bef\u7684\uff0c\u4f46\u6bd4\u54cd\u5e94 1 \u66f4\u957f\uff0c\u8fd9\u53ef\u80fd\u4f1a\u8bef\u5bfc\u5956\u52b1\u6a21\u578b\u4e3a\u54cd\u5e94 3 \u5206\u914d\u66f4\u9ad8\u7684\u5956\u52b1\u3002\u6a21\u578b\u9677\u5165\u201c\u6837\u5f0f\u504f\u89c1\u9677\u9631\u201d\uff1a\u957f\u7b54\u6848\u2248\u9ad8\u5206\uff0c\u4e8b\u5b9e\u9519\u8bef\u5168\u9760\u5305\u88c5<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u8868 1\uff1a\u5173\u4e8e\u859b\u5b9a\u8c14\u7684\u732b\u7684\u63d0\u793a\u7684\u4e09\u79cd\u4e0d\u540c\u54cd\u5e94\uff0c\u7531\u5956\u52b1\u6a21\u578b LxzGordon\/URM-LLaMa-3-8B \u9884\u6d4b\u5956\u52b1\u3002\u54cd\u5e94 1 \u63d0\u4f9b\u4e86\u6b63\u786e\u4fe1\u606f\uff0c\u800c\u54cd\u5e94 2 \u548c\u54cd\u5e94 3 \u5305\u542b\u4e8b\u5b9e\u9519\u8bef\u3002\u5956\u52b1\u6a21\u578b\u96be\u4ee5\u8fa8\u522b\u54cd\u5e94 1 \u548c\u54cd\u5e94 2 \u4e4b\u95f4\u7ec6\u5fae\u4f46\u5173\u952e\u7684\u5dee\u5f02\uff0c\u5e76\u4e14\u7531\u4e8e\u54cd\u5e94 3 \u66f4\u957f\uff0c\u503e\u5411\u4e8e\u66f4\u559c\u6b22\u54cd\u5e94 3\u3002<\/em><\/p>\n\n\n\n

\u4e3a\u6b64\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u7684 RM-Bench<\/strong> \u8fd9\u4e2a\u5168\u65b0\u7684\u57fa\u51c6\u6d4b\u8bd5\uff0c\u8fd9\u662f\u4e00\u79cd\u5168\u65b0\u7684\u57fa\u51c6\uff0c\u7528\u4e8e\u8bc4\u4f30\u5956\u52b1\u6a21\u578b\u533a\u5206\u5fae\u5999\u53d8\u5316\u548c\u62b5\u6297\u98ce\u683c\u504f\u5dee\u7684\u80fd\u529b<\/strong>\uff0c\u5bf9\u8bed\u8a00\u6a21\u578b\u7684”\u8ba4\u77e5\u654f\u9510\u5ea6”\u53d1\u8d77\u6311\u6218\uff01\u5b83\u4e3b\u8981\u4ece\u4e24\u4e2a\u5173\u952e\u65b9\u9762\u6765\u8bc4\u4f30\u5956\u52b1\u6a21\u578b\uff1a<\/p>\n\n\n\n

1.\u5bf9\u7ec6\u5fae\u5185\u5bb9\u5dee\u5f02\u7684\u654f\u611f\u5ea6\uff1a<\/strong>\u4e00\u4e2a\u53ef\u9760\u7684\u5956\u52b1\u6a21\u578b\u5e94\u80fd\u654f\u611f\u5730\u533a\u5206\u5fae\u5999\u53d8\u5316\uff0c\u5e76\u4e3a\u6b63\u786e\u54cd\u5e94\u5206\u914d\u66f4\u9ad8\u7684\u5956\u52b1\u3002<\/p>\n\n\n\n

2.\u5bf9\u98ce\u683c\u504f\u5dee\u7684\u9c81\u68d2\u6027\uff1a<\/strong>\u4e00\u4e2a\u5f3a\u5927\u7684\u5956\u52b1\u6a21\u578b\u5e94\u907f\u514d\u88ab\u98ce\u683c\u4e0e\u5185\u5bb9\u4e4b\u95f4\u7684\u865a\u5047\u76f8\u5173\u6027\u8bef\u5bfc\uff0c\u5e76\u4e14\u65e0\u8bba\u98ce\u683c\u5982\u4f55\uff0c\u90fd\u80fd\u59cb\u7ec8\u62d2\u7edd\u4e8b\u5b9e\u9519\u8bef\u7684\u54cd\u5e94\u3002\u4f7f\u7528\u98ce\u683c\u63a7\u5236\u7684\u63d0\u793a\u751f\u6210\u4e0d\u540c\u98ce\u683c\u7684\u54cd\u5e94\u53d8\u4f53\uff0c\u5305\u62ec\u7b80\u6d01\u3001\u8be6\u7ec6\u548c Markdown \u683c\u5f0f\u3002<\/p>\n\n\n\n

\u5c0f\u7ed3\uff1aRM-BENCH \u4e0e\u5bf9\u9f50\u8bed\u8a00\u6a21\u578b\uff08\u7b56\u7565\u6a21\u578b\uff09\u7684\u6027\u80fd\u9ad8\u5ea6\u76f8\u5173\uff0c\u80fd\u4e3a\u9009\u62e9\u5408\u9002\u7684\u5956\u52b1\u6a21\u578b\u6765\u6821\u51c6\u8bed\u8a00\u6a21\u578b\u63d0\u4f9b\u53ef\u9760\u53c2\u8003\u3002<\/p>\n\n\n\n

02.RM-Bench \u6784\u5efa<\/strong><\/p>\n\n\n\n

\u8fd9\u662f\u4e00\u4e2a\u7528\u4e8e\u8bc4\u4f30\u5956\u52b1\u6a21\u578b\u7684\u57fa\u51c6\uff0cRM-BENCH \u6db5\u76d6\u56db\u4e2a\u5173\u952e\u9886\u57df\uff0c\u5373\u804a\u5929\u3001\u4ee3\u7801\u3001\u6570\u5b66\u548c\u5b89\u5168\u3002\u8fd9\u4e9b\u9886\u57df\u6db5\u76d6\u4e86\u5e7f\u6cdb\u7684\u73b0\u5b9e\u573a\u666f\uff0c\u5305\u62ec\u5f00\u653e\u57df\u804a\u5929\u3001\u63a8\u7406\u4efb\u52a1\u548c\u5b89\u5168\u5173\u952e\u60c5\u51b5\u3002<\/p>\n\n\n\n

\u5bf9\u4e8e\u6bcf\u4e2a\u9886\u57df\uff0c\u6211\u4eec\u6784\u5efa\u4e00\u4e2a (x, yc, yr) \u5143\u7ec4\u7684\u6570\u636e\u96c6\uff0c\u5176\u4e2d x \u662f\u63d0\u793a\uff0cyc \u662f\u9009\u5b9a\u7684\u54cd\u5e94\uff0cyr \u662f\u88ab\u62d2\u7edd\u7684\u54cd\u5e94\u3002\u4e24\u4e2a\u54cd\u5e94\u5747\u7531\u76f8\u540c\u7684\u5f3a\u5927\u8bed\u8a00\u6a21\u578b\u751f\u6210\u3002\u6b64\u5916\uff0c\u6211\u4eec\u751f\u6210\u9009\u5b9a\u548c\u88ab\u62d2\u7edd\u54cd\u5e94\u7684\u98ce\u683c\u63a7\u5236\u53d8\u4f53\uff0c\u4ee5\u8bc4\u4f30\u4e0e\u98ce\u683c\u7279\u5f81\u76f8\u5173\u7684\u5956\u52b1\u6a21\u578b\u504f\u5dee\u3002\u54cd\u5e94\u7684\u6b63\u786e\u6027\u7531\u4eba\u5de5\u6807\u6ce8\u5458\u9a8c\u8bc1\uff0c\u4ee5\u786e\u4fdd\u6240\u6709\u9886\u57df\u7684\u6570\u636e\u8d28\u91cf\u3002<\/p>\n\n\n\n

\u4ee5\u4e0b\u5404\u8282\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u4e3a\u6bcf\u4e2a\u9886\u57df\u6536\u96c6\u63d0\u793a x\u3001\u751f\u6210\u9009\u5b9a\u548c\u88ab\u62d2\u7edd\u54cd\u5e94 yc \u548c yr \u4ee5\u5f62\u6210\u6d4b\u8bd5\u5143\u7ec4 (x, yc, yr) \u7684\u8fc7\u7a0b\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

2.1 \u804a\u5929<\/strong><\/p>\n\n\n\n

\u5728\u804a\u5929\u9886\u57df\uff0c\u5b83\u4f1a\u4ece AlpacaEval \u6536\u96c6\u63d0\u793a\uff0c\u7cbe\u5fc3\u7b5b\u9009\u51fa\u4e0e\u4e8b\u5b9e\u77e5\u8bc6\u76f8\u5173\u7684\u90e8\u5206\uff0c\u7136\u540e\u7528 gpt – 4o \u751f\u6210\u54cd\u5e94\uff0c\u5e76\u5229\u7528\u591a\u8f6e\u8d8a\u72f1\u6280\u672f\u7ed9\u62d2\u7edd\u54cd\u5e94\u6ce8\u5165\u9519\u8bef\uff0c\u7ecf\u8fc7\u4eba\u5de5\u4e25\u683c\u9a8c\u8bc1\u540e\uff0c\u4fdd\u7559\u6709\u6548\u6837\u672c\u8fdb\u884c\u6d4b\u8bd5\u3002<\/p>\n\n\n\n

2.2 \u4ee3\u7801\u4e0e\u6570\u5b66<\/strong><\/p>\n\n\n\n

\u5728\u4ee3\u7801\u548c\u6570\u5b66\u9886\u57df\uff0c\u5206\u522b\u4ece HumanEvalPack \u548c MATH \u57fa\u51c6\u83b7\u53d6\u63d0\u793a\uff0c\u501f\u52a9 gpt – 4o \u751f\u6210\u54cd\u5e94\uff0c\u5229\u7528\u5355\u5143\u6d4b\u8bd5\u548c\u4e8b\u5b9e\u7b54\u6848\u81ea\u52a8\u9a8c\u8bc1\u54cd\u5e94\u7684\u6b63\u786e\u6027\uff0c\u786e\u4fdd\u6d4b\u8bd5\u7684\u51c6\u786e\u6027\u548c\u5ba2\u89c2\u6027\u3002<\/p>\n\n\n\n

2.3 \u5b89\u5168<\/strong><\/p>\n\n\n\n

RM-BENCH \u7684\u5b89\u5168\u9886\u57df\u8bc4\u4f30\u5956\u52b1\u6a21\u578b\u533a\u5206\u5b89\u5168\u548c\u4e0d\u5b89\u5168\u54cd\u5e94\u7684\u80fd\u529b\u3002\u9274\u4e8e\u5b89\u5168\u6311\u6218\u7684\u72ec\u7279\u6027\u8d28\uff0c\u8be5\u9886\u57df\u5206\u4e3a\u4e24\u4e2a\u5b50\u57df\uff1a\u201c\u5e94\u54cd\u5e94\u5b89\u5168\u201d \u548c \u201c\u5e94\u62d2\u7edd\u5b89\u5168\u201d\u3002\u201c\u5e94\u54cd\u5e94\u5b89\u5168\u201d \u5305\u542b\u8868\u9762\u4e0a\u4ee4\u4eba\u62c5\u5fe7\u4f46\u5b9e\u9645\u4e0a\u65e0\u5bb3\u7684\u63d0\u793a\uff0c\u4f8b\u5982 \u201c\u5982\u4f55\u5728\u7535\u5b50\u6e38\u620f\u4e2d\u6740\u6b7b\u654c\u4eba\uff1f\u201d\uff1b \u201c\u5e94\u62d2\u7edd\u5b89\u5168\u201d \u5b50\u57df\u5305\u542b\u771f\u6b63\u6709\u5bb3\u7684\u63d0\u793a\uff0c\u4f8b\u5982 \u201c\u5982\u4f55\u5236\u9020\u70b8\u5f39\uff1f\u201d\u3002\u9488\u5bf9\u4e0d\u540c\u7c7b\u578b\u7684\u63d0\u793a\uff0c\u91c7\u7528\u4e0d\u540c\u7684\u65b9\u5f0f\u751f\u6210\u54cd\u5e94\u5e76\u8fdb\u884c\u8bc4\u4f30\u3002<\/p>\n\n\n\n

03.\u5982\u4f55\u7834\u89e3”\u5f62\u5f0f\u5927\u4e8e\u5185\u5bb9”\uff1f<\/strong><\/p>\n\n\n\n

\u5bf9\u8bed\u8a00\u6a21\u578b\u4e2d\u5f3a\u5316\u5b66\u4e60\u7684\u6279\u8bc4\u6307\u51fa\uff0c\u50cf\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09\u548c\u76f4\u63a5\u7b56\u7565\u4f18\u5316\uff08DPO\uff09\u8fd9\u6837\u7684\u7b97\u6cd5\u53ef\u80fd\u4f1a\u5f15\u5165 \u201c\u91cd\u5f62\u5f0f\u8f7b\u5185\u5bb9\u201d \u7684\u504f\u5dee\uff0c\u5bfc\u81f4\u6a21\u578b\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u826f\u597d\uff0c\u4f46\u5b9e\u9645\u4e0a\u5e76\u6ca1\u6709\u771f\u6b63\u89e3\u51b3\u4efb\u52a1\u3002\u4e3a\u4e86\u56de\u5e94\u8fd9\u4e9b\u62c5\u5fe7\uff0c\u7814\u7a76\u4eba\u5458\u5f15\u5165\u4e86\u6570\u636e\u96c6\u7684\u98ce\u683c\u63a7\u5236\u53d8\u4f53\uff0c\u201c\u5f3a\u5236\u201d\u5207\u6362\u4e09\u79cd\u8f93\u51fa\u6a21\u5f0f\uff0c\u4ee5\u63a2\u7a76\u5956\u52b1\u6a21\u578b\u5bf9\u54cd\u5e94\u98ce\u683c\u7684\u504f\u5dee\u3002<\/p>\n\n\n\n

\u9075\u5faa\u804a\u5929\u673a\u5668\u4eba\u7ade\u6280\u573a\uff08Chatbot Arena\uff09\u7684\u98ce\u683c\u63a7\u5236\u8bbe\u8ba1\uff0c\u8003\u8651\u4e24\u4e2a\u98ce\u683c\u7279\u5f81\uff1a\u957f\u5ea6\u548c Markdown \u683c\u5f0f\u3002\u6839\u636e\u8fd9\u4e9b\u7279\u5f81\uff0c\u54cd\u5e94\u88ab\u5206\u4e3a\u4e09\u79cd\u7c7b\u578b\uff0c\u5f3a\u5236\u6a21\u578b\u751f\u6210\u4e09\u79cd\u5f62\u6001\u7b54\u6848\uff0c\u7834\u89e3”\u5f62\u5f0f\u5927\u4e8e\u5185\u5bb9”\u9b54\u5492\uff1a<\/p>\n\n\n\n

1. y<\/em>\u2205\u2014\u2014\u7b80\u4ecb\u98ce\u683c\uff1a\u7b80\u6d01\u3001\u76f4\u63a5\u4e86\u5f53\u3002<\/p>\n\n\n\n

2. yL\u2014\u2014\u8be6\u7ec6\u98ce\u683c\u7eaf\u6587\u672c\u683c\u5f0f\uff1a\u5185\u5bb9\u4e30\u5bcc\u8be6\u7ec6\uff0c\u4f46\u65e0\u7ed3\u6784\u5316\u6392\u7248\u3002<\/p>\n\n\n\n

3. yL,M<\/em>\u2014\u2014\u8be6\u7ec6\u98ce\u683c Markdown \u683c\u5f0f\uff1a\u5185\u5bb9\u4e30\u5bcc\u4e14\u6392\u7248\u826f\u597d\uff0c\u6ce8\u91cd\u53ef\u8bfb\u6027\u548c\u5c42\u6b21\u611f\u3002<\/em><\/p>\n\n\n\n

\u5bf9\u4e8e\u6bcf\u4e2a\u63d0\u793ax<\/em>\uff0c\u6211\u4eec\u5728\u4e09\u4e2a\u98ce\u683c\u7ea7\u522b\u4e0a\u6bd4\u8f83\u9009\u62e9\u54cd\u5e94\u548c\u62d2\u7edd\u54cd\u5e94\uff1a\u6781\u7b80\u7684y<\/em>\u2205\u3001\u8be6\u7ec6\u7684yL<\/em>\u548c Markdown \u683c\u5f0f\u7684\u54cd\u5e94yL<\/em>,M<\/em>\u3002\u8fd9\u4f7f\u6211\u4eec\u80fd\u591f\u72ec\u7acb\u4e8e\u98ce\u683c\u5dee\u5f02\u8bc4\u4f30\u5956\u52b1\u6a21\u578b\u533a\u5206\u9009\u62e9\u54cd\u5e94\u548c\u62d2\u7edd\u54cd\u5e94\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u7cfb\u7edf\u5730\u8bc4\u4f30\u5956\u52b1\u6a21\u578b\u5e76\u5c3d\u91cf\u51cf\u5c11\u98ce\u683c\u7684\u5e72\u6270\uff0c\u6211\u4eec\u5c06\u7ed3\u679c\u6574\u7406\u6210\u4e00\u4e2a 3\u00d73 \u7684\u77e9\u9635\uff0c\u79f0\u4e3aStyle-Substance Eval Matrix\uff08\u98ce\u683c – \u5185\u5bb9\u8bc4\u4f30\u77e9\u9635\uff09\u3002<\/strong><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u8be5\u56fe\u5c55\u793a\u4e86 sfairXC\/FsfairX – LLaMA3 – RM – v0.1 \u5956\u52b1\u6a21\u578b\u5728\u804a\u5929\u9886\u57df\u7684\u8fd9\u4e2a\u77e9\u9635\u793a\u4f8b\u3002\u884c\u4ee3\u8868\u4e0d\u540c\u98ce\u683c\u7684\u9009\u62e9\u54cd\u5e94\uff0c\u5217\u4ee3\u8868\u4e0d\u540c\u98ce\u683c\u7684\u62d2\u7edd\u54cd\u5e94\u3002\u5bf9\u89d2\u5143\u7d20\u6bd4\u8f83\u76f8\u540c\u98ce\u683c\u7684\u54cd\u5e94\uff0c\u800c\u975e\u5bf9\u89d2\u5143\u7d20\u6bd4\u8f83\u4e0d\u540c\u8be6\u7ec6\u7a0b\u5ea6\u548c\u683c\u5f0f\u7684\u54cd\u5e94\u3002<\/em>

<\/p>\n\n\n\n

\u4ece\u8fd9\u4e2a\u77e9\u9635\u4e2d\uff0c\u6211\u4eec\u5f97\u51fa\u4e09\u4e2a\u51c6\u786e\u7387\u6307\u6807\uff1a<\/p>\n\n\n\n

    \n
  • \u7b80\u5355\u51c6\u786e\u7387\uff08Easy Accuracy\uff09\uff1a<\/strong>\u4e0b\u4e09\u89d2\u7684\u5e73\u5747\u503c\uff0c\u4ee3\u8868\u5956\u52b1\u6a21\u578b\u5728\u5b58\u5728\u98ce\u683c\u7ebf\u7d22\u65f6\u68c0\u6d4b\u5185\u5bb9\u5b9e\u8d28\u7684\u80fd\u529b\u3002<\/li>\n\n\n\n
  • \u666e\u901a\u51c6\u786e\u7387\uff08Normal Accuracy\uff09<\/strong>\uff1a\u5bf9\u89d2\u5143\u7d20\u7684\u5e73\u5747\u503c\uff0c\u53cd\u6620\u6a21\u578b\u5728\u4e24\u4e2a\u54cd\u5e94\u98ce\u683c\u76f8\u540c\u65f6\u8bc4\u4f30\u5185\u5bb9\u5b9e\u8d28\u7684\u80fd\u529b\u3002<\/li>\n\n\n\n
  • \u56f0\u96be\u51c6\u786e\u7387\uff08Hard Accuracy\uff09<\/strong>\uff1a\u4e0a\u4e09\u89d2\u7684\u5e73\u5747\u503c\uff0c\u8861\u91cf\u6a21\u578b\u5373\u4f7f\u5728\u62d2\u7edd\u54cd\u5e94\u98ce\u683c\u66f4\u6709\u5229\u7684\u60c5\u51b5\u4e0b\uff0c\u4ec5\u57fa\u4e8e\u5185\u5bb9\u5b9e\u8d28\u8bc6\u522b\u66f4\u597d\u54cd\u5e94\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n

    \u8fd9\u4e9b\u6307\u6807\u9488\u5bf9\u804a\u5929\u3001\u5b89\u5168\u3001\u4ee3\u7801\u548c\u6570\u5b66\u8fd9\u56db\u4e2a\u9886\u57df\u8fdb\u884c\u8ba1\u7b97\uff0c\u4ece\u800c\u5f97\u51fa\u7279\u5b9a\u9886\u57df\u7684\u6307\u6807\uff0c\u5982\u804a\u5929\u666e\u901a\u51c6\u786e\u7387\u6216\u5b89\u5168\u56f0\u96be\u51c6\u786e\u7387\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8ba1\u7b97\u6240\u6709\u9886\u57df\u7684\u5e73\u5747\u51c6\u786e\u7387\uff0c\u4e3a\u5956\u52b1\u6a21\u578b\u63d0\u4f9b\u4e00\u4e2a\u6574\u4f53\u6027\u80fd\u6307\u6807\u3002<\/p>\n\n\n\n

    04.\u8bc4\u4f30\u7ed3\u679c<\/strong><\/p>\n\n\n\n

    \u6211\u4eec\u5728 RM – BENCH \u4e0a\u5bf9\u5404\u79cd\u5956\u52b1\u6a21\u578b\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u4f30\uff0c\u8fd9\u4e9b\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u4ece 20 \u4ebf\u5230\u5927\u89c4\u6a21\u7684 3400 \u4ebf\u4e0d\u7b49\uff0c\u5b83\u4eec\u8981\u4e48\u88ab\u8bad\u7ec3\u4e3a\u5206\u7c7b\u5668\uff0c\u8981\u4e48\u5728\u6709\u53c2\u8003\u6a21\u578b\u65f6\u901a\u8fc7\u76f4\u63a5\u7b56\u7565\u4f18\u5316\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n

    4.1 \u6574\u4f53\u6027\u80fd<\/strong><\/p>\n\n\n\n

    \u6211\u4eec\u5c55\u793a\u4e86\u5956\u52b1\u6a21\u578b\u5728 RM – BENCH \u4e0a\u7684\u6574\u4f53\u6027\u80fd\uff0c\u7a81\u51fa\u8fdb\u5c55\u5e76\u627e\u51fa\u9700\u8981\u6539\u8fdb\u7684\u5730\u65b9\u3002\u4e0b\u8868\u5c55\u793a\u4e86 RM – BENCH \u4e0a\u524d 20 \u4e2a\u5956\u52b1\u6a21\u578b\u7684\u6027\u80fd\u3002\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u8868 3\uff1aRM-BENCH \u4e0a\u524d 20 \u4e2a\u5956\u52b1\u6a21\u578b\u3002\u804a\u5929\u3001\u6570\u5b66\u3001\u4ee3\u7801\u3001\u5b89\u5168\u5217\u5c55\u793a\u4e86\u6a21\u578b\u5728\u5404\u9886\u57df\u7684\u5e73\u5747\u51c6\u786e\u7387\u3002\u7b80\u5355\u3001\u666e\u901a\u3001\u56f0\u96be\u5217\u5c55\u793a\u4e86\u6a21\u578b\u5728\u6240\u6709\u9886\u57df\u4e0d\u540c\u96be\u5ea6\u7ea7\u522b\u7684\u51c6\u786e\u7387\u3002\u5e73\u5747\u5217\u5c55\u793a\u4e86\u6a21\u578b\u5728 RM-BENCH \u4e0a\u7684\u6574\u4f53\u5e73\u5747\u51c6\u786e\u7387\u3002<\/em><\/p>\n\n\n\n

    \u7814\u7a76\u4eba\u5458\u7528 RM-BENCH \u5bf9\u8fd1 40 \u79cd\u5956\u52b1\u6a21\u578b\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u4f30\uff0c\u53d1\u73b0\u4e86\u4e0d\u5c11\u6709\u610f\u601d\u7684\u7ed3\u679c\uff1a<\/p>\n\n\n\n

    1.RM – BENCH \u5177\u6709\u6311\u6218\u6027<\/strong>\uff1a<\/p>\n\n\n\n

    \u63d0\u9ad8\u5956\u52b1\u6a21\u578b\u7684\u6027\u80fd\u4ecd\u6709\u5f88\u5927\u7684\u8fdb\u6b65\u7a7a\u95f4\u3002\u6211\u4eec\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u5373\u4f7f\u662f\u6700\u5148\u8fdb\u7684\u6a21\u578b\uff0c\u5982 Skywork – Reward – Llama – 3.1 – 8B\uff0c\u5728 RM – BENCH \u4e0a\u7684\u5e73\u5747\u51c6\u786e\u7387\u4e5f\u4ec5\u4e3a 70.1%\uff0c\u56f0\u96be\u51c6\u786e\u7387\u4e3a 46.6% \u3002<\/p>\n\n\n\n

    Nemotron-340B-Reward\u8fd9\u6837\u7684\u5927\u578b\u5956\u52b1\u6a21\u578b\uff0c\u5728 RM-BENCH \u4e0a\u4e5f\u8868\u73b0\u4e0d\u4f73\uff0c\u5e73\u5747\u51c6\u786e\u7387\u4ec5\u4e3a 69.5%\u3002\u4e0e\u968f\u673a\u731c\u6d4b\u7684\u57fa\u7ebf\uff0850%\uff09\u76f8\u6bd4\uff0c\u7ed3\u679c\u8fdc\u4e0d\u80fd\u4ee4\u4eba\u6ee1\u610f\u3002<\/p>\n\n\n\n

    2.\u98ce\u683c\u504f\u5dee\u4e25\u91cd<\/strong>\uff1a<\/p>\n\n\n\n

    RM – BENCH \u4e0a\u7684\u56f0\u96be\u51c6\u786e\u7387\u660e\u663e\u4f4e\u4e8e\u666e\u901a\u51c6\u786e\u7387\uff0c\u5927\u591a\u6570\u5956\u52b1\u6a21\u578b\u5728 Hard \u6a21\u5f0f\u4e0b\uff0c\u51c6\u786e\u7387\u672a\u80fd\u8d85\u8fc7\u968f\u673a\u6c34\u5e73\uff0850%\uff09\u3002\u8fd9\u8868\u660e\u8bb8\u591a\u73b0\u6709\u7684\u5956\u52b1\u6a21\u578b\u66f4\u50cf\u662f\u98ce\u683c\u504f\u597d\u6a21\u578b\uff0c\u5728\u9884\u6d4b\u5956\u52b1\u65f6\uff0c\u5956\u52b1\u6a21\u578b\u5f88\u5bb9\u6613\u53d7\u5230\u54cd\u5e94\u98ce\u683c\u7684\u5f71\u54cd\uff0c\u504f\u79bb\u4e86\u54cd\u5e94\u7684\u5b9e\u8d28\u5185\u5bb9\uff0c\u597d\u4f3c\u201cAI\u7cbe\u795e\u5206\u88c2\u201d\u3002<\/p>\n\n\n\n

    \u6700\u5148\u8fdb\u7684\u5956\u52b1\u6a21\u578b\uff0c\u5982 Skyword-Reward\uff0c\u65e0\u6cd5\u62b5\u6297\u98ce\u683c\u504f\u5dee\uff0c\u51c6\u786e\u7387\u4ec5\u4e3a 46.6%\uff0c\u5728\u98ce\u683c\u5e72\u6270\u4e0b\u4f4e\u4e8e\u968f\u673a\u731c\u6d4b\u7684\u51c6\u786e\u7387\u3002\u51cf\u8f7b\u98ce\u683c\u504f\u5dee\u548c\u63d0\u9ad8\u5956\u52b1\u6a21\u578b\u7a33\u5065\u6027\u662f\u76ee\u524d\u7684\u8feb\u5207\u9700\u6c42\u3002<\/p>\n\n\n\n

    3.\u6570\u5b66\u548c\u4ee3\u7801\u9886\u57df\u51fa\u73b0\u5927\u5d29\u76d8<\/strong>\uff1a<\/p>\n\n\n\n

    \u6570\u5b66\u548c\u4ee3\u7801\u9886\u57df\u5bf9\u5956\u52b1\u6a21\u578b\u6784\u6210\u4e86\u6700\u5927\u7684\u6311\u6218\uff0c\u5373\u4f7f\u662f\u5e73\u5747\u51c6\u786e\u7387\u4e5f\u96be\u4ee5\u8d85\u8fc7\u968f\u673a\u6c34\u5e73\uff0850%\uff09\u3002\u5728\u56f0\u96be\u51c6\u786e\u7387\u65b9\u9762\uff0c\u5956\u52b1\u6a21\u578b\u7684\u8868\u73b0\u66f4\u5dee\u3002<\/p>\n\n\n\n

    \u6700\u5148\u8fdb\u7684 Skywork – Reward – Llama – 3.1 – 8B \u5728\u6570\u5b66\u548c\u4ee3\u7801\u9886\u57df\u7684\u56f0\u96be\u51c6\u786e\u7387\u5206\u522b\u4ec5\u4e3a 28.4% \u548c 30.7%\uff0c\u582a\u6bd4\u629b\u786c\u5e01\u51b3\u7b56\uff0c\u8fd9\u4e00\u6027\u80fd\u8fdc\u8fdc\u843d\u540e\u4e8e\u968f\u673a\u731c\u6d4b\u7684\u57fa\u7ebf\uff0850%\uff09\uff0c\u8868\u660e\u5f53\u524d\u7684\u5956\u52b1\u6a21\u578b\u5728\u8fd9\u4e9b\u9886\u57df\u53ef\u80fd\u4f1a\u4f7f\u7b56\u7565\u6a21\u578b\u504f\u79bb\u6b63\u786e\u65b9\u5411\u3002<\/p>\n\n\n\n

    4.2 DPO \u6a21\u578b\u4e0e\u5e8f\u5217\u5206\u7c7b\u5668<\/strong><\/p>\n\n\n\n

    \u6211\u4eec\u65e8\u5728\u6bd4\u8f83\u4e24\u79cd\u5e7f\u6cdb\u91c7\u7528\u7684\u5956\u52b1\u5efa\u6a21\u8303\u5f0f\uff0c\u5373\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u6a21\u578b\u548c\u5e8f\u5217\u5206\u7c7b\u5668\u3002DPO \u662f\u4e00\u79cd\u6d41\u884c\u7684\u65e0\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\uff0c\u4f7f\u7528\u504f\u597d\u6570\u636e\u96c6\uff0c\u901a\u8fc7\u7b56\u7565\u6a21\u578b\u81ea\u8eab\u7684\u9690\u5f0f\u5956\u52b1\u4fe1\u53f7\u76f4\u63a5\u4f18\u5316\u7b56\u7565\u6a21\u578b\u3002<\/p>\n\n\n\n

    \u7531\u4e8e DPO \u6a21\u578b\u548c\u5e8f\u5217\u5206\u7c7b\u5668\u5956\u52b1\u6a21\u578b\u90fd\u53ef\u4ee5\u5728\u76f8\u540c\u7684\u504f\u597d\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6211\u4eec\u8fdb\u884c\u4e86\u4e00\u9879\u5bf9\u6bd4\u7814\u7a76\uff0c\u4ee5\u8bc4\u4f30\u4f7f\u7528 DPO \u6a21\u578b\u4f5c\u4e3a\u5956\u52b1\u6a21\u578b\u7684\u6709\u6548\u6027\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u4f7f\u7528 tulu – v2.5 \u7cfb\u5217\u4e2d\u7684\u5e8f\u5217\u5206\u7c7b\u5668\u548c DPO \u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u5728 HH – RLHF\u3001StackExchange\u3001Chatbot Arena 2023\u548c Nectar\u7b49\u504f\u597d\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n

    \u6211\u4eec\u5728 RM – BENCH \u4e0a\u8bc4\u4f30\u8fd9\u4e9b\u5e8f\u5217\u5206\u7c7b\u5668\u3002\u5bf9\u4e8e\u5b83\u4eec\u5bf9\u5e94\u7684 DPO \u6a21\u578b\uff0c\u6211\u4eec\u5728\u6709\u548c\u6ca1\u6709\u53c2\u8003\u6a21\u578b tulu – 2 – 13b \u7684\u60c5\u51b5\u4e0b\uff0c\u8bc4\u4f30\u5b83\u4eec\u5728 RM – BENCH \u4e0a\u7684\u5e73\u5747\u51c6\u786e\u7387\u3002\u7ed3\u679c\u5982\u4e0b\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5728 RM-BENCH \u4e0a\uff0c\u4f7f\u7528\u4e0d\u540c\u504f\u597d\u6570\u636e\u96c6\u8bad\u7ec3\u7684 DPO \u6a21\u578b\u548c\u5e8f\u5217\u5206\u7c7b\u5668\u7684\u5e73\u5747\u51c6\u786e\u7387\u6bd4\u8f83\u3002\u53c2\u8003\u6a21\u578b\u662f tulu-2-13b<\/em><\/p>\n\n\n\n

    \u5982\u8be5\u8868\u6240\u793a\uff0c\u5728\u76f8\u540c\u504f\u597d\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u65f6\uff0cDPO \u6a21\u578b\u7684\u8868\u73b0\u4f18\u4e8e\u5e8f\u5217\u5206\u7c7b\u5668\u3002\u6211\u4eec\u5047\u8bbe\u8fd9\u79cd\u6539\u8fdb\u6e90\u4e8e\u53c2\u8003\u6a21\u578b\u7684\u5f71\u54cd\uff0cDPO \u6a21\u578b\u7684\u5956\u52b1\u4fe1\u53f7\u7531\u53c2\u8003\u6a21\u578b\u7684\u4fe1\u53f7\u8fdb\u884c\u7f29\u653e\u3002\u6570\u636e\u652f\u6301\u8fd9\u4e00\u5047\u8bbe\uff0c\u56e0\u4e3a\u5f53\u53c2\u8003\u6a21\u578b\u4e0d\u53ef\u7528\u65f6\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u6027\u80fd\u663e\u8457\u4e0b\u964d\uff0c\u8fd9\u8868\u660e\u53c2\u8003\u6a21\u578b\u8d77\u7740\u5173\u952e\u4f5c\u7528\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    RM-BENCH \u4e2d\u54cd\u5e94\u7684\u6b63\u786e\u6027\u548c\u5197\u957f\u6027\u5206\u6570\u7684\u6563\u70b9\u56fe<\/em><\/p>\n\n\n\n

    \u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u6a21\u578b\u5728\u5956\u52b1\u5efa\u6a21\u65b9\u9762\u663e\u793a\u51fa\u66f4\u5927\u7684\u6f5c\u529b\u3002<\/strong>\u4e0e\u5e8f\u5217\u5206\u7c7b\u5956\u52b1\u6a21\u578b\u76f8\u6bd4\uff0cDPO \u6a21\u578b\u5728 RM-BENCH \u4e0a\u8868\u73b0\u66f4\u597d\u3002DPO\u6a21\u578b\u5f02\u519b\u7a81\u8d77\uff1a\u5728\u5956\u52b1\u5efa\u6a21\u65b9\u9762\u6f5c\u529b\u5de8\u5927\uff0c\u5f88\u53ef\u80fd\u6210\u4e3a\u66f4\u4f18\u7684\u5956\u52b1\u6a21\u578b\u9009\u62e9\u3002<\/p>\n\n\n\n

    05.\u4e0e\u7b56\u7565\u6a21\u578b\u7684\u76f8\u5173\u6027<\/strong><\/p>\n\n\n\n

    RM-BENCH \u4e0e\u7b56\u7565\u6a21\u578b\u6027\u80fd\u5177\u6709\u5f88\u5f3a\u7684\u76f8\u5173\u6027\uff0c\u4f7f\u5176\u6210\u4e3a\u9009\u62e9\u5956\u52b1\u6a21\u578b\u8fdb\u884c\u8bed\u8a00\u6a21\u578b\u5bf9\u9f50\u7684\u53ef\u9760\u53c2\u8003\u3002\u5956\u52b1\u6a21\u578b\u7684\u4e3b\u8981\u76ee\u6807\u662f\u63d0\u9ad8\u7b56\u7565\u6a21\u578b\u7684\u6027\u80fd\u3002\u56e0\u6b64\uff0c\u4e00\u4e2a\u597d\u7684\u5956\u52b1\u6a21\u578b\u57fa\u51c6\u5e94\u8be5\u4e0e\u7b56\u7565\u6a21\u578b\u7684\u6027\u80fd\u5448\u6b63\u76f8\u5173\u3002<\/p>\n\n\n\n

    \u4f7f\u7528 Tulu – v2.5 \u7cfb\u5217\u4e2d\u7684\u5956\u52b1\u6a21\u578b\u53ca\u5176\u76f8\u5e94\u7684\u7b56\u7565\u6a21\u578b\u8fdb\u884c\u5b9e\u9a8c\u3002\u8fd9\u56db\u4e2a\u5956\u52b1\u6a21\u578b\u5728\u4e0d\u540c\u7684\u504f\u597d\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u5305\u62ec HH – RLHF\u3001StackExchange\u3001Chatbot Arena 2023\u548c Nectar\u3002\u6240\u6709\u6570\u636e\u96c6\u90fd\u91c7\u6837\u5230 60k \u4e2a\u793a\u4f8b\uff0c\u4ee5\u786e\u4fdd\u8bad\u7ec3\u6570\u636e\u5927\u5c0f\u5177\u6709\u53ef\u6bd4\u6027\u3002\u7b56\u7565\u6a21\u578b\u4f7f\u7528\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f7f\u7528\u76f8\u540c\u7684\u8bad\u7ec3\u6570\u636e\u548c\u8d85\u53c2\u6570\u3002<\/p>\n\n\n\n

    5.1 \u98ce\u683c\u63a7\u5236\u76f8\u5173\u6027<\/strong><\/p>\n\n\n\n

    \u9996\u5148\uff0c\u6211\u4eec\u7814\u7a76 RM – BENCH \u4e0a\u5956\u52b1\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u7b56\u7565\u6a21\u578b\u5728\u98ce\u683c\u63a7\u5236\u8bc4\u4f30\u4e2d\u7684\u6027\u80fd\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u63a2\u7a76\u5728 RM – BENCH \u4e0a\u56f0\u96be\u51c6\u786e\u7387\u8868\u73b0\u826f\u597d\u7684\u5956\u52b1\u6a21\u578b\uff0c\u662f\u5426\u80fd\u5728\u98ce\u683c\u63a7\u5236\u7684\u8bbe\u7f6e\u4e2d\u4f7f\u7b56\u7565\u6a21\u578b\u8868\u73b0\u66f4\u597d\u3002<\/p>\n\n\n\n

    \u4e3a\u4e86\u6d4b\u8bd5\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u4f7f\u7528 Arena – Hard – Auto\u4f5c\u4e3a\u7b56\u7565\u6a21\u578b\u7684\u98ce\u683c\u63a7\u5236\u8bc4\u4f30\u57fa\u51c6\u3002\u8fd9\u4e2a\u57fa\u51c6\u4e0e RM – BENCH \u7c7b\u4f3c\uff0c\u5c06\u957f\u5ea6\u548c Markdown \u683c\u5f0f\u4f5c\u4e3a\u98ce\u683c\u7279\u5f81\u3002\u6211\u4eec\u5c06\u7b56\u7565\u6a21\u578b\u7684\u98ce\u683c\u63a7\u5236\u5206\u6570\u5b9a\u4e49\u4e3a\u5728\u98ce\u683c\u63a7\u5236\u8bc4\u4f30\u4e2d\u76f8\u5bf9\u4e8e\u65e0\u98ce\u683c\u63a7\u5236\u8bc4\u4f30\u7684\u6027\u80fd\u76f8\u5bf9\u4e0b\u964d\u3002\u8f83\u9ad8\u7684\u98ce\u683c\u63a7\u5236\u5206\u6570\u8868\u660e\u7b56\u7565\u6a21\u578b\u5bf9\u98ce\u683c\u7279\u5f81\u7684\u504f\u5dee\u8f83\u5c0f\u3002<\/p>\n\n\n\n

    \u5bf9\u4e8e\u5956\u52b1\u6a21\u578b\uff0c\u6211\u4eec\u4f7f\u7528 RM – BENCH \u804a\u5929\u9886\u57df\u7684\u56f0\u96be\u51c6\u786e\u7387\u4f5c\u4e3a\u8bc4\u4f30\u6307\u6807\uff0c\u56e0\u4e3a\u5b83\u76f4\u63a5\u8861\u91cf\u4e86\u6a21\u578b\u4f18\u5148\u8003\u8651\u5185\u5bb9\u5b9e\u8d28\u800c\u975e\u98ce\u683c\u7684\u80fd\u529b\uff0c\u8fd9\u5bf9\u4e8e\u51cf\u5c11\u98ce\u683c\u504f\u5dee\u81f3\u5173\u91cd\u8981\u3002<\/p>\n\n\n\n

    \u5982\u56fe\u6240\u793a\uff0cRM – BENCH \u4e0a\u56f0\u96be\u51c6\u786e\u7387\u7684\u63d0\u9ad8\u4e0e\u7b56\u7565\u6a21\u578b\u98ce\u683c\u63a7\u5236\u5206\u6570\u7684\u663e\u8457\u63d0\u9ad8\u76f8\u5173\u3002\u8fd9\u8868\u660e\uff0c\u5f3a\u8c03\u5185\u5bb9\u5b9e\u8d28\u800c\u975e\u98ce\u683c\u7684\u5956\u52b1\u6a21\u578b\u4f1a\u4f7f\u7b56\u7565\u6a21\u578b\u7684\u98ce\u683c\u504f\u5dee\u964d\u4f4e\u3002<\/p>\n\n\n\n

    \u7b56\u7565\u6a21\u578b\u98ce\u683c\u504f\u5dee\u5206\u6570\u4e0e RM-BENCH \u804a\u5929\u9886\u57df\u5956\u52b1\u6a21\u578b\u56f0\u96be\u51c6\u786e\u7387\u7684\u6298\u7ebf\u56fe<\/em><\/p>\n\n\n\n

    5.2 \u4e0b\u6e38\u4efb\u52a1\u76f8\u5173\u6027<\/strong><\/p>\n\n\n\n

    \u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u7814\u7a76 RM – BENCH \u4e0a\u5956\u52b1\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u7b56\u7565\u6a21\u578b\u5728\u5404\u79cd\u4e0b\u6e38\u4efb\u52a1\uff08\u5305\u62ec\u6570\u5b66\u3001\u4ee3\u7801\u548c\u5b89\u5168\uff09\u4e2d\u7684\u6027\u80fd\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u3002\u6570\u5b66\u4efb\u52a1\u4f7f\u7528 GSM8k\u548c Big Bench Hard\u8fdb\u884c\u8bc4\u4f30\u3002\u4ee3\u7801\u4efb\u52a1\u4f7f\u7528 HumanEval+\u548c MBPP+\u8fdb\u884c\u8bc4\u4f30\u3002\u5b89\u5168\u4efb\u52a1\u5728 ToxiGen\u548c XSTest\u4e0a\u8fdb\u884c\u8bc4\u4f30\u3002<\/p>\n\n\n\n

    \u5bf9\u4e8e\u5956\u52b1\u6a21\u578b\uff0c\u6211\u4eec\u6839\u636e\u4efb\u52a1\u7684\u6027\u8d28\u9009\u62e9\u6307\u6807\u3002\u5bf9\u4e8e\u6570\u5b66\u548c\u5b89\u5168\u4efb\u52a1\uff0c\u6211\u4eec\u4f7f\u7528\u56f0\u96be\u51c6\u786e\u7387\uff0c\u56e0\u4e3a\u6b63\u786e\u6027\u81f3\u5173\u91cd\u8981\uff0c\u5e76\u4e14\u8fd9\u4e9b\u4efb\u52a1\u901a\u5e38\u6d89\u53ca\u4e0d\u540c\u7684\u6587\u672c\u98ce\u683c\uff0c\u9700\u8981\u533a\u5206\u5185\u5bb9\u5b9e\u8d28\u548c\u98ce\u683c\u3002<\/p>\n\n\n\n

    \u5bf9\u4e8e\u4ee3\u7801\u4efb\u52a1\uff0c\u8bed\u8a00\u6a21\u578b\u503e\u5411\u4e8e\u751f\u6210\u98ce\u683c\u4e00\u81f4\u7684\u6587\u672c\uff08\u7279\u522b\u662f Markdown \u683c\u5f0f\uff09\uff0c\u56e0\u4e3a\u6765\u81ea GitHub \u548c StackOverflow \u7b49\u6765\u6e90\u7684\u8bad\u7ec3\u6570\u636e\u5927\u591a\u662f Markdown \u683c\u5f0f\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4f7f\u7528\u666e\u901a\u51c6\u786e\u7387\uff0c\u4ee5\u4fbf\u66f4\u597d\u5730\u4e0e\u4ee3\u7801\u98ce\u683c\u7684\u5185\u5728\u4e00\u81f4\u6027\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n\n\n\n

    \u4e3a\u4e86\u8fdb\u4e00\u6b65\u5c55\u793a\u76f8\u5173\u6027\uff0c\u6211\u4eec\u9996\u5148\u5c06\u7b56\u7565\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u57fa\u7840 SFT \u6a21\u578b tulu – 2 – 13b\u8fdb\u884c\u6bd4\u8f83\uff0c\u5bf9\u5176\u8fdb\u884c\u5f52\u4e00\u5316\u3002RM – BENCH \u4e0a\u5956\u52b1\u6a21\u578b\u7684\u5206\u6570\u4f7f\u7528\u5176\u6027\u80fd\u7684\u5747\u503c\u548c\u6807\u51c6\u5dee\u8fdb\u884c\u6807\u51c6\u5316\u3002RM – BENCH \u4e0a\u5956\u52b1\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u4e0b\u6e38\u4efb\u52a1\u4e2d\u7b56\u7565\u6a21\u578b\u6027\u80fd\u7684\u5bf9\u6bd4\u56fe\u5982\u4e0b\uff1a<\/p>\n\n\n\n

    RM-BENCH \u4e0a\u5956\u52b1\u6a21\u578b\u6027\u80fd\u4e0e\u4e0b\u6e38\u4efb\u52a1\u4e2d\u7b56\u7565\u6a21\u578b\u6027\u80fd\u7684\u76f8\u5173\u6027<\/em><\/p>\n\n\n\n

    \u76ae\u5c14\u900a\u76f8\u5173\u7cfb\u6570\u4e3a 0.55\uff08p<\/em>=0.07\uff09\uff0c\u8868\u660e\u5b58\u5728\u4e2d\u7b49\u7a0b\u5ea6\u7684\u6b63\u76f8\u5173\uff0c\u4e14\u63a5\u8fd1\u663e\u8457\u6c34\u5e73\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cRewardBench\u62a5\u544a\u7684\u76ae\u5c14\u900a\u76f8\u5173\u7cfb\u6570\u4e3ar<\/em>=0.21\uff08p<\/em>=0.51\uff09\u3002\u8fd9\u7a81\u51fa\u8868\u660e\uff0cRM – BENCH \u5728\u6210\u4e3a\u4e0e\u5956\u52b1\u6a21\u578b\u8bc4\u4f30\u76f8\u5173\u6027\u66f4\u597d\u7684\u57fa\u51c6\u65b9\u9762\u5411\u524d\u8fc8\u8fdb\u4e86\u4e00\u6b65\u3002<\/p>\n\n\n\n

    \u201c\u4e0d\u662f\u8981\u6dd8\u6c70\u73b0\u6709\u6a21\u578b\uff0c\u800c\u662f\u5efa\u7acb\u65b0\u7684\u8bc4\u4f30\u8303\u5f0f\u2014\u2014\u5c31\u50cf\u4ece\u2018\u6bd4\u8c01\u8dd1\u5f97\u5feb\u2019\u8f6c\u5411\u2018\u6d4b\u8c01\u4e0d\u6454\u8de4\u2019\u201d\u3002<\/strong>\u5e0c\u671b RM-BENCH \u80fd\u591f\u9f13\u52b1\u793e\u533a\u6279\u5224\u6027\u5730\u5ba1\u89c6\u5956\u52b1\u6a21\u578b\u57fa\u51c6\u7684\u8bbe\u8ba1\uff0c\u5e76\u6fc0\u53d1\u672a\u6765\u5f00\u53d1\u66f4\u51c6\u786e\u3001\u66f4\u7cfb\u7edf\u7684\u8bc4\u4f30\u3002<\/p>\n\n\n\n

    <\/p>\n","protected":false},"excerpt":{"rendered":"

    \u524d\u8a00\uff1a\u5f53\u6a21\u578b\u7528\u534e\u4e3d\u8f9e\u85fb\u5305\u88c5\u9519\u8bef\u7b54\u6848\u3001\u4e00\u672c\u6b63\u7ecf\u80e1\u8bf4\u516b\u9053\u65f6\uff0c\u73b0\u6709\u7684\u8bc4\u6d4b\u4f53\u7cfb\u5374\u6d51\u7136\u4e0d\u89c9\uff0c\u75af\u72c2\u6253Call\uff01\u6e05\u534e\u5927\u5b66\u3001\u590d […]<\/p>\n","protected":false},"author":1,"featured_media":1480,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[],"class_list":["post-1479","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-8"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1479","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=1479"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1479\/revisions"}],"predecessor-version":[{"id":1488,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1479\/revisions\/1488"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/1480"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=1479"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=1479"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=1479"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}