{"id":37321,"date":"2025-01-26T08:00:15","date_gmt":"2025-01-26T00:00:15","guid":{"rendered":"https:\/\/17aitech.com\/?p=37321"},"modified":"2025-01-26T08:00:15","modified_gmt":"2025-01-26T00:00:15","slug":"3b%e6%a8%a1%e5%9e%8b%e9%95%bf%e6%80%9d%e8%80%83%e5%90%8e%e5%87%bb%e8%b4%a570b%ef%bc%81huggingface%e9%80%86%e5%90%91%e5%87%bao1%e8%83%8c%e5%90%8e%e6%8a%80%e6%9c%af%e7%bb%86%e8%8a%82%e5%b9%b6%e5%bc%80","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=37321","title":{"rendered":"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-17-7\" target=\"_blank\">3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"30\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>\u5982\u679c\u7ed9\u5c0f\u6a21\u578b\u66f4\u957f\u7684\u601d\u8003\u65f6\u95f4\uff0c\u5b83\u4eec\u6027\u80fd\u53ef\u4ee5\u8d85\u8d8a\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u3002<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<p>\u6700\u8fd1\u4e00\u6bb5\u65f6\u95f4\uff0c\u4e1a\u5185\u5bf9\u5c0f\u6a21\u578b\u7684\u7814\u7a76\u70ed\u60c5\u7a7a\u524d\u5730\u9ad8\u6da8\uff0c\u901a\u8fc7\u4e00\u4e9b\u300c\u5b9e\u7528\u6280\u5de7\u300d\u8ba9\u5b83\u4eec\u5728\u6027\u80fd\u4e0a\u8d85\u8d8a\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u3002<\/p>\n<p>\u53ef\u4ee5\u8bf4\uff0c\u5c06\u76ee\u5149\u653e\u5230\u63d0\u5347\u8f83\u5c0f\u6a21\u578b\u7684\u6027\u80fd\u4e0a\u6765\u6709\u5176\u5fc5\u7136\u6027\u3002\u5bf9\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u800c\u8a00\uff0c\u8bad\u7ec3\u65f6\u8ba1\u7b97\uff08train-time compute\uff09\u7684\u6269\u5c55\u4e3b\u5bfc\u4e86\u5b83\u4eec\u7684\u53d1\u5c55\u3002\u5c3d\u7ba1\u8fd9\u79cd\u6a21\u5f0f\u5df2\u88ab\u8bc1\u660e\u975e\u5e38\u6709\u6548\uff0c\u4f46\u8d8a\u6765\u8d8a\u5927\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u6240\u9700\u7684\u8d44\u6e90\u5374\u53d8\u5f97\u5f02\u5e38\u6602\u8d35\uff0c\u6570\u5341\u4ebf\u7f8e\u5143\u7684\u96c6\u7fa4\u5df2\u7ecf\u51fa\u73b0\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u8fd9\u4e00\u8d8b\u52bf\u5f15\u53d1\u4e86\u4eba\u4eec\u5bf9\u53e6\u5916\u4e00\u79cd\u4e92\u8865\u65b9\u6cd5\u7684\u6781\u5927\u5174\u8da3\uff0c\u5373\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u6269\u5c55\uff08test-time compute scaling\uff09\u3002\u6d4b\u8bd5\u65f6\u65b9\u6cd5\u4e0d\u4f9d\u8d56\u4e8e\u8d8a\u6765\u8d8a\u5927\u7684\u9884\u8bad\u7ec3\u9884\u7b97\uff0c\u800c\u662f\u4f7f\u7528\u52a8\u6001\u63a8\u7406\u7b56\u7565\uff0c\u8ba9\u6a21\u578b\u5728\u66f4\u96be\u7684\u95ee\u9898\u4e0a\u300c\u601d\u8003\u66f4\u957f\u65f6\u95f4\u300d\u3002\u4e00\u4e2a\u7a81\u51fa\u7684\u4f8b\u5b50\u662f OpenAI \u7684 o1 \u6a21\u578b\uff0c\u968f\u7740\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u91cf\u7684\u589e\u52a0\uff0c\u5b83\u5728\u56f0\u96be\u6570\u5b66\u95ee\u9898\u4e0a\u8868\u73b0\u51fa\u6301\u7eed\u7684\u8fdb\u6b65\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8d824c26223cb12949e12f33a675e983.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8d824c26223cb12949e12f33a675e983.png\"><\/a><\/p>\n<p>\u867d\u7136\u6211\u4eec\u4e0d\u6e05\u695a o1 \u662f\u5982\u4f55\u8bad\u7ec3\u7684\uff0c\u4f46 DeepMind \u6700\u8fd1\u7684\u7814\u7a76\u8868\u660e\uff0c\u53ef\u4ee5\u901a\u8fc7\u8fed\u4ee3\u81ea\u6211\u6539\u8fdb\u6216\u4f7f\u7528\u5956\u52b1\u6a21\u578b\u5728\u89e3\u51b3\u65b9\u6848\u7a7a\u95f4\u4e0a\u8fdb\u884c\u641c\u7d22\u7b49\u7b56\u7565\u6765\u5b9e\u73b0\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u7684\u6700\u4f73\u6269\u5c55\u3002\u901a\u8fc7\u81ea\u9002\u5e94\u5730\u6309 prompt \u5206\u914d\u6d4b\u8bd5\u65f6\u8ba1\u7b97\uff0c\u8f83\u5c0f\u7684\u6a21\u578b\u53ef\u4ee5\u4e0e\u8f83\u5927\u3001\u8d44\u6e90\u5bc6\u96c6\u578b\u6a21\u578b\u76f8\u5ab2\u7f8e\uff0c\u6709\u65f6\u751a\u81f3\u8d85\u8d8a\u5b83\u4eec\u3002\u5f53\u5185\u5b58\u53d7\u9650\u4e14\u53ef\u7528\u786c\u4ef6\u4e0d\u8db3\u4ee5\u8fd0\u884c\u8f83\u5927\u6a21\u578b\u65f6\uff0c\u6269\u5c55\u65f6\u95f4\u65f6\u8ba1\u7b97\u5c24\u5176\u6709\u5229\u3002\u7136\u800c\u8fd9\u79cd\u6709\u524d\u9014\u7684\u65b9\u6cd5\u662f\u7528\u95ed\u6e90\u6a21\u578b\u6f14\u793a\u7684\uff0c\u6ca1\u6709\u53d1\u5e03\u4efb\u4f55\u5b9e\u73b0\u7ec6\u8282\u6216\u4ee3\u7801\u3002<\/p>\n<p>DeepMind \u8bba\u6587\uff1ahttps:\/\/arxiv.org\/pdf\/2408.03314<\/p>\n<p>\u5728\u8fc7\u53bb\u51e0\u4e2a\u6708\u91cc\uff0cHuggingFace \u4e00\u76f4\u5728\u6df1\u5165\u7814\u7a76\uff0c\u8bd5\u56fe\u5bf9\u8fd9\u4e9b\u7ed3\u679c\u8fdb\u884c\u9006\u5411\u5de5\u7a0b\u5e76\u590d\u73b0\u3002\u4ed6\u4eec\u5728\u8fd9\u7bc7\u535a\u6587\u5c06\u4ecb\u7ecd\uff1a<\/p>\n<ul>\n<li>\n<section><strong>\u8ba1\u7b97\u6700\u4f18\u6269\u5c55\uff08compute-optimal scaling\uff09\uff1a<\/strong>\u901a\u8fc7\u5b9e\u73b0 DeepMind \u7684\u6280\u5de7\u6765\u63d0\u5347\u6d4b\u8bd5\u65f6\u5f00\u653e\u6a21\u578b\u7684\u6570\u5b66\u80fd\u529b\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u591a\u6837\u6027\u9a8c\u8bc1\u5668\u6811\u641c\u7d22 (DVTS)\uff1a<\/strong>\u5b83\u662f\u4e3a\u9a8c\u8bc1\u5668\u5f15\u5bfc\u6811\u641c\u7d22\u6280\u672f\u5f00\u53d1\u7684\u6269\u5c55\u3002\u8fd9\u79cd\u7b80\u5355\u9ad8\u6548\u7684\u65b9\u6cd5\u63d0\u9ad8\u4e86\u591a\u6837\u6027\u5e76\u63d0\u4f9b\u4e86\u66f4\u597d\u7684\u6027\u80fd\uff0c\u7279\u522b\u662f\u5728\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u9884\u7b97\u8f83\u5927\u7684\u60c5\u51b5\u4e0b\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u641c\u7d22\u548c\u5b66\u4e60\uff1a<\/strong>\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u5de5\u5177\u5305\uff0c\u7528\u4e8e\u4f7f\u7528 LLM \u5b9e\u73b0\u641c\u7d22\u7b56\u7565\uff0c\u5e76\u4f7f\u7528 vLLM \u5b9e\u73b0\u901f\u5ea6\u63d0\u5347\u3002<\/section>\n<\/li>\n<\/ul>\n<p>\u90a3\u4e48\uff0c\u8ba1\u7b97\u6700\u4f18\u6269\u5c55\u5728\u5b9e\u8df5\u4e2d\u6548\u679c\u5982\u4f55\u5462\uff1f\u5728\u4e0b\u56fe\u4e2d\uff0c\u5982\u679c\u4f60\u7ed9\u5b83\u4eec\u8db3\u591f\u7684\u300c\u601d\u8003\u65f6\u95f4\u300d\uff0c\u89c4\u6a21\u5f88\u5c0f\u7684 1B \u548c 3B Llama Instruct \u6a21\u578b\u5728\u5177\u6709\u6311\u6218\u6027\u7684 MATH-500 \u57fa\u51c6\u4e0a\uff0c\u8d85\u8d8a\u4e86\u6bd4\u5b83\u4eec\u5927\u5f97\u591a\u7684 8B\u300170B \u6a21\u578b\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-71236e7f6def47cf4d544cbd9fb4602a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-71236e7f6def47cf4d544cbd9fb4602a.png\"><\/a><\/p>\n<p>HuggingFace \u8054\u5408\u521b\u59cb\u4eba\u517c CEO Clem Delangue \u8868\u793a\uff0c\u5728 OpenAI o1 \u516c\u5f00\u4eae\u76f8\u4ec5 10 \u5929\u540e\uff0c\u6211\u4eec\u5f88\u9ad8\u5174\u5730\u63ed\u6653\u4e86\u5176\u6210\u529f\u80cc\u540e\u7684\u7a81\u7834\u6027\u6280\u672f\u7684\u5f00\u6e90\u7248\u672c\uff1a\u6269\u5c55\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u3002\u901a\u8fc7\u7ed9\u6a21\u578b\u66f4\u957f\u7684\u300c\u601d\u8003\u65f6\u95f4\u300d\uff0c1B \u6a21\u578b\u53ef\u4ee5\u51fb\u8d25 8B\u30013B \u6a21\u578b\u53ef\u4ee5\u51fb\u8d25 70B\u3002\u5f53\u7136\uff0c\u5b8c\u6574\u7684\u6280\u672f\u914d\u65b9\u662f\u5f00\u6e90\u7684\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f74e1fe7a283682c5d14396f6d4661ee.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f74e1fe7a283682c5d14396f6d4661ee.png\"><\/a><\/p>\n<p>\u5404\u8def\u7f51\u53cb\u770b\u5230\u8fd9\u4e9b\u7ed3\u679c\u4e5f\u4e0d\u6de1\u5b9a\u4e86\uff0c\u76f4\u547c\u4e0d\u53ef\u601d\u8bae\uff0c\u5e76\u8ba4\u4e3a\u8fd9\u662f\u5c0f\u6a21\u578b\u7684\u80dc\u5229\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d7e523c5669b60f036f36633f04dc3b1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d7e523c5669b60f036f36633f04dc3b1.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-27e93bc885541a8314d6bd4a1a4ab853.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-27e93bc885541a8314d6bd4a1a4ab853.png\"><\/a><\/p>\n<p>\u63a5\u4e0b\u6765\uff0cHuggingFace \u6df1\u5165\u63a2\u8ba8\u4e86\u4ea7\u751f\u4e0a\u8ff0\u7ed3\u679c\u80cc\u540e\u7684\u539f\u56e0\uff0c\u5e76\u5e2e\u52a9\u8bfb\u8005\u4e86\u89e3\u5b9e\u73b0\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u6269\u5c55\u7684\u5b9e\u7528\u7b56\u7565\u3002<\/p>\n<p><strong>\u6269\u5c55\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u7b56\u7565<\/strong><\/p>\n<p>\u6269\u5c55\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u4e3b\u8981\u6709\u4ee5\u4e0b\u4e24\u79cd\u4e3b\u8981\u7b56\u7565\uff1a<\/p>\n<ul>\n<li>\n<section>\u81ea\u6211\u6539\u8fdb\uff1a\u6a21\u578b\u901a\u8fc7\u5728\u540e\u7eed\u8fed\u4ee3\u4e2d\u8bc6\u522b\u548c\u7ea0\u9519\u6765\u8fed\u4ee3\u6539\u8fdb\u81ea\u5df1\u7684\u8f93\u51fa\u6216\u300c\u60f3\u6cd5\u300d\u3002\u867d\u7136\u8fd9\u79cd\u7b56\u7565\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u6709\u6548\uff0c\u4f46\u901a\u5e38\u8981\u6c42\u6a21\u578b\u5177\u6709\u5185\u7f6e\u7684\u81ea\u6211\u6539\u8fdb\u673a\u5236\uff0c\u8fd9\u53ef\u80fd\u4f1a\u9650\u5236\u5176\u9002\u7528\u6027\u3002<\/section>\n<\/li>\n<li>\n<section>\u9488\u5bf9\u9a8c\u8bc1\u5668\u8fdb\u884c\u641c\u7d22\uff1a\u8fd9\u79cd\u65b9\u6cd5\u4fa7\u91cd\u4e8e\u751f\u6210\u591a\u4e2a\u5019\u9009\u7b54\u6848\u5e76\u4f7f\u7528\u9a8c\u8bc1\u5668\u9009\u62e9\u6700\u4f73\u7b54\u6848\u3002\u9a8c\u8bc1\u5668\u53ef\u4ee5\u662f\u57fa\u4e8e\u786c\u7f16\u7801\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\uff0c\u4e5f\u53ef\u4ee5\u662f\u5b66\u5f97\u7684\u5956\u52b1\u6a21\u578b\u3002\u672c\u6587\u5c06\u91cd\u70b9\u4ecb\u7ecd\u5b66\u5f97\u7684\u9a8c\u8bc1\u5668\uff0c\u5b83\u5305\u62ec\u4e86 Best-of-N \u91c7\u6837\u548c\u6811\u641c\u7d22\u7b49\u6280\u672f\u3002\u8fd9\u79cd\u641c\u7d22\u7b56\u7565\u66f4\u7075\u6d3b\uff0c\u53ef\u4ee5\u9002\u5e94\u95ee\u9898\u7684\u96be\u5ea6\uff0c\u4e0d\u8fc7\u5b83\u4eec\u7684\u6027\u80fd\u53d7\u5230\u9a8c\u8bc1\u5668\u8d28\u91cf\u7684\u9650\u5236\u3002<\/section>\n<\/li>\n<\/ul>\n<p>HuggingFace \u4e13\u6ce8\u4e8e\u57fa\u4e8e\u641c\u7d22\u7684\u65b9\u6cd5\uff0c\u5b83\u4eec\u662f\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u4f18\u5316\u7684\u5b9e\u7528\u4e14\u53ef\u6269\u5c55\u7684\u89e3\u51b3\u65b9\u6848\u3002\u4e0b\u9762\u662f\u4e09\u79cd\u7b56\u7565\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-62c75270aeafa49fab48f2c3da7b0f60.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-62c75270aeafa49fab48f2c3da7b0f60.png\"><\/a><\/p>\n<ul>\n<li>\n<section>Best-of-N\uff1a\u901a\u5e38\u4f7f\u7528\u5956\u52b1\u6a21\u578b\u4e3a\u6bcf\u4e2a\u95ee\u9898\u751f\u6210\u591a\u4e2a\u54cd\u5e94\u5e76\u4e3a\u6bcf\u4e2a\u5019\u9009\u7b54\u6848\u5206\u914d\u5206\u6570\uff0c\u7136\u540e\u9009\u62e9\u5956\u52b1\u6700\u9ad8\u7684\u7b54\u6848\uff08\u6216\u7a0d\u540e\u8ba8\u8bba\u7684\u52a0\u6743\u53d8\u4f53\uff09\u3002\u8fd9\u79cd\u65b9\u6cd5\u5f3a\u8c03\u7b54\u6848\u8d28\u91cf\u800c\u975e\u9891\u7387\u3002<\/section>\n<\/li>\n<li>\n<section>\u96c6\u675f\u641c\u7d22\uff1a\u4e00\u79cd\u63a2\u7d22\u89e3\u51b3\u65b9\u6848\u7a7a\u95f4\u7684\u7cfb\u7edf\u641c\u7d22\u65b9\u6cd5\uff0c\u901a\u5e38\u4e0e\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b (PRM) \u7ed3\u5408\u4f7f\u7528\uff0c\u4ee5\u4f18\u5316\u95ee\u9898\u89e3\u51b3\u4e2d\u95f4\u6b65\u9aa4\u7684\u91c7\u6837\u548c\u8bc4\u4f30\u3002\u4e0e\u5bf9\u6700\u7ec8\u7b54\u6848\u4ea7\u751f\u5355\u4e00\u5206\u6570\u7684\u4f20\u7edf\u5956\u52b1\u6a21\u578b\u4e0d\u540c\uff0cPRM \u4f1a\u63d0\u4f9b\u4e00\u7cfb\u5217\u5206\u6570\uff0c\u5176\u4e2d\u63a8\u7406\u8fc7\u7a0b\u7684\u6bcf\u4e2a\u6b65\u9aa4\u90fd\u6709\u4e00\u4e2a\u5206\u6570\u3002\u8fd9\u79cd\u7ec6\u7c92\u5ea6\u53cd\u9988\u80fd\u529b\u4f7f\u5f97 PRM \u6210\u4e3a LLM \u641c\u7d22\u65b9\u6cd5\u7684\u81ea\u7136\u9009\u62e9\u3002<\/section>\n<\/li>\n<li>\n<section>\u591a\u6837\u6027\u9a8c\u8bc1\u5668\u6811\u641c\u7d22 (DVTS)\uff1aHuggingFace \u5f00\u53d1\u7684\u96c6\u675f\u641c\u7d22\u6269\u5c55\uff0c\u5c06\u521d\u59cb\u96c6\u675f\u62c6\u5206\u4e3a\u72ec\u7acb\u7684\u5b50\u6811\uff0c\u7136\u540e\u4f7f\u7528 PRM \u8d2a\u5a6a\u5730\u6269\u5c55\u8fd9\u4e9b\u5b50\u6811\u3002\u8fd9\u79cd\u65b9\u6cd5\u63d0\u9ad8\u4e86\u89e3\u51b3\u65b9\u6848\u7684\u591a\u6837\u6027\u548c\u6574\u4f53\u6027\u80fd\uff0c\u5c24\u5176\u662f\u5728\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u9884\u7b97\u8f83\u5927\u7684\u60c5\u51b5\u4e0b\u3002<\/section>\n<\/li>\n<\/ul>\n<p><strong>\u5b9e\u9a8c\u8bbe\u7f6e<\/strong><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-dda424e18732c9de8f4903c82304f544.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-dda424e18732c9de8f4903c82304f544.png\"><\/a><\/p>\n<p>\u5b9e\u9a8c\u8bbe\u7f6e\u5305\u62ec\u4ee5\u4e0b\u6b65\u9aa4\uff1a<\/p>\n<ul>\n<li>\n<section>\u9996\u5148\u7ed9 LLM \u63d0\u4f9b\u4e00\u4e2a\u6570\u5b66\u95ee\u9898\uff0c\u8ba9\u5176\u751f\u6210 N \u4e2a\u90e8\u5206\u89e3\uff0c\u4f8b\u5982\uff0c\u63a8\u5bfc\u8fc7\u7a0b\u4e2d\u7684\u4e2d\u95f4\u6b65\u9aa4\u3002<\/section>\n<\/li>\n<li>\n<section>\u6bcf\u4e2a step \u90fd\u7531 PRM \u8bc4\u5206\uff0cPRM \u4f30\u8ba1\u6bcf\u4e2a\u6b65\u9aa4\u6700\u7ec8\u8fbe\u5230\u6b63\u786e\u7b54\u6848\u7684\u6982\u7387\u3002<\/section>\n<\/li>\n<li>\n<section>\u4e00\u65e6\u641c\u7d22\u7b56\u7565\u7ed3\u675f\uff0c\u6700\u7ec8\u5019\u9009\u89e3\u51b3\u65b9\u6848\u5c06\u7531 PRM \u6392\u5e8f\u4ee5\u4ea7\u751f\u6700\u7ec8\u7b54\u6848\u3002<\/section>\n<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u6bd4\u8f83\u5404\u79cd\u641c\u7d22\u7b56\u7565\uff0c\u672c\u6587\u4f7f\u7528\u4e86\u4ee5\u4e0b\u5f00\u6e90\u6a21\u578b\u548c\u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li>\n<section>\u6a21\u578b\uff1a\u4f7f\u7528 meta-llama\/Llama-3.2-1B-Instruct \u4e3a\u4e3b\u8981\u6a21\u578b\uff0c\u7528\u4e8e\u6269\u5c55\u6d4b\u8bd5\u65f6\u8ba1\u7b97\uff1b<\/section>\n<\/li>\n<li>\n<section>\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b PRM\uff1a\u4e3a\u4e86\u6307\u5bfc\u641c\u7d22\u7b56\u7565\uff0c\u672c\u6587\u4f7f\u7528\u4e86 RLHFlow\/Llama3.1-8B-PRM-Deepseek-Data\uff0c\u8fd9\u662f\u4e00\u4e2a\u7ecf\u8fc7\u8fc7\u7a0b\u76d1\u7763\u8bad\u7ec3\u7684 80 \u4ebf\u5956\u52b1\u6a21\u578b\u3002\u8fc7\u7a0b\u76d1\u7763\u662f\u4e00\u79cd\u8bad\u7ec3\u65b9\u6cd5\uff0c\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u7684\u6bcf\u4e00\u6b65\u90fd\u4f1a\u6536\u5230\u53cd\u9988\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u6700\u7ec8\u7ed3\u679c\uff1b<\/section>\n<\/li>\n<li>\n<section>\u6570\u636e\u96c6\uff1a\u672c\u6587\u5728 MATH-500 \u5b50\u96c6\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u8fd9\u662f OpenAI \u4f5c\u4e3a\u8fc7\u7a0b\u76d1\u7763\u7814\u7a76\u7684\u4e00\u90e8\u5206\u53d1\u5e03\u7684 MATH \u57fa\u51c6\u6570\u636e\u96c6\u3002\u8fd9\u4e9b\u6570\u5b66\u95ee\u9898\u6db5\u76d6\u4e86\u4e03\u4e2a\u79d1\u76ee\uff0c\u5bf9\u4eba\u7c7b\u548c\u5927\u591a\u6570\u5927\u8bed\u8a00\u6a21\u578b\u6765\u8bf4\u90fd\u5177\u6709\u6311\u6218\u6027\u3002<\/section>\n<\/li>\n<\/ul>\n<p>\u672c\u6587\u5c06\u4ece\u4e00\u4e2a\u7b80\u5355\u7684\u57fa\u7ebf\u5f00\u59cb\uff0c\u7136\u540e\u9010\u6b65\u7ed3\u5408\u5176\u4ed6\u6280\u672f\u6765\u63d0\u9ad8\u6027\u80fd\u3002<\/p>\n<p><strong>\u591a\u6570\u6295\u7968<\/strong><\/p>\n<p>\u591a\u6570\u6295\u7968\u662f\u805a\u5408 LLM \u8f93\u51fa\u7684\u6700\u76f4\u63a5\u65b9\u6cd5\u3002\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u6570\u5b66\u95ee\u9898\uff0c\u4f1a\u751f\u6210 N \u4e2a\u5019\u9009\u89e3\uff0c\u7136\u540e\u9009\u62e9\u51fa\u73b0\u9891\u7387\u6700\u9ad8\u7684\u7b54\u6848\u3002\u5728\u6240\u6709\u7684\u5b9e\u9a8c\u4e2d\uff0c\u672c\u6587\u91c7\u6837\u4e86\u591a\u8fbe N=256 \u4e2a\u5019\u9009\u89e3\uff0c\u6e29\u5ea6\u53c2\u6570 T=0.8\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u95ee\u9898\u751f\u6210\u4e86\u6700\u591a 2048 \u4e2a token\u3002<\/p>\n<p>\u4ee5\u4e0b\u662f\u591a\u6570\u6295\u7968\u5e94\u7528\u4e8e Llama 3.2 1B Instruct \u65f6\u7684\u8868\u73b0\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0f34551343cd8ba51b5433339dcdb52d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0f34551343cd8ba51b5433339dcdb52d.png\"><\/a><\/p>\n<p>\u7ed3\u679c\u8868\u660e\uff0c\u591a\u6570\u6295\u7968\u6bd4\u8d2a\u5a6a\u89e3\u7801\u57fa\u7ebf\u6709\u663e\u8457\u7684\u6539\u8fdb\uff0c\u4f46\u5176\u6536\u76ca\u5728\u5927\u7ea6 N=64 generation \u540e\u5f00\u59cb\u8d8b\u4e8e\u5e73\u7a33\u3002\u8fd9\u79cd\u9650\u5236\u7684\u51fa\u73b0\u662f\u56e0\u4e3a\u591a\u6570\u6295\u7968\u96be\u4ee5\u89e3\u51b3\u9700\u8981\u7ec6\u81f4\u63a8\u7406\u7684\u95ee\u9898\u3002<\/p>\n<p>\u57fa\u4e8e\u591a\u6570\u6295\u7968\u7684\u5c40\u9650\u6027\uff0c\u8ba9\u6211\u4eec\u770b\u770b\u5982\u4f55\u7ed3\u5408\u5956\u52b1\u6a21\u578b\u6765\u63d0\u9ad8\u6027\u80fd\u3002<\/p>\n<p><strong>\u8d85\u8d8a\u591a\u6570\uff1aBest-of-N<\/strong><\/p>\n<p>Best-of-N \u662f\u591a\u6570\u6295\u7968\u7b97\u6cd5\u7684\u7b80\u5355\u4e14\u6709\u6548\u7684\u6269\u5c55\uff0c\u5b83\u4f7f\u7528\u5956\u52b1\u6a21\u578b\u6765\u786e\u5b9a\u6700\u5408\u7406\u7684\u7b54\u6848\u3002\u8be5\u65b9\u6cd5\u6709\u4e24\u79cd\u4e3b\u8981\u53d8\u4f53\uff1a<\/p>\n<p>\u666e\u901a\u7684 Best-of-N\uff1a\u751f\u6210 N \u4e2a\u72ec\u7acb\u54cd\u5e94\uff0c\u9009\u62e9 RM \u5956\u52b1\u6700\u9ad8\u7684\u4e00\u4e2a\u4f5c\u4e3a\u6700\u7ec8\u56de\u7b54\u3002\u8fd9\u786e\u4fdd\u4e86\u9009\u62e9\u7f6e\u4fe1\u5ea6\u6700\u9ad8\u7684\u54cd\u5e94\uff0c\u4f46\u5b83\u5e76\u6ca1\u6709\u8003\u8651\u5230\u56de\u7b54\u4e4b\u95f4\u7684\u4e00\u81f4\u6027\u3002<\/p>\n<p>\u52a0\u6743 Best-of-N\uff1a\u6c47\u603b\u6240\u6709\u76f8\u540c\u54cd\u5e94\u7684\u5f97\u5206\uff0c\u5e76\u9009\u62e9\u603b\u5956\u52b1\u6700\u9ad8\u7684\u56de\u7b54\u3002\u8fd9\u79cd\u65b9\u6cd5\u901a\u8fc7\u91cd\u590d\u51fa\u73b0\u6765\u63d0\u9ad8\u5206\u6570\uff0c\u4ece\u800c\u4f18\u5148\u8003\u8651\u9ad8\u8d28\u91cf\u7684\u56de\u7b54\u3002\u4ece\u6570\u5b66\u4e0a\u8bb2\uff0c\u56de\u7b54\u7684\u6743\u91cd a_i\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-5c6dec21b417eec7ff9e8cb90719f366.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-5c6dec21b417eec7ff9e8cb90719f366.png\"><\/a><\/p>\n<p>\u5176\u4e2d\uff0cRM (p,s_i) \u662f\u5bf9\u4e8e\u95ee\u9898 p \u7684\u7b2c i \u4e2a\u89e3\u51b3\u65b9\u6848 s_i \u7684\u5956\u52b1\u6a21\u578b\u5206\u6570\u3002<\/p>\n<p>\u901a\u5e38\uff0c\u4eba\u4eec\u4f7f\u7528\u7ed3\u679c\u5956\u52b1\u6a21\u578b (ORM) \u6765\u83b7\u5f97\u5355\u4e2a\u89e3\u51b3\u65b9\u6848\u7ea7\u522b\u7684\u5206\u6570\u3002\u4f46\u4e3a\u4e86\u4e0e\u5176\u4ed6\u641c\u7d22\u7b56\u7565\u8fdb\u884c\u516c\u5e73\u6bd4\u8f83\uff0c\u4f7f\u7528\u76f8\u540c\u7684 PRM \u5bf9 Best-of-N \u7684\u89e3\u51b3\u65b9\u6848\u8fdb\u884c\u8bc4\u5206\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0cPRM \u4e3a\u6bcf\u4e2a\u89e3\u51b3\u65b9\u6848\u751f\u6210\u4e00\u4e2a\u7d2f\u79ef\u7684\u6b65\u9aa4\u7ea7\u5206\u6570\u5e8f\u5217\uff0c\u56e0\u6b64\u9700\u8981\u5bf9\u6b65\u9aa4\u8fdb\u884c\u89c4\u7ea6\uff08reduction\uff09\u4ee5\u83b7\u5f97\u5355\u4e2a\u89e3\u51b3\u65b9\u6848\u7ea7\u5206\u6570\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7f1494d1a99155a8035ce36d0ff81c9e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7f1494d1a99155a8035ce36d0ff81c9e.png\"><\/a><\/p>\n<p>\u6700\u5e38\u89c1\u7684\u89c4\u7ea6\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\n<p>Min\uff1a\u4f7f\u7528\u6240\u6709\u6b65\u9aa4\u4e2d\u7684\u6700\u4f4e\u5206\u6570\u3002<\/p>\n<\/li>\n<li>\n<p>Prod\uff1a\u4f7f\u7528\u9636\u68af\u5206\u6570\u7684\u4e58\u79ef\u3002<\/p>\n<\/li>\n<li>\n<p>Last\uff1a\u4f7f\u7528\u6b65\u9aa4\u4e2d\u7684\u6700\u7ec8\u5206\u6570\u3002\u8be5\u5206\u6570\u5305\u542b\u6240\u6709\u5148\u524d\u6b65\u9aa4\u7684\u7d2f\u79ef\u4fe1\u606f\uff0c\u56e0\u6b64\u5c06 PRM \u6709\u6548\u5730\u89c6\u4e3a\u80fd\u591f\u5bf9\u90e8\u5206\u89e3\u51b3\u65b9\u6848\u8fdb\u884c\u8bc4\u5206\u7684 ORM\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4ee5\u4e0b\u662f\u5e94\u7528 Best-of-N \u7684\u4e24\u79cd\u53d8\u4f53\u5f97\u5230\u7684\u7ed3\u679c\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-cd788c8aaaf5b8ed491bb6f9c131eeff.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-cd788c8aaaf5b8ed491bb6f9c131eeff.png\"><\/a><\/p>\n<p>\u7ed3\u679c\u63ed\u793a\u4e86\u4e00\u4e2a\u660e\u663e\u7684\u4f18\u52bf\uff1a\u52a0\u6743\u7684 Best-of-N \u59cb\u7ec8\u4f18\u4e8e\u666e\u901a\u7684 Best-of-N\uff0c\u7279\u522b\u662f\u5728\u53d1\u7535\u9884\u7b97\u8f83\u5927\u7684\u60c5\u51b5\u4e0b\u3002\u5b83\u80fd\u591f\u6c47\u603b\u76f8\u540c\u7b54\u6848\u7684\u5206\u6570\uff0c\u786e\u4fdd\u5373\u4f7f\u9891\u7387\u8f83\u4f4e\u4f46\u8d28\u91cf\u8f83\u9ad8\u7684\u7b54\u6848\u4e5f\u80fd\u5f97\u5230\u6709\u6548\u7684\u4f18\u5148\u5904\u7406\u3002<\/p>\n<p>\u7136\u800c\uff0c\u5c3d\u7ba1\u6709\u8fd9\u4e9b\u6539\u8fdb\uff0c\u4ecd\u7136\u8fbe\u4e0d\u5230 Llama 8B \u6a21\u578b\u6240\u8fbe\u5230\u7684\u6027\u80fd\uff0c\u5e76\u4e14\u5728 N=256 \u65f6 Best-of-N \u65b9\u6cd5\u5f00\u59cb\u8d8b\u4e8e\u7a33\u5b9a\u3002<\/p>\n<p>\u53ef\u4ee5\u901a\u8fc7\u9010\u6b65\u76d1\u7763\u641c\u7d22\u8fc7\u7a0b\u6765\u8fdb\u4e00\u6b65\u7a81\u7834\u754c\u9650\u5417\uff1f<\/p>\n<p><strong>\u4f7f\u7528 PRM \u7684\u96c6\u675f\u641c\u7d22<\/strong><\/p>\n<p>\u4f5c\u4e3a\u4e00\u79cd\u7ed3\u6784\u5316\u641c\u7d22\u65b9\u6cd5\uff0c\u96c6\u675f\u641c\u7d22\u53ef\u4ee5\u7cfb\u7edf\u5730\u63a2\u7d22\u89e3\u51b3\u65b9\u6848\u7a7a\u95f4\uff0c\u4f7f\u5176\u6210\u4e3a\u5728\u6d4b\u8bd5\u65f6\u6539\u8fdb\u6a21\u578b\u8f93\u51fa\u7684\u5f3a\u5927\u5de5\u5177\u3002\u4e0e PRM \u7ed3\u5408\u4f7f\u7528\u65f6\uff0c\u96c6\u675f\u641c\u7d22\u53ef\u4ee5\u4f18\u5316\u95ee\u9898\u89e3\u51b3\u4e2d\u4e2d\u95f4\u6b65\u9aa4\u7684\u751f\u6210\u548c\u8bc4\u4f30\u3002\u96c6\u675f\u641c\u7d22\u7684\u5de5\u4f5c\u65b9\u5f0f\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\n<section>\u901a\u8fc7\u4fdd\u6301\u56fa\u5b9a\u6570\u91cf\u7684\u300c\u96c6\u675f\u300d\u6216\u6d3b\u52a8\u8def\u5f84 N \uff0c\u8fed\u4ee3\u751f\u6210\u591a\u4e2a\u5019\u9009\u89e3\u51b3\u65b9\u6848\u3002<\/section>\n<\/li>\n<li>\n<section>\u5728\u7b2c\u4e00\u6b21\u8fed\u4ee3\u4e2d\uff0c\u4ece\u6e29\u5ea6\u4e3a T \u7684 LLM \u4e2d\u62bd\u53d6 N \u4e2a\u72ec\u7acb\u6b65\u9aa4\uff0c\u4ee5\u5f15\u5165\u54cd\u5e94\u7684\u591a\u6837\u6027\u3002\u8fd9\u4e9b\u6b65\u9aa4\u901a\u5e38\u7531\u505c\u6b62\u6807\u51c6\u5b9a\u4e49\uff0c\u4f8b\u5982\u7ec8\u6b62\u4e8e\u65b0\u884c n \u6216\u53cc\u65b0\u884c nn\u3002<\/section>\n<\/li>\n<li>\n<section>\u4f7f\u7528 PRM \u5bf9\u6bcf\u4e2a\u6b65\u9aa4\u8fdb\u884c\u8bc4\u5206\uff0c\u5e76\u9009\u62e9\u524d N\/M \u4e2a\u6b65\u9aa4\u4f5c\u4e3a\u4e0b\u4e00\u8f6e\u751f\u6210\u7684\u5019\u9009\u3002\u8fd9\u91cc M \u8868\u793a\u7ed9\u5b9a\u6d3b\u52a8\u8def\u5f84\u7684\u300c\u96c6\u675f\u5bbd\u5ea6\u300d\u3002\u4e0e Best-of-N \u4e00\u6837\uff0c\u4f7f\u7528\u300c\u6700\u540e\u300d\u7684\u89c4\u7ea6\u6765\u5bf9\u6bcf\u6b21\u8fed\u4ee3\u7684\u90e8\u5206\u89e3\u51b3\u65b9\u6848\u8fdb\u884c\u8bc4\u5206\u3002<\/section>\n<\/li>\n<li>\n<section>\u901a\u8fc7\u5728\u89e3\u51b3\u65b9\u6848\u4e2d\u91c7\u6837 M \u4e2a\u540e\u7eed\u6b65\u9aa4\u6765\u6269\u5c55\u5728\u6b65\u9aa4 (3) \u4e2d\u9009\u62e9\u7684\u6b65\u9aa4\u3002<\/section>\n<\/li>\n<li>\n<section>\u91cd\u590d\u6b65\u9aa4 (3) \u548c (4)\uff0c\u76f4\u5230\u8fbe\u5230 EOS token \u6216\u8d85\u8fc7\u6700\u5927\u641c\u7d22\u6df1\u5ea6\u3002<\/section>\n<\/li>\n<\/ol>\n<p>\u901a\u8fc7\u5141\u8bb8 PRM \u8bc4\u4f30\u4e2d\u95f4\u6b65\u9aa4\u7684\u6b63\u786e\u6027\uff0c\u96c6\u675f\u641c\u7d22\u53ef\u4ee5\u5728\u6d41\u7a0b\u65e9\u671f\u8bc6\u522b\u5e76\u4f18\u5148\u8003\u8651\u6709\u5e0c\u671b\u7684\u8def\u5f84\u3002\u8fd9\u79cd\u9010\u6b65\u8bc4\u4f30\u7b56\u7565\u5bf9\u4e8e\u6570\u5b66\u7b49\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u7279\u522b\u6709\u7528\uff0c\u8fd9\u662f\u56e0\u4e3a\u9a8c\u8bc1\u90e8\u5206\u89e3\u51b3\u65b9\u6848\u53ef\u4ee5\u663e\u8457\u6539\u5584\u6700\u7ec8\u7ed3\u679c\u3002<\/p>\n<p><strong>\u5b9e\u73b0\u7ec6\u8282<\/strong><\/p>\n<p>\u5728\u5b9e\u9a8c\u4e2d\uff0cHuggingFace \u9075\u5faa DeepMind \u7684\u8d85\u53c2\u6570\u9009\u62e9\uff0c\u5e76\u6309\u7167\u4ee5\u4e0b\u65b9\u5f0f\u8fd0\u884c\u96c6\u675f\u641c\u7d22\uff1a<\/p>\n<ul>\n<li>\n<section>\u8ba1\u7b97\u6269\u5c55\u4e3a 4\u300116\u300164\u3001256 \u65f6\u7684 N \u4e2a\u96c6\u675f<\/section>\n<\/li>\n<li>\n<section>\u56fa\u5b9a\u96c6\u675f\u5bbd\u5ea6 M=4<\/section>\n<\/li>\n<li>\n<section>\u5728\u6e29\u5ea6 T=0.8 \u65f6\u91c7\u6837<\/section>\n<\/li>\n<li>\n<section>\u6700\u591a 40 \u6b21\u8fed\u4ee3\uff0c\u5373\u6700\u5927\u6df1\u5ea6\u4e3a 40 \u6b65\u7684\u6811<\/section>\n<\/li>\n<\/ul>\n<p>\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u7ed3\u679c\u4ee4\u4eba\u9707\u60ca\uff1a\u5728 N=4 \u7684\u6d4b\u8bd5\u65f6\u9884\u7b97\u4e0b\uff0c\u96c6\u675f\u641c\u7d22\u5b9e\u73b0\u4e86\u4e0e N=16 \u65f6 Best-of-N \u76f8\u540c\u7684\u51c6\u786e\u7387\uff0c\u5373\u8ba1\u7b97\u6548\u7387\u63d0\u9ad8\u4e86 4 \u500d\uff01\u6b64\u5916\uff0c\u96c6\u675f\u641c\u7d22\u7684\u6027\u80fd\u4e0e Llama 3.1 8B \u76f8\u5f53\uff0c\u6bcf\u4e2a\u95ee\u9898\u4ec5\u9700 N=32 \u89e3\u51b3\u65b9\u6848\u3002\u8ba1\u7b97\u673a\u79d1\u5b66\u535a\u58eb\u751f\u5728\u6570\u5b66\u65b9\u9762\u7684\u5e73\u5747\u8868\u73b0\u7ea6\u4e3a 40%\uff0c\u56e0\u6b64\u5bf9\u4e8e 1B \u6a21\u578b\u6765\u8bf4\uff0c\u63a5\u8fd1 55% \u5df2\u7ecf\u5f88\u4e0d\u9519\u4e86\uff01<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-316b688aeaeabb8d6a637f2576fde939.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-316b688aeaeabb8d6a637f2576fde939.png\"><\/a><\/p>\n<p><strong>\u54ea\u4e9b\u95ee\u9898\u96c6\u675f\u641c\u7d22\u89e3\u51b3\u5f97\u6700\u597d<\/strong><\/p>\n<p>\u867d\u7136\u603b\u4f53\u4e0a\u5f88\u660e\u663e\uff0c\u96c6\u675f\u641c\u7d22\u662f\u4e00\u79cd\u6bd4 Best-of-N \u6216\u591a\u6570\u6295\u7968\u66f4\u597d\u7684\u641c\u7d22\u7b56\u7565\uff0c\u4f46 DeepMind \u7684\u8bba\u6587\u8868\u660e\uff0c\u6bcf\u79cd\u7b56\u7565\u90fd\u6709\u6743\u8861\uff0c\u8fd9\u53d6\u51b3\u4e8e\u95ee\u9898\u7684\u96be\u5ea6\u548c\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u9884\u7b97\u3002<\/p>\n<p>\u4e3a\u4e86\u4e86\u89e3\u54ea\u4e9b\u95ee\u9898\u6700\u9002\u5408\u54ea\u79cd\u7b56\u7565\uff0cDeepMind \u8ba1\u7b97\u4e86\u4f30\u8ba1\u95ee\u9898\u96be\u5ea6\u7684\u5206\u5e03\uff0c\u5e76\u5c06\u7ed3\u679c\u5206\u6210\u4e94\u7b49\u5206\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u6bcf\u4e2a\u95ee\u9898\u88ab\u5206\u914d\u5230 5 \u4e2a\u7ea7\u522b\u4e4b\u4e00\uff0c\u5176\u4e2d\u7ea7\u522b 1 \u8868\u793a\u8f83\u5bb9\u6613\u7684\u95ee\u9898\uff0c\u7ea7\u522b 5 \u8868\u793a\u6700\u96be\u7684\u95ee\u9898\u3002\u4e3a\u4e86\u4f30\u8ba1\u95ee\u9898\u96be\u5ea6\uff0cDeepMind \u4e3a\u6bcf\u4e2a\u95ee\u9898\u751f\u6210\u4e86 2048 \u4e2a\u5019\u9009\u89e3\u51b3\u65b9\u6848\uff0c\u5e76\u8fdb\u884c\u4e86\u6807\u51c6\u91c7\u6837\uff0c\u7136\u540e\u63d0\u51fa\u4e86\u4ee5\u4e0b\u542f\u53d1\u5f0f\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li>\n<section>Oracle\uff1a\u4f7f\u7528\u57fa\u672c\u4e8b\u5b9e\u6807\u7b7e\u4f30\u8ba1\u6bcf\u4e2a\u95ee\u9898\u7684 pass@1 \u5206\u6570\uff0c\u5bf9 pass@1 \u5206\u6570\u7684\u5206\u5e03\u8fdb\u884c\u5206\u7c7b\u4ee5\u786e\u5b9a\u4e94\u5206\u4f4d\u6570\u3002<\/section>\n<\/li>\n<li>\n<section>\u6a21\u578b\uff1a\u4f7f\u7528\u6bcf\u4e2a\u95ee\u9898\u7684\u5e73\u5747 PRM \u5206\u6570\u5206\u5e03\u6765\u786e\u5b9a\u4e94\u5206\u4f4d\u6570\u3002\u8fd9\u91cc\u7684\u76f4\u89c9\u662f\uff1a\u66f4\u96be\u7684\u95ee\u9898\u5206\u6570\u4f1a\u66f4\u4f4e\u3002<\/section>\n<\/li>\n<\/ul>\n<p>\u4e0b\u56fe\u662f\u6839\u636e pass@1 \u5206\u6570\u548c\u56db\u4e2a\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u9884\u7b97 N=[4,16,64,256] \u5bf9\u5404\u79cd\u65b9\u6cd5\u7684\u7ec6\u5206\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-46c3d1f8aec64429bad5df7acc9fa189.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-46c3d1f8aec64429bad5df7acc9fa189.png\"><\/a><\/p>\n<p>\u53ef\u4ee5\u770b\u5230\uff0c\u6bcf\u4e2a\u6761\u5f62\u8868\u793a\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u9884\u7b97\uff0c\u5e76\u4e14\u5728\u6bcf\u4e2a\u6761\u5f62\u5185\u663e\u793a\u6bcf\u79cd\u65b9\u6cd5\u7684\u76f8\u5bf9\u51c6\u786e\u5ea6\u3002\u4f8b\u5982\u5728\u96be\u5ea6\u7b49\u7ea7 2 \u7684\u56db\u4e2a\u6761\u5f62\u4e2d\uff1a<\/p>\n<p>\u591a\u6570\u6295\u7968\u662f\u6240\u6709\u8ba1\u7b97\u9884\u7b97\u4e2d\u8868\u73b0\u6700\u5dee\u7684\u65b9\u6cd5\uff0c\u9664\u4e86 N=256\uff08\u96c6\u675f\u641c\u7d22\u8868\u73b0\u6700\u5dee\uff09\u3002<\/p>\n<p>\u96c6\u675f\u641c\u7d22\u6700\u9002\u5408 N=[4,16,64]\uff0c\u4f46 Best-of-N \u6700\u9002\u5408 N=256\u3002<\/p>\n<p>\u5e94\u8be5\u8bf4\uff0c\u96c6\u675f\u641c\u7d22\u5728\u4e2d\u7b49\u96be\u5ea6\u548c\u56f0\u96be\u96be\u5ea6\u95ee\u9898\uff083-5 \u7ea7\uff09\u4e2d\u53d6\u5f97\u4e86\u6301\u7eed\u7684\u8fdb\u5c55\uff0c\u4f46\u5728\u8f83\u7b80\u5355\u95ee\u9898\u4e0a\uff0c\u5c24\u5176\u662f\u5728\u8ba1\u7b97\u9884\u7b97\u8f83\u5927\u7684\u60c5\u51b5\u4e0b\uff0c\u5b83\u7684\u8868\u73b0\u5f80\u5f80\u6bd4 Best-of-N\uff08\u751a\u81f3\u591a\u6570\u6295\u7968\uff09\u66f4\u5dee\u3002<\/p>\n<p>\u901a\u8fc7\u89c2\u5bdf\u96c6\u675f\u641c\u7d22\u751f\u6210\u7684\u7ed3\u679c\u6811\uff0cHuggingFace \u610f\u8bc6\u5230\uff0c\u5982\u679c\u5355\u4e2a\u6b65\u9aa4\u88ab\u5206\u914d\u4e86\u9ad8\u5956\u52b1\uff0c\u90a3\u4e48\u6574\u68f5\u6811\u5c31\u5728\u8be5\u8f68\u8ff9\u4e0a\u5d29\u6e83\uff0c\u4ece\u800c\u5f71\u54cd\u591a\u6837\u6027\u3002\u8fd9\u4fc3\u4f7f\u4ed6\u4eec\u63a2\u7d22\u4e00\u79cd\u6700\u5927\u5316\u591a\u6837\u6027\u7684\u96c6\u675f\u641c\u7d22\u6269\u5c55\u3002<\/p>\n<p><strong>DVTS\uff1a\u901a\u8fc7\u591a\u6837\u6027\u63d0\u5347\u6027\u80fd<\/strong><\/p>\n<p>\u6b63\u5982\u4e0a\u9762\u6240\u770b\u5230\u7684\uff0c\u96c6\u675f\u641c\u7d22\u6bd4 Best-of-N \u5177\u6709\u66f4\u597d\u7684\u6027\u80fd\uff0c\u4f46\u5728\u5904\u7406\u7b80\u5355\u95ee\u9898\u548c\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u9884\u7b97\u8f83\u5927\u65f6\u5f80\u5f80\u8868\u73b0\u4e0d\u4f73\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cHuggingFace \u5f00\u53d1\u4e86\u4e00\u4e2a\u6269\u5c55\uff0c\u79f0\u4e4b\u4e3a\u300c\u591a\u6837\u6027\u9a8c\u8bc1\u5668\u6811\u641c\u7d22\u300d\uff08DVTS\uff09\uff0c\u65e8\u5728\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8 N \u8f83\u5927\u65f6\u7684\u591a\u6837\u6027\u3002<\/p>\n<p>DVTS \u7684\u5de5\u4f5c\u65b9\u5f0f\u4e0e\u96c6\u675f\u641c\u7d22\u7c7b\u4f3c\uff0c\u4f46\u6709\u4ee5\u4e0b\u4fee\u6539\uff1a<\/p>\n<ul>\n<li>\n<section>\u5bf9\u4e8e\u7ed9\u5b9a\u7684 N \u548c M\uff0c\u5c06\u521d\u59cb\u96c6\u675f\u6269\u5c55\u4e3a N\/M \u4e2a\u72ec\u7acb\u5b50\u6811\u3002<\/section>\n<\/li>\n<li>\n<section>\u5bf9\u4e8e\u6bcf\u4e2a\u5b50\u6811\uff0c\u9009\u62e9\u5177\u6709\u6700\u9ad8 PRM \u5206\u6570\u7684\u6b65\u9aa4\u3002<\/section>\n<\/li>\n<li>\n<section>\u4ece\u6b65\u9aa4 (2) \u4e2d\u9009\u62e9\u7684\u8282\u70b9\u751f\u6210 M \u4e2a\u65b0\u6b65\u9aa4\uff0c\u5e76\u9009\u62e9\u5177\u6709\u6700\u9ad8 PRM \u5206\u6570\u7684\u6b65\u9aa4\u3002<\/section>\n<\/li>\n<li>\n<section>\u91cd\u590d\u6b65\u9aa4 (3)\uff0c\u76f4\u5230\u8fbe\u5230 EOS token \u6216\u6700\u5927\u6811\u6df1\u5ea6\u3002<\/section>\n<\/li>\n<\/ul>\n<p>\u4e0b\u56fe\u662f\u5c06 DVTS \u5e94\u7528\u4e8e Llama 1B \u7684\u7ed3\u679c\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-70bd79e8641190bf4b5b0c18fe5b85aa.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-70bd79e8641190bf4b5b0c18fe5b85aa.png\"><\/a><\/p>\n<p>\u53ef\u4ee5\u770b\u5230\uff0cDVTS \u4e3a\u96c6\u675f\u641c\u7d22\u63d0\u4f9b\u4e86\u4e00\u79cd\u8865\u5145\u7b56\u7565\uff1a\u5728 N \u8f83\u5c0f\u65f6\uff0c\u96c6\u675f\u641c\u7d22\u66f4\u6709\u6548\u5730\u627e\u5230\u6b63\u786e\u7684\u89e3\u51b3\u65b9\u6848\uff1b\u4f46\u5728 N \u8f83\u5927\u65f6\uff0cDVTS \u5019\u9009\u7684\u591a\u6837\u6027\u5f00\u59cb\u53d1\u6325\u4f5c\u7528\uff0c\u53ef\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n<p>\u6b64\u5916\u5728\u95ee\u9898\u96be\u5ea6\u7ec6\u5206\u4e2d\uff0cDVTS \u5728 N \u8f83\u5927\u65f6\u63d0\u9ad8\u4e86\u7b80\u5355 \/ \u4e2d\u7b49\u95ee\u9898\u7684\u6027\u80fd\uff0c\u800c\u96c6\u675f\u641c\u7d22\u5728 N \u8f83\u5c0f\u65f6\u8868\u73b0\u6700\u4f73\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-b64fb797ae787c4081bfe0f4b04d6d9d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-b64fb797ae787c4081bfe0f4b04d6d9d.png\"><\/a><\/p>\n<p><strong>\u8ba1\u7b97 &#8211; \u6700\u4f18\u6269\u5c55\uff08compute-optimal scaling\uff09<\/strong><\/p>\n<p>\u6709\u4e86\u5404\u79cd\u5404\u6837\u7684\u641c\u7d22\u7b56\u7565\uff0c\u4e00\u4e2a\u81ea\u7136\u7684\u95ee\u9898\u662f\u54ea\u4e00\u4e2a\u662f\u6700\u597d\u7684\uff1f\u5728 DeepMind \u7684\u8bba\u6587\u4e2d\uff08\u53ef\u53c2\u8003\u300aScaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters \u300b\uff09\uff0c\u4ed6\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u8ba1\u7b97 &#8211; \u6700\u4f18\u6269\u5c55\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u53ef\u4ee5\u9009\u62e9\u641c\u7d22\u65b9\u6cd5\u548c\u8d85\u53c2\u6570 \u03b8\uff0c\u4ee5\u4fbf\u5728\u7ed9\u5b9a\u7684\u8ba1\u7b97\u9884\u7b97 N \u4e0b\u8fbe\u5230\u6700\u4f73\u6027\u80fd\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0b02fbf92972ad180d617740ff53f52c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0b02fbf92972ad180d617740ff53f52c.png\"><\/a><\/p>\n<p>\u5176\u4e2d<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7fa6328b53836f654a03ab8e34b5e616.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7fa6328b53836f654a03ab8e34b5e616.png\"><\/a>\u662f\u95ee\u9898 q \u7684\u6b63\u786e\u7b54\u6848\u3002<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-fa649d9d6373a408662dd252c19bf9a6.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-fa649d9d6373a408662dd252c19bf9a6.png\"><\/a>\u8868\u793a\u8ba1\u7b97 &#8211; \u6700\u4f18\u7684\u6269\u5c55\u7b56\u7565\u3002\u7531\u4e8e\u76f4\u63a5\u8ba1\u7b97<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4047c16c48cc723ddd40c12ff9d4e7d7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4047c16c48cc723ddd40c12ff9d4e7d7.png\"><\/a>\u6709\u4e9b\u68d8\u624b\uff0cDeepMind \u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u95ee\u9898\u96be\u5ea6\u7684\u8fd1\u4f3c\u65b9\u6cd5\uff0c\u5373\u6839\u636e\u54ea\u79cd\u641c\u7d22\u7b56\u7565\u5728\u7ed9\u5b9a\u96be\u5ea6\u7ea7\u522b\u4e0a\u8fbe\u5230\u6700\u4f73\u6027\u80fd\u6765\u5206\u914d\u6d4b\u8bd5\u65f6\u7684\u8ba1\u7b97\u8d44\u6e90\u3002<\/p>\n<p>\u4f8b\u5982\uff0c\u5bf9\u4e8e\u8f83\u7b80\u5355\u7684\u95ee\u9898\u548c\u8f83\u4f4e\u7684\u8ba1\u7b97\u9884\u7b97\uff0c\u6700\u597d\u4f7f\u7528 Best-of-N \u7b49\u7b56\u7565\uff0c\u800c\u5bf9\u4e8e\u8f83\u96be\u7684\u95ee\u9898\uff0c\u96c6 shu \u641c\u7d22\u662f\u66f4\u597d\u7684\u9009\u62e9\u3002\u4e0b\u56fe\u4e3a\u8ba1\u7b97 &#8211; \u6700\u4f18\u66f2\u7ebf\uff01<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-1ca843960a125522a0b30ac930166198.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-1ca843960a125522a0b30ac930166198.png\"><\/a><\/p>\n<p><strong>\u6269\u5c55\u5230\u66f4\u5927\u7684\u6a21\u578b<\/strong><\/p>\n<p>\u672c\u6587\u8fd8\u63a2\u7d22\u4e86\u5c06\u8ba1\u7b97 &#8211; \u6700\u4f18\uff08compute-optimal\uff09\u7684\u65b9\u6cd5\u6269\u5c55\u5230 Llama 3.2 3B Instruct \u6a21\u578b\uff0c\u4ee5\u89c2\u5bdf PRM \u5728\u4e0e\u7b56\u7565\u81ea\u8eab\u5bb9\u91cf\u76f8\u6bd4\u65f6\u5728\u54ea\u4e2a\u70b9\u5f00\u59cb\u51cf\u5f31\u3002\u7ed3\u679c\u663e\u793a\uff0c\u8ba1\u7b97 &#8211; \u6700\u4f18\u7684\u6269\u5c55\u6548\u679c\u975e\u5e38\u597d\uff0c3B \u6a21\u578b\u7684\u6027\u80fd\u8d85\u8fc7\u4e86 Llama 3.1 70B Instruct\uff08\u540e\u8005\u662f\u524d\u8005\u5927\u5c0f\u7684 22 \u500d\uff01\uff09\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2f39513b491cf50bb950739b996db529.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2f39513b491cf50bb950739b996db529.png\"><\/a><\/p>\n<p><strong>\u63a5\u4e0b\u6765\u8be5\u600e\u4e48\u529e\uff1f<\/strong><\/p>\n<p>\u5bf9\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u6269\u5c55\u7684\u63a2\u7d22\u63ed\u793a\u4e86\u5229\u7528\u57fa\u4e8e\u641c\u7d22\u7684\u65b9\u6cd5\u7684\u6f5c\u529b\u548c\u6311\u6218\u3002\u5c55\u671b\u672a\u6765\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u51e0\u4e2a\u4ee4\u4eba\u5174\u594b\u7684\u65b9\u5411\uff1a<\/p>\n<ul>\n<li>\n<section>\u5f3a\u9a8c\u8bc1\u5668\uff1a\u5f3a\u9a8c\u8bc1\u5668\u5728\u63d0\u9ad8\u6027\u80fd\u65b9\u9762\u53d1\u6325\u7740\u5173\u952e\u4f5c\u7528\uff0c\u63d0\u9ad8\u9a8c\u8bc1\u5668\u7684\u7a33\u5065\u6027\u548c\u901a\u7528\u6027\u5bf9\u4e8e\u63a8\u8fdb\u8fd9\u4e9b\u65b9\u6cd5\u81f3\u5173\u91cd\u8981\uff1b<\/section>\n<\/li>\n<li>\n<section>\u81ea\u6211\u9a8c\u8bc1\uff1a\u6700\u7ec8\u76ee\u6807\u662f\u5b9e\u73b0\u81ea\u6211\u9a8c\u8bc1\uff0c\u5373\u6a21\u578b\u53ef\u4ee5\u81ea\u4e3b\u9a8c\u8bc1\u81ea\u5df1\u7684\u8f93\u51fa\u3002\u8fd9\u79cd\u65b9\u6cd5\u4f3c\u4e4e\u662f o1 \u7b49\u6a21\u578b\u6b63\u5728\u505a\u7684\u4e8b\u60c5\uff0c\u4f46\u5728\u5b9e\u8df5\u4e2d\u4ecd\u7136\u96be\u4ee5\u5b9e\u73b0\u3002\u4e0e\u6807\u51c6\u76d1\u7763\u5fae\u8c03 (SFT) \u4e0d\u540c\uff0c\u81ea\u6211\u9a8c\u8bc1\u9700\u8981\u66f4\u7ec6\u81f4\u7684\u7b56\u7565\uff1b<\/section>\n<\/li>\n<li>\n<section>\u5c06\u601d\u7ef4\u878d\u5165\u8fc7\u7a0b\uff1a\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u878d\u5165\u660e\u786e\u7684\u4e2d\u95f4\u6b65\u9aa4\u6216\u601d\u7ef4\u53ef\u4ee5\u8fdb\u4e00\u6b65\u589e\u5f3a\u63a8\u7406\u548c\u51b3\u7b56\u80fd\u529b\u3002\u901a\u8fc7\u5c06\u7ed3\u6784\u5316\u63a8\u7406\u878d\u5165\u641c\u7d22\u8fc7\u7a0b\uff0c\u53ef\u4ee5\u5728\u590d\u6742\u4efb\u52a1\u4e0a\u5b9e\u73b0\u66f4\u597d\u7684\u8868\u73b0\uff1b<\/section>\n<\/li>\n<li>\n<section>\u641c\u7d22\u4f5c\u4e3a\u6570\u636e\u751f\u6210\u5de5\u5177\uff1a\u8be5\u65b9\u6cd5\u8fd8\u53ef\u4ee5\u5145\u5f53\u5f3a\u5927\u7684\u6570\u636e\u751f\u6210\u8fc7\u7a0b\uff0c\u521b\u5efa\u9ad8\u8d28\u91cf\u7684\u8bad\u7ec3\u6570\u636e\u96c6\u3002\u4f8b\u5982\uff0c\u6839\u636e\u641c\u7d22\u4ea7\u751f\u7684\u6b63\u786e\u8f68\u8ff9\u5bf9 Llama 1B \u7b49\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u53ef\u4ee5\u5e26\u6765\u663e\u8457\u7684\u6536\u76ca\u3002\u8fd9\u79cd\u57fa\u4e8e\u7b56\u7565\u7684\u65b9\u6cd5\u7c7b\u4f3c\u4e8e ReST \u6216 V-StaR \u7b49\u6280\u672f\uff0c\u4f46\u5177\u6709\u641c\u7d22\u7684\u989d\u5916\u4f18\u52bf\uff0c\u4e3a\u8fed\u4ee3\u6539\u8fdb\u63d0\u4f9b\u4e86\u4e00\u4e2a\u6709\u5e0c\u671b\u7684\u65b9\u5411\uff1b<\/section>\n<\/li>\n<li>\n<section>\u8c03\u7528\u66f4\u591a\u7684 PRM\uff1aPRM \u76f8\u5bf9\u8f83\u5c11\uff0c\u9650\u5236\u4e86\u5176\u66f4\u5e7f\u6cdb\u7684\u5e94\u7528\u3002\u4e3a\u4e0d\u540c\u9886\u57df\u5f00\u53d1\u548c\u5171\u4eab\u66f4\u591a PRM \u662f\u793e\u533a\u53ef\u4ee5\u505a\u51fa\u91cd\u5927\u8d21\u732e\u7684\u5173\u952e\u9886\u57df\u3002<\/section>\n<\/li>\n<\/ul>\n<p><em><sup>\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/spaces\/HuggingFaceH4\/blogpost-scaling-test-time-compute<\/sup><\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-17-7\" target=\"_blank\">3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d257 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-37321","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=37321\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90\",\"datePublished\":\"2025-01-26T00:00:15+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321\"},\"wordCount\":245,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-8d824c26223cb12949e12f33a675e983.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=37321#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=37321\",\"name\":\"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-8d824c26223cb12949e12f33a675e983.png\",\"datePublished\":\"2025-01-26T00:00:15+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=37321\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-8d824c26223cb12949e12f33a675e983.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-8d824c26223cb12949e12f33a675e983.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37321#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=37321","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=37321#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=37321"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90","datePublished":"2025-01-26T00:00:15+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=37321"},"wordCount":245,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=37321#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8d824c26223cb12949e12f33a675e983.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=37321#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=37321","url":"https:\/\/17aitech.com\/?p=37321","name":"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=37321#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=37321#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8d824c26223cb12949e12f33a675e983.png","datePublished":"2025-01-26T00:00:15+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=37321#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=37321"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=37321#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8d824c26223cb12949e12f33a675e983.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8d824c26223cb12949e12f33a675e983.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=37321#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"3B\u6a21\u578b\u957f\u601d\u8003\u540e\u51fb\u8d2570B\uff01HuggingFace\u9006\u5411\u51fao1\u80cc\u540e\u6280\u672f\u7ec6\u8282\u5e76\u5f00\u6e90"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/37321","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=37321"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/37321\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=37321"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=37321"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=37321"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}