{"id":730,"date":"2024-07-15T14:06:46","date_gmt":"2024-07-15T06:06:46","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=730"},"modified":"2024-07-19T14:40:04","modified_gmt":"2024-07-19T06:40:04","slug":"202407-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb%e6%b1%87%e6%8a%a5ppt-22%e7%ba%a7%e5%88%98%e6%99%93-truthfulqa-measuring-how-models-mimic-human-falsehoods","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202407-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb%e6%b1%87%e6%8a%a5ppt-22%e7%ba%a7%e5%88%98%e6%99%93-truthfulqa-measuring-how-models-mimic-human-falsehoods\/","title":{"rendered":"202407 \u8bba\u6587\u7814\u8bfb-TruthfulQA Measuring How Models Mimic Human Falsehoods"},"content":{"rendered":"\n

\u4e00. \u80cc\u666f<\/h2>\n\n\n\n

\u4ee5GPT-3-175B\u4e3a\u4f8b\uff0c\u5927\u6a21\u578b\u7684\u8bad\u7ec3\u76ee\u6807\u5bb9\u6613\u6fc0\u52b1\u5176\u751f\u6210\u9ad8\u6a21\u4eff\u7684\u9519\u8bef\u7b54\u6848\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u6839\u636e\u7f29\u653e\u51c6\u5219\uff0c\u8fd9\u4e2a\u95ee\u9898\u65e0\u6cd5\u901a\u8fc7\u6269\u5927\u89c4\u6a21\u6765\u89e3\u51b3\u3002\u73b0\u6709\u7684\u95ee\u7b54\u57fa\u51c6\u4e5f\u65e0\u6cd5\u8986\u76d6\u5230\u6b64\u5e7b\u89c9\u3002<\/p>\n\n\n\n

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

\u4e8c. \u521b\u65b0<\/h2>\n\n\n\n
    \n
  1. \u63d0\u51fa\u4e86\u4e00\u4e2a\u65b0\u7684\u6d4b\u8bd5\u57fa\u51c6\uff1aTruthfulQA<\/a>\uff0c\u586b\u8865\u4e86\u73b0\u6709\u95ee\u7b54\u57fa\u51c6\u4e2d\u7f3a\u4e4f\u5bf9\u6a21\u578b\u771f\u5b9e\u6027\u8bc4\u4f30\u7684\u7a7a\u767d\u3002<\/li>\n\n\n\n
  2. \u5728\u96f6\u6837\u672c\u8bbe\u5b9a\u4e0b\u6d4b\u8bd5\u4e86\u4e0d\u540c\u89c4\u6a21\u5927\u5c0f\u548c\u4e0d\u540c\u63d0\u793a\u4e0b\u6d4b\u8bd5\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u4eba\u7c7b\u8bc4\u4f30\u8fd9\u4e9b\u6a21\u578b\u5728TruthfulQA\u4e0a\u7b54\u6848\u7684\u771f\u5b9e\u6027\u3002<\/li>\n\n\n\n
  3. \u5fae\u8c03GPT-3\u6a21\u578b\uff0c\u4f7f\u5176\u80fd\u591f\u81ea\u52a8\u8bc4\u4f30\u7b54\u6848\u7684\u771f\u5b9e\u6027\uff0c\u5e76\u53d6\u5f97\u4e8690-96%\u7684\u51c6\u786e\u7387\u3002<\/li>\n<\/ol>\n\n\n\n

    \u4e09. \u7b97\u6cd5\u4e0e\u5b9e\u73b0<\/h2>\n\n\n\n
      \n
    1. \u6784\u5efa\u6570\u636e\u96c6<\/strong>\n
        \n
      • \u4f5c\u8005\u4ece\u591a\u4e2a\u9886\u57df\uff08\u5982\u5065\u5eb7\u3001\u6cd5\u5f8b\u3001\u9634\u8c0b\u8bba\u7b49\uff09\u4e2d\u9009\u62e9\u95ee\u9898\uff0c\u5e76\u901a\u8fc7\u5b9e\u9a8c\u786e\u5b9a\u54ea\u4e9b\u95ee\u9898\u5bb9\u6613\u5bfc\u81f4\u6a21\u578b\u751f\u6210\u865a\u5047\u9648\u8ff0\u3002<\/li>\n\n\n\n
      • \u771f\u7406\u5224\u65ad\u6807\u51c6\uff1a\u7c7b\u4f3c\u4e8e\u7ef4\u57fa\u767e\u79d1\u548c\u79d1\u5b66\u6587\u6458\u6240\u4f7f\u7528\u7684\u6807\u51c6\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n
      • \u6a21\u578b\u8bc4\u4f30<\/strong>\n
          \n
        • \u5728\u4e0d\u540c\u6a21\u578b\u65cf\uff08\u5982GPT-3\u3001GPT-Neo\/J\u3001GPT-2\u548cT5\uff09\u4e2d\uff0c\u6700\u5927\u7684\u6a21\u578b\u901a\u5e38\u4e0d\u771f\u5b9e\uff0c\u8fd9\u79cd\u201c\u53cd\u5411\u7f29\u653e\u201d\u8d8b\u52bf\u4e0eNLP\u4e2d\u7684\u5927\u591a\u6570\u4efb\u52a1\u5f62\u6210\u5bf9\u6bd4\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n

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

            \n
          • \u8bc4\u4f30\u6807\u51c6\uff1a\n
              \n
            • \u771f\u5b9e\u6027\uff1a\u62d2\u7edd\u56de\u7b54\u95ee\u9898\u3001\u8868\u8fbe\u4e0d\u786e\u5b9a\u6216\u8005\u7ed9\u51fa\u771f\u5b9e\u4f46\u4e0d\u76f8\u5173\u7684\u7b54\u6848 <\/li>\n\n\n\n
            • \u4fe1\u606f\u4e30\u5bcc\u6027\uff1a\u7b54\u6848\u4e0e\u95ee\u9898\u6f5c\u5728\u7684\u76f8\u5173\u6027<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n

              3. \u81ea\u52a8\u5316\u8bc4\u4f30<\/strong><\/p>\n\n\n\n

              \uff081\uff09\u5f15\u5165\u4e86\u4eba\u7c7b\u8bc4\u4f30\u548c\u81ea\u52a8\u5316\u8bc4\u4f30<\/p>\n\n\n\n

              \uff082\uff09\u8bbe\u8ba1\u4e86\u65b0\u7684\u5ea6\u91cf\u6807\u51c6GPT-judge\uff0c\u8fd9\u662f\u4e00\u4e2a\u7ecf\u8fc7\u5fae\u8c03\u7684GPT-3-6.7B\u6a21\u578b<\/p>\n\n\n\n

              \uff083\uff09\u6d4b\u8bd5\u6a21\u578b\u5728TruthfulQA\u95ee\u9898\u7684\u540c\u4e49\u53e5\u4e0a\u7684\u8868\u73b0\uff0c\u9a8c\u8bc1\u5982\u679c\u4e00\u4e2a\u95ee\u9898\u5bfc\u81f4\u6a21\u4eff\u6027\u865a\u5047\u9648\u8ff0\uff0c\u90a3\u4e48\u5176\u540c\u4e49\u53e5\u662f\u5426\u4e5f\u4f1a\u5bfc\u81f4\u76f8\u540c\u7684\u865a\u5047\u9648\u8ff0\u3002\u603b\u4f53\u4e0a\uff0c\u6a21\u578b\u5728\u540c\u4e49\u53e5\u95ee\u9898\u4e0a\u7684\u771f\u5b9e\u6027\u8bc4\u5206\u6ca1\u6709\u663e\u8457\u53d8\u5316\u3002\u7279\u522b\u662f\uff0c\u6700\u5927\u7684GPT-3\u548cGPT-Neo\/J\u6a21\u578b\u5728\u540c\u4e49\u53e5\u4e0a\u7684\u8868\u73b0\u4ecd\u7136\u6bd4\u5b83\u4eec\u7684\u8f83\u5c0f\u7248\u672c\u5dee\u3002<\/p>\n\n\n\n

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

              4. \u6539\u8fdb<\/strong><\/p>\n\n\n\n

              \u6392\u9664\u6389\u6269\u5927\u6a21\u578b\u7684\u65b9\u6cd5\u540e\uff0c\u53ef\u4ee5\u5c1d\u8bd5<\/p>\n\n\n\n

              \uff081\uff09\u6269\u5927\u6a21\u578b\u89c4\u6a21\u518d\u7ed3\u5408\u5176\u4ed6\u6280\u672f\uff08\u5982\u63d0\u793a\u5de5\u7a0b\u6216\u5fae\u8c03\uff09\u3002<\/p>\n\n\n\n

              \uff082\uff09\u63d0\u793aGPT-3\u8bda\u5b9e\u4f5c\u7b54<\/p>\n\n\n\n

              \uff083\uff09\u9009\u62e9\u793a\u8303\u771f\u5b9e\u6027\u7684\u4f8b\u5b50\u8fdb\u884c\u5fae\u8c03\uff0c\u6216\u8005\u901a\u8fc7\u4ece\u4eba\u7c7b\u53cd\u9988\u4e2d\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8fdb\u884c\u5fae\u8c03<\/p>\n\n\n\n

              \"\"
              \u7ecf\u8fc7\u5fae\u8c03\u7684GPT-judge\u6a21\u578b\u80fd\u591f\u9884\u6d4b\u4eba\u7c7b\u5bf9\u771f\u5b9e\u6027\u7684\u8bc4\u4ef7\uff0c\u9a8c\u8bc1\u51c6\u786e\u7387\u4e3a90-96%<\/figcaption><\/figure>\n\n\n\n

              \u56db. \u603b\u7ed3\u4e0e\u5bf9\u9f50<\/h2>\n\n\n\n
                \n
              1. \u7ed3\u8bba\u4e00\uff1a<\/strong>\u4e3a\u4e86\u5f00\u53d1\u771f\u5b9e\u7684\u6a21\u578b\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u5957\u8861\u91cf\u771f\u5b9e\u6027\u7684\u57fa\u51c6\u548c\u5de5\u5177\uff1aTruthfulQA\u4e13\u6ce8\u4e8e\u6d4b\u91cf\u6a21\u4eff\u6027\u8c0e\u8a00\uff0c\u4e14\u5177\u6709\u901a\u7528\u6027\u548c\u4e13\u7528\u6027\u3002<\/li>\n\n\n\n
              2. \u7ed3\u8bba\u4e8c\uff1a<\/strong>\u5728\u96f6\u6837\u672c\u4e0a\u8bc4\u4f30\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u5982\u4eca\u5927\u6a21\u578b\u7684\u771f\u5b9e\u6027\u8fdc\u8fdc\u4e0d\u5982\u4eba\u7c7b\u4e14\u5177\u6709\u53cd\u5411\u7f29\u653e\u6027\uff0c\u56e0\u6b64\u63d0\u51fa\u5e76\u9a8c\u8bc1\u4e86\u6269\u5927\u6a21\u578b\u89c4\u6a21\u518d\u7ed3\u5408\u5176\u4ed6\u6280\u672f\uff08\u5982\u63d0\u793a\u5de5\u7a0b\u6216\u5fae\u8c03\uff09\u7b49\u6539\u8fdb\u65b9\u6cd5\u3002<\/li>\n\n\n\n
              3. \u542f\u53d1\u4e00\uff1a<\/strong>\u672c\u6587\u786e\u5b9a\u4e86\u540c\u4e49\u53e5\u6d4b\u8bd5\u6a21\u578b\u7684\u5fc5\u8981\u6027\uff0c\u5373\u6a21\u578b\u5728\u4e0d\u540c\u8bed\u6cd5\u7ed3\u6784\u6216\u8868\u8ff0\u65b9\u5f0f\u4e0b\u4f9d\u7136\u51c6\u786e\u3002\u6240\u4ee5\u53ef\u4ee5\u7528\u540c\u4e49\u53e5\u6765\u6d4b\u8bd5\u6a21\u578b\u7684\u9c81\u68d2\u6027\u3002<\/li>\n\n\n\n
              4. \u542f\u53d1\u4e8c\uff1a<\/strong>\u53ef\u4ee5\u5f00\u53d1\u65b0\u7684\u8bc4\u4f30\u6307\u6807\u6216\u57fa\u51c6\uff0c\u4e13\u95e8\u68c0\u6d4b\u6a21\u578b\u751f\u6210\u7684\u56de\u7b54\u662f\u5426\u771f\u5b9e\uff0c\u5e76\u533a\u522b\u4e8e\u8868\u9762\u4e0a\u770b\u4f3c\u6b63\u786e\u4f46\u5b9e\u9645\u9519\u8bef\u7684\u56de\u7b54\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"

                \u4e00. \u80cc\u666f \u4ee5GPT-3-175B\u4e3a\u4f8b\uff0c\u5927\u6a21\u578b\u7684\u8bad\u7ec3\u76ee\u6807\u5bb9\u6613\u6fc0\u52b1\u5176\u751f\u6210\u9ad8\u6a21\u4eff\u7684\u9519\u8bef\u7b54\u6848\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u6839\u636e\u7f29\u653e\u51c6 […]<\/p>\n","protected":false},"author":18,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-730","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/730","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/18"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=730"}],"version-history":[{"count":4,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/730\/revisions"}],"predecessor-version":[{"id":943,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/730\/revisions\/943"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=730"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=730"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=730"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}