{"id":16829,"date":"2023-11-18T18:27:07","date_gmt":"2023-11-18T10:27:07","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=16829"},"modified":"2023-11-18T18:28:29","modified_gmt":"2023-11-18T10:28:29","slug":"16829","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2023\/11\/18\/16829\/","title":{"rendered":"13B\u6a21\u578b\u5168\u65b9\u4f4d\u78be\u538bGPT-4\uff1f\u8fd9\u80cc\u540e\u6709\u4ec0\u4e48\u732b\u817b"},"content":{"rendered":"
\u5982\u679c\u4f60\u7ec6\u7ec6\u67e5\u770b\u56fe\u4e2d\u7684\u6a21\u578b\uff0c\u53d1\u73b0\u53ea\u8981\u5e26\u6709\u300crephraser\u300d\u8fd9\u4e2a\u5355\u8bcd\uff0c\u6a21\u578b\u6027\u80fd\u90fd\u6bd4\u8f83\u9ad8\u3002\u8fd9\u80cc\u540e\u5230\u5e95\u6709\u4f55\u732b\u817b\uff1f\u539f\u6765\u662f\u6570\u636e\u6c61\u67d3\u4e86\uff0c\u5373\u6d4b\u8bd5\u96c6\u4fe1\u606f\u5728\u8bad\u7ec3\u96c6\u4e2d\u906d\u5230\u6cc4\u6f0f\uff0c\u800c\u4e14\u8fd9\u79cd\u6c61\u67d3\u8fd8\u4e0d\u6613\u88ab\u68c0\u6d4b\u5230\u3002\u5c3d\u7ba1\u8fd9\u4e00\u95ee\u9898\u975e\u5e38\u5173\u952e\uff0c\u4f46\u7406\u89e3\u548c\u68c0\u6d4b\u6c61\u67d3\u4ecd\u7136\u662f\u4e00\u4e2a\u5f00\u653e\u4e14\u5177\u6709\u6311\u6218\u6027\u7684\u96be\u9898\u3002<\/div>\n
\n
\n

\u4e00\u4e2a\u53c2\u6570\u91cf\u4e3a 13B \u7684\u6a21\u578b\u7adf\u7136\u6253\u8d25\u4e86\u9876\u6d41 GPT-4\uff1f\u5c31\u50cf\u4e0b\u56fe\u6240\u5c55\u793a\u7684\uff0c\u5e76\u4e14\u4e3a\u4e86\u786e\u4fdd\u7ed3\u679c\u7684\u6709\u6548\u6027\uff0c\u8fd9\u9879\u6d4b\u8bd5\u8fd8\u9075\u5faa\u4e86 OpenAI \u7684\u6570\u636e\u53bb\u6c61\u65b9\u6cd5\uff0c\u66f4\u5173\u952e\u7684\u662f\u6ca1\u6709\u53d1\u73b0\u6570\u636e\u6c61\u67d3\u7684\u8bc1\u636e\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u5982\u679c\u4f60\u7ec6\u7ec6\u67e5\u770b\u56fe\u4e2d\u7684\u6a21\u578b\uff0c\u53d1\u73b0\u53ea\u8981\u5e26\u6709\u300crephraser\u300d\u8fd9\u4e2a\u5355\u8bcd\uff0c\u6a21\u578b\u6027\u80fd\u90fd\u6bd4\u8f83\u9ad8\u3002<\/p>\n

\u8fd9\u80cc\u540e\u5230\u5e95\u6709\u4f55\u732b\u817b\uff1f\u539f\u6765\u662f\u6570\u636e\u6c61\u67d3\u4e86\uff0c\u5373\u6d4b\u8bd5\u96c6\u4fe1\u606f\u5728\u8bad\u7ec3\u96c6\u4e2d\u906d\u5230\u6cc4\u6f0f\uff0c\u800c\u4e14\u8fd9\u79cd\u6c61\u67d3\u8fd8\u4e0d\u6613\u88ab\u68c0\u6d4b\u5230\u3002\u5c3d\u7ba1\u8fd9\u4e00\u95ee\u9898\u975e\u5e38\u5173\u952e\uff0c\u4f46\u7406\u89e3\u548c\u68c0\u6d4b\u6c61\u67d3\u4ecd\u7136\u662f\u4e00\u4e2a\u5f00\u653e\u4e14\u5177\u6709\u6311\u6218\u6027\u7684\u96be\u9898\u3002<\/p>\n

\u73b0\u9636\u6bb5\uff0c\u53bb\u6c61\u6700\u5e38\u7528\u7684\u65b9\u6cd5\u662f n-gram \u91cd\u53e0\u548c\u5d4c\u5165\u76f8\u4f3c\u6027\u641c\u7d22\uff1aN-gram \u91cd\u53e0\u4f9d\u8d56\u4e8e\u5b57\u7b26\u4e32\u5339\u914d\u6765\u68c0\u6d4b\u6c61\u67d3\uff0c\u662f GPT-4\u3001PaLM \u548c Llama-2 \u7b49\u6a21\u578b\u5e38\u7528\u65b9\u6cd5\uff1b\u5d4c\u5165\u76f8\u4f3c\u6027\u641c\u7d22\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\uff08\u4f8b\u5982 BERT\uff09\u7684\u5d4c\u5165\u6765\u67e5\u627e\u76f8\u4f3c\u4e14\u53ef\u80fd\u53d7\u5230\u6c61\u67d3\u7684\u793a\u4f8b\u3002<\/p>\n

\u7136\u800c\uff0c\u6765\u81ea UC \u4f2f\u514b\u5229\u3001\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66\u7684\u7814\u7a76\u8868\u660e\u6d4b\u8bd5\u6570\u636e\u7684\u7b80\u5355\u53d8\u5316\uff08\u4f8b\u5982\uff0c\u6539\u5199\u3001\u7ffb\u8bd1\uff09\u5c31\u53ef\u4ee5\u8f7b\u677e\u7ed5\u8fc7\u73b0\u6709\u7684\u68c0\u6d4b\u65b9\u6cd5\u3002\u4ed6\u4eec\u5e76\u5c06\u6d4b\u8bd5\u7528\u4f8b\u7684\u6b64\u7c7b\u53d8\u4f53\u79f0\u4e3a\u300c\u6539\u5199\u6837\u672c\uff08Rephrased Samples\uff09\u300d\u3002<\/p>\n

\u4e0b\u9762\u6f14\u793a\u4e86 MMLU \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6539\u5199\u6837\u672c\u3002\u7ed3\u679c\u8bc1\u660e\uff0c\u5982\u679c\u8bad\u7ec3\u96c6\u4e2d\u5305\u542b\u6b64\u7c7b\u6837\u672c\uff0c13B \u6a21\u578b\u53ef\u4ee5\u8fbe\u5230\u6781\u9ad8\u7684\u6027\u80fd (MMLU 85.9)\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u73b0\u6709\u7684\u68c0\u6d4b\u65b9\u6cd5\uff08\u4f8b\u5982\uff0cn-gram \u91cd\u53e0\u3001\u5d4c\u5165\u76f8\u4f3c\u6027\uff09\u65e0\u6cd5\u68c0\u6d4b\u5230\u8fd9\u79cd\u6c61\u67d3\u3002\u6bd4\u5982\u5d4c\u5165\u76f8\u4f3c\u6027\u65b9\u6cd5\u5f88\u96be\u5c06\u6539\u5199\u7684\u95ee\u9898\u4e0e\u540c\u4e00\u4e3b\u9898\u4e2d\u7684\u5176\u4ed6\u95ee\u9898\u533a\u5206\u5f00\u6765\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u901a\u8fc7\u7c7b\u4f3c\u7684\u6539\u5199\u6280\u672f\uff0c\u672c\u6587\u5728\u5e7f\u6cdb\u4f7f\u7528\u7684\u7f16\u7801\u548c\u6570\u5b66\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u89c2\u5bdf\u5230\u4e00\u81f4\u7684\u7ed3\u679c\uff0c\u4f8b\u5982 HumanEval \u548c GSM-8K\uff08\u5982\u6587\u7ae0\u5f00\u5934\u56fe\u4e2d\u6240\u793a\uff09\u3002\u56e0\u6b64\uff0c\u80fd\u591f\u68c0\u6d4b\u6b64\u7c7b\u6539\u5199\u6837\u672c\u53d8\u5f97\u81f3\u5173\u91cd\u8981\u3002<\/p>\n

\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u770b\u770b\u8fd9\u9879\u7814\u7a76\u662f\u5982\u4f55\u8fdb\u884c\u7684\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

    \n
  • \u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2311.04850.pdf<\/li>\n
  • \u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/lm-sys\/llm-decontaminator#detect<\/li>\n<\/ul>\n

    \u8bba\u6587\u4ecb\u7ecd<\/h4>\n

    \u6587\u4e2d\u8868\u793a\uff0c\u5927\u6a21\u578b\uff08LLM\uff09\u5728\u5feb\u901f\u53d1\u5c55\u7684\u540c\u65f6\uff0c\u5173\u4e8e\u6d4b\u8bd5\u96c6\u6c61\u67d3\u7684\u95ee\u9898\u88ab\u8d8a\u6765\u8d8a\u591a\u7684\u91cd\u89c6\u8d77\u6765\uff0c\u5f88\u591a\u4eba\u5bf9\u516c\u5171\u57fa\u51c6\u7684\u53ef\u4fe1\u5ea6\u8868\u793a\u62c5\u5fe7\u3002<\/p>\n

    \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6709\u4e9b\u4eba\u91c7\u7528\u4f20\u7edf\u7684\u53bb\u6c61\u65b9\u6cd5\u5982\u5b57\u7b26\u4e32\u5339\u914d\uff08\u4f8b\u5982\uff0cn-gram \u91cd\u53e0\uff09\u6765\u5220\u9664\u57fa\u51c6\u6570\u636e\uff0c\u4f46\u8fd9\u4e9b\u64cd\u4f5c\u8fd8\u8fdc\u8fdc\u4e0d\u591f\uff0c\u56e0\u4e3a\u5bf9\u6d4b\u8bd5\u6570\u636e\u8fdb\u884c\u4e00\u4e9b\u7b80\u5355\u7684\u66f4\u6539\uff08\u4f8b\u5982\uff0c\u6539\u5199\u3001\u7ffb\u8bd1\uff09\u5c31\u53ef\u4ee5\u8f7b\u677e\u7ed5\u8fc7\u8fd9\u4e9b\u51c0\u5316\u63aa\u65bd\u3002<\/p>\n

    \u66f4\u91cd\u8981\u7684\u662f\uff0c\u5982\u679c\u4e0d\u6d88\u9664\u6d4b\u8bd5\u6570\u636e\u7684\u8fd9\u79cd\u66f4\u6539\uff0c13B \u6a21\u578b\u5f88\u5bb9\u6613\u8fc7\u5ea6\u62df\u5408\u6d4b\u8bd5\u57fa\u51c6\u5e76\u5b9e\u73b0\u4e0e GPT-4 \u76f8\u5f53\u7684\u6027\u80fd\u3002\u4ed6\u4eec\u5728 MMLU\u3001GSK8k \u548c HumanEval \u7b49\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u9a8c\u8bc1\u4e86\u8fd9\u4e9b\u89c2\u5bdf\u7ed3\u679c\u3002<\/p>\n

    \u540c\u65f6\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u65e5\u76ca\u589e\u957f\u7684\u98ce\u9669\uff0c\u672c\u6587\u8fd8\u63d0\u51fa\u4e86\u4e00\u79cd\u66f4\u4e3a\u5f3a\u5927\u7684\u57fa\u4e8e LLM \u7684\u53bb\u6c61\u65b9\u6cd5 LLM decontaminator\uff0c\u5e76\u5c06\u5176\u5e94\u7528\u4e8e\u6d41\u884c\u7684\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u6570\u636e\u96c6\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u672c\u6587\u63d0\u51fa\u7684 LLM \u65b9\u6cd5\u5728\u5220\u9664\u6539\u5199\u6837\u672c\u65b9\u9762\u660e\u663e\u4f18\u4e8e\u73b0\u6709\u65b9\u6cd5\u3002<\/p>\n

    \u8fd9\u4e00\u505a\u6cd5\u4e5f\u63ed\u9732\u4e86\u4e00\u4e9b\u5148\u524d\u672a\u77e5\u7684\u6d4b\u8bd5\u91cd\u53e0\uff08test overlap\uff09\u3002\u4f8b\u5982\uff0c\u5728 RedPajamaData-1T \u548c StarCoder-Data \u7b49\u9884\u8bad\u7ec3\u96c6\u4e2d\uff0c\u672c\u6587\u53d1\u73b0 HumanEval \u57fa\u51c6\u6709 8-18% \u91cd\u53e0\u3002\u6b64\u5916\uff0c\u672c\u6587\u8fd8\u5728 GPT-3.5\/4 \u751f\u6210\u7684\u5408\u6210\u6570\u636e\u96c6\u4e2d\u53d1\u73b0\u4e86\u8fd9\u79cd\u6c61\u67d3\uff0c\u8fd9\u4e5f\u8bf4\u660e\u4e86\u5728 AI \u9886\u57df\u5b58\u5728\u6f5c\u5728\u7684\u610f\u5916\u6c61\u67d3\u98ce\u9669\u3002<\/p>\n

    \u672c\u6587\u5e0c\u671b\uff0c\u793e\u533a\u5728\u4f7f\u7528\u516c\u5171\u57fa\u51c6\u65f6\u91c7\u53d6\u66f4\u5f3a\u6709\u529b\u7684\u51c0\u5316\u65b9\u6cd5\uff0c\u5e76\u547c\u5401\u793e\u533a\u79ef\u6781\u5f00\u53d1\u65b0\u7684\u4e00\u6b21\u6027\u6d4b\u8bd5\uff08one-time exams\uff09\u6848\u4f8b\u6765\u51c6\u786e\u8bc4\u4f30\u6a21\u578b\u3002<\/p>\n

    \u6539\u5199\u6837\u672c<\/strong><\/p>\n

    \u672c\u6587\u7684\u76ee\u6807\u662f\u8c03\u67e5\u8bad\u7ec3\u96c6\u4e2d\u5305\u542b\u6d4b\u8bd5\u96c6\u7684\u7b80\u5355\u53d8\u5316\u662f\u5426\u4f1a\u5f71\u54cd\u6700\u7ec8\u7684\u57fa\u51c6\u6027\u80fd\uff0c\u5e76\u5c06\u6d4b\u8bd5\u7528\u4f8b\u7684\u8fd9\u79cd\u53d8\u5316\u79f0\u4e3a\u300c\u6539\u5199\u6837\u672c\u300d\u3002\u5b9e\u9a8c\u4e2d\u8003\u8651\u4e86\u57fa\u51c6\u7684\u5404\u4e2a\u9886\u57df\uff0c\u5305\u62ec\u6570\u5b66\u3001\u77e5\u8bc6\u548c\u7f16\u7801\u3002\u793a\u4f8b 1 \u662f\u6765\u81ea GSM-8k \u7684\u6539\u5199\u6837\u672c\uff0c\u5176\u4e2d\u6709 10-gram \u91cd\u53e0\u65e0\u6cd5\u68c0\u6d4b\u5230\uff0c\u4fee\u6539\u540e\u548c\u539f\u59cb\u6587\u672c\u4fdd\u6301\u76f8\u540c\u7684\u8bed\u4e49\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u57fa\u51c6\u6c61\u67d3\u5177\u6709\u4e0d\u540c\u7684\u5f62\u5f0f\uff0c\u56e0\u6b64\u6539\u5199\u6280\u672f\u5b58\u5728\u4e00\u4e9b\u7ec6\u5fae\u7684\u5dee\u5f02\u3002\u5bf9\u4e8e\u57fa\u4e8e\u6587\u672c\u7684\u57fa\u51c6\uff0c\u672c\u6587\u5728\u4e0d\u6539\u53d8\u8bed\u4e49\u7684\u60c5\u51b5\u4e0b\u6539\u5199\u6d4b\u8bd5\u7528\u4f8b\uff0c\u4f8b\u5982\u901a\u8fc7\u91cd\u65b0\u6392\u5217\u8bcd\u5e8f\u6216\u7528\u540c\u4e49\u672f\u8bed\u66ff\u6362\uff1b\u5bf9\u4e8e\u57fa\u4e8e\u4ee3\u7801\u7684\u57fa\u51c6\u6d4b\u8bd5\uff0c\u672c\u6587\u6539\u53d8\u7f16\u7801\u98ce\u683c\u3001\u547d\u540d\u65b9\u5f0f\u7b49\u3002<\/p>\n

    \u5982\u4e0b\u6240\u793a\uff0c\u7b97\u6cd5 1 \u4e2d\u9488\u5bf9\u7ed9\u5b9a\u7684\u6d4b\u8bd5\u96c6\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u7684\u7b97\u6cd5\u3002\u8be5\u65b9\u6cd5\u53ef\u4ee5\u5e2e\u52a9\u6d4b\u8bd5\u6837\u672c\u9003\u907f\u68c0\u6d4b\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u63a5\u4e0b\u6765\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u6c61\u67d3\u68c0\u6d4b\u65b9\u6cd5\uff0c\u53ef\u4ee5\u51c6\u786e\u5730\u4ece\u76f8\u5bf9\u4e8e\u57fa\u51c6\u7684\u6570\u636e\u96c6\u4e2d\u5220\u9664\u6539\u5199\u6837\u672c\u3002<\/p>\n

    \u5177\u4f53\u800c\u8a00\uff0c\u672c\u6587\u5f15\u5165\u4e86 LLM decontaminator\u3002\u9996\u5148\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u6d4b\u8bd5\u7528\u4f8b\uff0c\u5b83\u4f7f\u7528\u5d4c\u5165\u76f8\u4f3c\u5ea6\u641c\u7d22\u6765\u8bc6\u522b\u5177\u6709\u6700\u9ad8\u76f8\u4f3c\u5ea6\u7684 top-k \u8bad\u7ec3\u9879\uff0c\u4e4b\u540e\u901a\u8fc7 LLM\uff08\u4f8b\u5982 GPT-4\uff09\u8bc4\u4f30\u6bcf\u4e00\u5bf9\u662f\u5426\u76f8\u540c\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u52a9\u4e8e\u786e\u5b9a\u6570\u636e\u96c6\u4e2d\u6709\u591a\u5c11\u6539\u5199\u6837\u672c\u3002<\/p>\n

    \u56fe 4 \u5c55\u793a\u4e86\u4e0d\u540c\u6c61\u67d3\u4ee5\u53ca\u4e0d\u540c\u68c0\u6d4b\u65b9\u6cd5\u7684\u7ef4\u6069\u56fe\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u5b9e\u9a8c<\/h4>\n

    \u5728\u7b2c 5.1 \u8282\u4e2d\uff0c\u5b9e\u9a8c\u8bc1\u660e\u4e86\u5728\u6539\u5199\u6837\u672c\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u53ef\u4ee5\u53d6\u5f97\u663e\u7740\u7684\u9ad8\u5206\uff0c\u5728\u4e09\u4e2a\u5e7f\u6cdb\u4f7f\u7528\u7684\u57fa\u51c6\uff08MMLU\u3001HumanEval \u548c GSM-8k\uff09\u4e2d\u5b9e\u73b0\u4e0e GPT-4 \u76f8\u5f53\u7684\u6027\u80fd\uff0c\u8fd9\u8868\u660e\u6539\u5199\u6837\u672c\u5e94\u88ab\u89c6\u4e3a\u6c61\u67d3\uff0c\u5e94\u4ece\u8bad\u7ec3\u6570\u636e\u4e2d\u5220\u9664\u3002\u5728\u7b2c 5.2 \u8282\u4e2d\uff0c\u672c\u6587\u6839\u636e MMLU\/HumanEval \u4e2d\u6539\u5199\u6837\u672c\u8bc4\u4f30\u4e0d\u540c\u7684\u6c61\u67d3\u68c0\u6d4b\u65b9\u6cd5\u3002\u5728\u7b2c 5.3 \u8282\u4e2d\uff0c\u672c\u6587\u5c06 LLM decontaminator \u5e94\u7528\u4e8e\u5e7f\u6cdb\u4f7f\u7528\u7684\u8bad\u7ec3\u96c6\u5e76\u53d1\u73b0\u4ee5\u524d\u672a\u77e5\u7684\u6c61\u67d3\u3002<\/p>\n

    \u63a5\u4e0b\u6765\u6211\u4eec\u770b\u770b\u4e00\u4e9b\u4e3b\u8981\u7ed3\u679c\u3002<\/p>\n

    \u6539\u5199\u6837\u672c\u6c61\u67d3\u57fa\u51c6<\/strong><\/p>\n

    \u5982\u8868 2 \u6240\u793a\uff0c\u5728\u6539\u5199\u6837\u672c\u4e0a\u8bad\u7ec3\u7684 Llama-2 7B \u548c 13B \u5728 MMLU \u4e0a\u53d6\u5f97\u663e\u7740\u7684\u9ad8\u5206\uff0c\u4ece 45.3 \u5230 88.5\u3002\u8fd9\u8868\u660e\u7ecf\u8fc7\u6539\u5199\u7684\u6837\u672c\u53ef\u80fd\u4f1a\u4e25\u91cd\u626d\u66f2\u57fa\u51c6\u6570\u636e\uff0c\u5e94\u88ab\u89c6\u4e3a\u6c61\u67d3\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u672c\u6587\u8fd8\u5bf9 HumanEval \u6d4b\u8bd5\u96c6\u8fdb\u884c\u4e86\u6539\u5199\uff0c\u5e76\u5c06\u5176\u7ffb\u8bd1\u6210\u4e94\u79cd\u7f16\u7a0b\u8bed\u8a00\uff1aC\u3001JavaScript\u3001Rust\u3001Go \u548c Java\u3002\u7ed3\u679c\u663e\u793a\uff0c\u5728\u6539\u5199\u6837\u672c\u4e0a\u8bad\u7ec3\u7684 CodeLlama 7B \u548c 13B \u5728 HumanEval \u4e0a\u53ef\u4ee5\u53d6\u5f97\u6781\u9ad8\u7684\u5206\u6570\uff0c\u5206\u522b\u4ece 32.9 \u5230 67.7 \u4ee5\u53ca 36.0 \u5230 81.1\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cGPT-4 \u5728 HumanEval \u4e0a\u53ea\u80fd\u8fbe\u5230 67.0\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u4e0b\u8868 4 \u53d6\u5f97\u4e86\u540c\u6837\u7684\u6548\u679c\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u5bf9\u68c0\u6d4b\u6c61\u67d3\u65b9\u6cd5\u7684\u8bc4\u4f30<\/strong><\/p>\n

    \u5982\u8868 5 \u6240\u793a\uff0c\u9664 LLM decontaminator \u5916\uff0c\u6240\u6709\u5176\u4ed6\u68c0\u6d4b\u65b9\u6cd5\u90fd\u4f1a\u5f15\u5165\u4e00\u4e9b\u8bef\u62a5\u3002\u6539\u5199\u548c\u7ffb\u8bd1\u7684\u6837\u672c\u90fd\u4e0d\u4f1a\u88ab n-gram \u91cd\u53e0\u68c0\u6d4b\u5230\u3002\u4f7f\u7528 multi-qa BERT\uff0c\u5d4c\u5165\u76f8\u4f3c\u6027\u641c\u7d22\u88ab\u8bc1\u660e\u5bf9\u7ffb\u8bd1\u6837\u672c\u5b8c\u5168\u65e0\u6548\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u6570\u636e\u96c6\u6c61\u67d3\u60c5\u51b5<\/strong><\/p>\n

    \u8868 7 \u663e\u793a\u4e86\u6bcf\u4e2a\u8bad\u7ec3\u6570\u636e\u96c6\u4e2d\u4e0d\u540c\u57fa\u51c6\u7684\u6570\u636e\u6c61\u67d3\u767e\u5206\u6bd4\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    LLM decontaminator \u63ed\u793a\u4e86 79 \u4e2a\u81ea\u6539\u5199\u6837\u672c\u7684\u5b9e\u4f8b\uff0c\u5360 MATH \u6d4b\u8bd5\u96c6\u7684 1.58%\u3002\u793a\u4f8b 5 \u662f MATH \u8bad\u7ec3\u6570\u636e\u4e2d MATH \u6d4b\u8bd5\u7684\u6539\u5199\u793a\u4f8b\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u4e86\u89e3\u66f4\u591a\u5185\u5bb9\uff0c\u8bf7\u67e5\u770b\u539f\u8bba\u6587\u3002<\/p>\n<\/div>\n<\/div>\n

    <\/div>\n

    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

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

    \u5982\u679c\u4f60\u7ec6\u7ec6\u67e5\u770b\u56fe\u4e2d\u7684\u6a21\u578b\uff0c\u53d1\u73b0\u53ea\u8981\u5e26\u6709\u300crephraser\u300d\u8fd9\u4e2a\u5355\u8bcd\uff0c\u6a21\u578b\u6027\u80fd\u90fd\u6bd4\u8f83\u9ad8\u3002\u8fd9\u80cc\u540e\u5230\u5e95\u6709\u4f55\u732b\u817b\uff1f\u539f […]<\/p>\n

    <\/div>\n

    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

    <\/div>\n","protected":false},"author":56,"featured_media":16830,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,80,1],"tags":[124,820],"class_list":["post-16829","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-80","category-uncategorized","tag-ai","tag-820"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/16829","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=16829"}],"version-history":[{"count":3,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/16829\/revisions"}],"predecessor-version":[{"id":16856,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/16829\/revisions\/16856"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/16830"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=16829"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=16829"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=16829"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}