{"id":60877,"date":"2021-09-12T22:12:00","date_gmt":"2021-09-12T13:12:00","guid":{"rendered":"https:\/\/smilegate.ai\/?p=60877"},"modified":"2021-10-01T18:24:39","modified_gmt":"2021-10-01T09:24:39","slug":"instruction-tuning-flan","status":"publish","type":"post","link":"https:\/\/smilegate.ai\/cn\/2021\/09\/12\/instruction-tuning-flan\/","title":{"rendered":"\u6307\u4ee4\u8c03\u4f18 \u2013 FLAN"},"content":{"rendered":"

[\u878d\u5408\u7814\u7a76\u7ec4Hongmae Shim]<\/p>\n\n\n\n

\u5982\u679c\u4f60\u9009\u51fa 2020 \u5e74 NLP \u9886\u57df\u7684\u524d 10 \u4e2a\u5173\u952e\u8bcd\uff0c GPT-3 (\u8bed\u8a00\u6a21\u578b\u662f\u5c11\u6570\u5b66\u4e60\u8005\uff09<\/strong><\/strong> \u5f53\u7136\uff0c\u5b83\u4f1a\u5728\u6392\u540d\u4e2d\u3002\u65f6\u81f3\u4eca\u65e5\uff0cGPT-3 \u5e9e\u5927\u7684\u53c2\u6570\u91cf\u548c\u5353\u8d8a\u7684\u6027\u80fd\u4ecd\u7136\u5728 NLP \u9886\u57df\u5185\u5916\u8d8a\u6765\u8d8a\u53d7\u6b22\u8fce\u3002\u7136\u800c\uff0c\u4f5c\u4e3a\u4e00\u540d NLP \u7814\u7a76\u4eba\u5458\uff0cGPT-3 \u5bf9 state-of-the-art \u7814\u7a76\u7684\u6700\u5927\u8d21\u732e\u662f\u5728\u4e00\u822c\u5de5\u4f5c\u4e2d\uff08\u5c24\u5176\u662f\u96f6\u6837\u672c\u548c\u5c11\u6837\u672c\uff09\u3002 \u5373\u65f6\u8c03\u4f18<\/strong><\/em> \u6211\u8ba4\u4e3a\u8fd9\u8bc1\u660e\u4e86\u8be5\u6280\u672f\u7684\u9002\u7528\u6027\u3002\u5728 GPT-3 \u4e4b\u524d\uff0cprompt-tuning \u4e3b\u8981\u7528\u4e8e\u63a2\u7d22\u8bed\u8a00\u6a21\u578b\u4e2d\u5d4c\u5165\uff08latent\uff09\u7684\u77e5\u8bc6\uff0c\u8fd1\u4e24\u5e74\u5927\u91cf\u76f8\u5173\u8bba\u6587\u4f5c\u4e3a\u4e00\u4e2a\u975e\u5e38\u70ed\u95e8\u7684\u5173\u952e\u8bcd\u6d8c\u51fa\u3002<\/p>\n\n\n\n

Prompt-tuning \u548c GPT-3 \u7684\u5171\u540c\u6210\u5c31\u5728 NLP \u7684\u53d1\u5c55\u53f2\u4e0a\u662f\u4e0d\u53ef\u5ffd\u89c6\u7684\u3002\u57fa\u4e8e\u63d0\u793a\u8c03\u4f18\u7684 GPT-3 \u5728\u6267\u884c\u5404\u79cd\u7c7b\u578b\u7684\u4efb\u52a1\u65f6\u8868\u73b0\u51fa\u826f\u597d\u7684\u6027\u80fd\uff0c\u4f46\u4ed4\u7ec6\u60f3\u60f3\uff0cGPT-3 \u4e0d\u662f\u5177\u6709\u66f4\u597d\u7684\u96f6\u6837\u672c\u548c\u5c11\u6837\u672c\u5b66\u4e60\u80fd\u529b\u6765\u6267\u884c\u8fd9\u4e9b\u60ca\u4eba\u7684\u4efb\u52a1\u5417\uff1f\u63d0\u793a\u8c03\u4f18\u662f\u4f7f\u7528 GPT-3 \u7684\u6700\u4f73\u65b9\u5f0f\u5417\uff1f<\/p>\n\n\n\n

\u6700\u8fd1\uff0c\u8c37\u6b4c\u7814\u7a76\u4eba\u5458\u5f00\u53d1\u4e86\u4e00\u79cd\u6307\u4ee4\u8c03\u4f18\u65b9\u6cd5\uff0c\u5728\u4f7f\u7528\u6bd4 GPT-3 (175B) \u66f4\u5c11\u7684\u53c2\u6570 (137B) \u7684 25 \u9879\u4efb\u52a1\u4e2d\uff0c\u6709 19 \u9879\u7684\u6027\u80fd\u660e\u663e\u4f18\u4e8e GPT-3\u3002 \u6c34\u679c\u9985\u997c <\/em><\/strong><\/em><\/strong>(F<\/strong>\u8c03\u548c \u5c40\u57df\u7f51<\/strong>guage \u6a21\u578b\u662f\u96f6\u6837\u672c\u5b66\u4e60\u8005<\/a><\/em>)<\/em><\/strong>\u901a\u8fc7\u6697\u793a GPT-3 \u53ef\u4ee5\u53d8\u5f97\u66f4\u5f3a\u5927\u3002<\/p>\n\n\n\n

GPT-3<\/strong> ( \u5c40\u57df\u7f51<\/strong>guage \u6a21\u578b\u662f\u96f6\u6837\u672c\u5b66\u4e60\u8005<\/a>\uff09 \u76f8\u6bd4 \u6c34\u679c\u9985\u997c<\/em><\/strong>\u8ba9\u6211\u4eec\u4e0e\u4f17\u4e0d\u540c\u7684\u662f\u6211\u4eec\u7684\u5fae\u8c03\u3002 \u6c34\u679c\u9985\u997c<\/em><\/strong>\u6838\u5fc3\u601d\u60f3\u662f\u5fae\u8c03\u5404\u79cdNLP\u4efb\u52a1\uff0c\u901a\u8fc7\u5c06\u5b83\u4eec\u8f6c\u5316\u4e3a\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\uff08\u4e00\u79cd\u4efb\u52a1\u6307\u4ee4\u6216\u6307\u4ee4\uff09\u6765\u89e3\u51b3\u8fd9\u4e9b\u4efb\u52a1\u3002 \uff08\u53c2\u8003\u4e0b\u9762[\u56fe1]\u4e2d\u7684\uff08C\uff09\uff09<\/p>\n\n\n\n

\"\"
\u3010\u56fe1\u3011Instruction Tuning\u4e0ePretrain-finetune\u548cPrompting\u5bf9\u6bd4<\/figcaption><\/figure>\n\n\n\n

\u4e3a\u4e86\u66f4\u8be6\u7ec6\u5730\u89e3\u91ca\uff0cFLAN \u9996\u5148\u5bf9\u9884\u8bad\u7ec3\u7684 LM \u8fdb\u884c\u5fae\u8c03\uff0c\u4f7f\u5176\u53ef\u4ee5\u6267\u884c\u8bb8\u591a\u4e0d\u540c\u7684 NLP \u4efb\u52a1\uff0c\u5305\u62ec\u7ffb\u8bd1\u3001\u5e38\u8bc6\u63a8\u7406\u3001\u60c5\u611f\u5206\u7c7b\u7b49\u3002\u4f8b\u5982\uff0c\u5982\u4e0b\u9762\u7684 [\u56fe 2] \u6240\u793a\uff0c\u5bf9\u4e8e\u7ffb\u8bd1\u4efb\u52a1\u201c\u5c06\u8fd9\u53e5\u8bdd\u7ffb\u8bd1\u6210\u897f\u73ed\u7259\u8bed\u201d\uff0c\u5bf9\u4e8e\u60c5\u611f\u5206\u7c7b\u4efb\u52a1\uff0c\u4f7f\u7528\u547d\u4ee4\/\u6307\u4ee4\u201c\u8be5\u7535\u5f71\u8bc4\u8bba\u7684\u60c5\u611f\u662f\u6b63\u9762\u7684\u8fd8\u662f\u8d1f\u9762\u7684\uff1f\u201d\u5fae\u8c03\u5b8c\u6210\u540e\uff0c\u6a21\u578b\u53ef\u4ee5\u4f7f\u7528\u5305\u542b\u8fd9\u4e9b\u547d\u4ee4\/\u6307\u5357\u7684\u4fe1\u606f\u6267\u884c\u5404\u79cd\u4efb\u52a1\uff0c\u5c06\u73b0\u6709\u77e5\u8bc6\u5e94\u7528\u4e8e\u547d\u4ee4\u201c\u524d\u63d0\u662f\u5426\u5305\u542b\u5047\u8bbe\uff1f\u201d\u60a8\u53ef\u4ee5\u66f4\u597d\u5730\u4f7f\u7528\u5b83\u6765\u56de\u7b54\u5b83\u3002<\/p>\n\n\n\n

\"\"
\u3010\u56fe2\u3011NLP\u4efb\u52a1\u548c\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\u793a\u4f8b<\/figcaption><\/figure>\n\n\n\n

\u5728\u8bba\u6587\u4e2d\uff0c\u4f5c\u8005\u53d1\u73b0 FLAN \u751a\u81f3\u53ef\u4ee5\u5728\u5b66\u4e60\u4e86\u7f51\u9875\u3001\u7f16\u7a0b\u8bed\u8a00\u3001\u5bf9\u8bdd\u548c\u7ef4\u57fa\u767e\u79d1\u53e5\u5b50\u540e\u6ca1\u6709\u660e\u786e\u5b66\u4e60\u7684\u4efb\u52a1\u4e0a\u5b66\u4e60\u64cd\u4f5c\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6307\u4ee4\u8c03\u4f18\u53ef\u4ee5\u901a\u8fc7\u6559\u6388\u6a21\u578b\u5982\u4f55\u6267\u884c\u8868\u793a\u4e3a\u4e00\u79cd\u547d\u4ee4\/\u6307\u4ee4\u7684 NLP \u4efb\u52a1\u6765\u63d0\u9ad8\u5904\u7406\u548c\u7406\u89e3\u81ea\u7136\u8bed\u8a00\u7684\u80fd\u529b\u3002\u8fd9\u610f\u5473\u7740\u6211\u4eec\u53ef\u4ee5\u90e8\u5206\u7406\u89e3\u81ea\u7136\u8bed\u8a00\u7684\u771f\u5b9e\u610f\u56fe\u3002<\/p>\n\n\n\n

\u5728 FLAN \u7684\u8bba\u6587\u4e2d\uff0c\u901a\u8fc7\u9009\u53d6 12 \u4e2a\u7c7b\u522b\u5171 62 \u4e2a\u5e38\u89c1\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u751f\u6210\u4efb\u52a1\u76f8\u5173\u6570\u636e\u8fdb\u884c\u4e86\u8c03\u4f18\u5b9e\u9a8c\u3002 \uff08\u53c2\u8003\u3010\u56fe3\u3011\uff09<\/p>\n\n\n\n

\"\"
\u3010\u56fe3\u3011\u4efb\u52a1\u96c6\u7fa4\u5217\u8868\uff08\u84dd\u8272\u2014\u2014NLU\u4efb\u52a1\uff0c\u7eff\u677e\u77f3\u2014\u2014NLG\u4efb\u52a1\uff09 <\/figcaption><\/figure>\n\n\n\n

\u4f5c\u8005\u4f7f\u7528\u5927\u5c0f\u4e3a 137B \u7684\u81ea\u56de\u5f52\u8bed\u8a00\u6a21\u578b (Base LM) \u4f5c\u4e3a\u57fa\u7840\u8bed\u8a00\u6a21\u578b\u3002\u6307\u4ee4\u8c03\u6574\u7ba1\u9053\u6df7\u5408\u4e86 60 \u591a\u4e2a nlp \u4efb\u52a1\u7684\u6240\u6709\u6570\u636e\u96c6\uff0c\u5e76\u4ece\u6bcf\u4e2a\u6570\u636e\u96c6\u4e2d\u968f\u673a\u91c7\u6837\u3002\u6bcf\u4e2a\u6570\u636e\u96c6\u4e2d\u7684\u6837\u672c\u6570\u91cf\u5dee\u5f02\u5f88\u5927\uff0c\u6709\u4e9b\u6570\u636e\u96c6\u7684\u8bad\u7ec3\u6837\u672c\uff08\u4f8b\u5982\u7ffb\u8bd1\uff09\u8d85\u8fc71000\u4e07\u4e2a\uff0c\u5c06\u6bcf\u4e2a\u6570\u636e\u96c6\u4e2d\u7684\u6700\u7ec8\u8bad\u7ec3\u6837\u672c\u6570\u91cf\u9650\u5236\u572830,000\u4e2a\u3002\u5728\u5b9e\u9a8c\u4e2d\uff0cT5-11B\u548cGPT-3\u88ab\u7528\u4f5c\u53c2\u8003\u6a21\u578b\u3002<\/p>\n\n\n\n

\u4f5c\u4e3a\u5b9e\u9a8c\u7684\u7ed3\u679c\uff0cFLAN \u5728\u81ea\u7136\u8bed\u8a00\u63a8\u7406\u4efb\u52a1\u548c QA \u4efb\u52a1\u4e2d\u7684\u96f6\u6837\u672c\u573a\u666f\u4e2d\u7684\u8868\u73b0\u5df2\u7ecf\u4f18\u4e8e\u5c11\u6837\u672c GPT-3\uff0c\u5e76\u4e14\u5728\u8bb8\u591a\u4efb\u52a1\u4e2d\u53d6\u5f97\u4e86\u4e0e\u76d1\u7763\u6a21\u578b\u76f8\u4f3c\u7684\u6027\u80fd\uff08[\u56fe 4] , [\u4e0b\u56fe 5] ] \u6ce8.) \u8bba\u6587\u4e2d\u8fd8\u5305\u542b\u4e86\u5176\u4ed6\u5404\u79cd\u4efb\u52a1\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u5982\u6709\u5fc5\u8981\uff0c\u8bf7\u81ea\u884c\u68c0\u67e5\u8bba\u6587\u3002<\/p>\n\n\n\n

\"\"
[\u56fe 4] \u81ea\u7136\u8bed\u8a00\u63a8\u7406\u548c QA \u4efb\u52a1\u7684\u5b9e\u9a8c\u7ed3\u679c
<\/figcaption><\/figure>\n\n\n\n
\"\"
[\u56fe 5] \u96f6\u6837\u672c FLAN \u5bf9\u672a\u77e5\u4efb\u52a1\u7c7b\u578b\u7684\u6027\u80fd <\/figcaption><\/figure>\n\n\n\n

\u719f\u6089NLP\u7684\u4eba\u53ef\u80fd\u4f1a\u8ba4\u4e3a\u8fd9\u7bc7\u6587\u7ae0\u662f\u53e6\u4e00\u4e2a\u201cA+B\u201d\u4efb\u52a1\uff08A=\u63d0\u793a\u8c03\u4f18\uff0cB=\u591a\u4efb\u52a1\u5b66\u4e60\uff09\u3002\u4f46\u662f\uff0c\u8fd9\u4e9b A+B \u5c06\u662f \u901a\u7528\u81ea\u7136\u8bed\u8a00\u5904\u7406\u6a21\u578b<\/strong>\u53bb\u505a \u89e3\u51b3\u65b9\u6848\/\u65b9\u6cd5<\/strong>\u6211\u60f3\u8fd9\u53ef\u80fd\u662f\u3002\u9996\u5148\uff0c\u901a\u8fc7\u5927\u91cf\u672a\u6807\u6ce8\u7684\u8bed\u6599\uff0c\u8bad\u7ec3\u4e00\u4e2a\u5343\u4ebf\u53c2\u6570\u7684\u5927\u89c4\u6a21\u81ea\u56de\u5f52\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u6216\u8005\u9009\u62e9\u4e00\u4e2a\u73b0\u6709\u7684\u8bad\u7ec3\u6a21\u578b\uff0c\u5728\u7b2c\u4e8c\u6b65\uff0c\u8fd9\u4e9b\u6a21\u578b\u662f\u53ef\u4ee5\u5fae\u8c03\u7406\u89e3\u7684\u548c\u521b\u4f5c\u4efb\u52a1\u3002\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\uff0c\u4f7f\u7528\u7c7b\u4f3c\u4e8e\u8bfe\u7a0b\u5b66\u4e60\u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u5148\u8bad\u7ec3\u4e00\u4e2a\u8f83\u4f4e\u7ea7\u522b\u7684\u4efb\u52a1\uff08\u4f8b\u5982 NER \u8bc6\u522b\u3001\u5e8f\u5217\u8bed\u4e49\u6ce8\u91ca\uff09\uff0c\u7136\u540e\u518d\u8bad\u7ec3\u4e00\u4e2a\u8f83\u9ad8\u7ea7\u522b\u7684\u4efb\u52a1\uff08\u4f8b\u5982\u903b\u8f91\u63a8\u7406\u3001QA\uff09\u3002\u5b83\u8fd8\u9996\u5148\u5b66\u4e60\u8d44\u6e90\u4e30\u5bcc\u7684\u4efb\u52a1\uff08\u4f8b\u5982\u82f1\u8bed\/\u5927\u6570\u636e\u4efb\u52a1\uff09\uff0c\u7136\u540e\u5b66\u4e60\u8f83\u5c11\u7684\u8d44\u6e90\uff08\u4f8b\u5982\u5176\u4ed6\u8bed\u8a00\/\u4f4e\u6570\u636e\u4efb\u52a1\uff09\uff0c\u5e76\u4f7f\u7528\u9002\u914d\u5668\u5c06\u6bcf\u4e2a\u4efb\u52a1\u7684\u76f8\u5173\u90e8\u5206\u4fdd\u7559\u5728\u6a21\u578b\u4e2d\u3002\u6700\u540e\uff0c\u6211\u4eec\u63d0\u4f9b\u547d\u4ee4\/\u6307\u4ee4\uff0c\u4ee5\u4fbf\u6a21\u578b\u53ef\u4ee5\u63a8\u7406\u65b0\u6570\u636e\u548c\u65b0\u4efb\u52a1\u3002\u5982\u679c\u8fd9\u79cd\u901a\u7528\u65b9\u6cd5\u5f97\u5230\u5145\u5206\u5229\u7528\uff0c\u6211\u5f88\u671f\u5f85\u5b83\u80fd\u591f\u6267\u884c\u54ea\u4e9b\u65b0\u4efb\u52a1\uff01<\/p>\n\n\n\n

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

\u53c2\u8003\uff1a<\/p>\n\n\n\n

https:\/\/arxiv.org\/pdf\/2109.01652.pdf<\/a><\/p>\n

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

[\uc735\ud569\uc5f0\uad6c\ud300 \uc2ec\ud64d\ub9e4] NLP \ubd84\uc57c\uc5d0\uc11c 2020\ub144 Top 10 \ud0a4\uc6cc\ub4dc\ub97c \ubf51\ub294\ub2e4\uba74 GPT-3(Language Models are Few shot Learners) \ub2f9\uc5f0\ud788 \uc21c\uc704 \ub0b4\uc5d0 \uc788\uc744 \uac81\ub2c8\ub2e4. \ud604\uc7ac\uae4c\uc9c0\ub3c4 GPT-3\uc758 \uc5c4\uccad\ub09c \uc591\uc758 \ub9e4\uac1c\ubcc0\uc218\uc640 \uc6b0\uc218\ud55c \uc131\ub2a5\uc740 \uc5ec\uc804\ud788 \u200b\u200bNLP \ubd84\uc57c \ub0b4\uc678\ub97c \ub9c9\ub860\ud558\uace0 \uc778\uae30\ub97c \ub354\ud574\uac00\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 NLP \uc5f0\uad6c\uc6d0\uc73c\ub85c\uc11c \ucd5c\ucca8\ub2e8 \uc5f0\uad6c\uc5d0 \ub300\ud55c GPT-3\uc758 \uac00\uc7a5 \ud070 \uae30\uc5ec\ub294 \uc77c\ubc18 \uc791\uc5c5( \ud2b9\ud788 zero-shot , few-shot ) \uc5d0\uc11c Prompt-tuning \uae30\uc220\uc758…<\/p>\n

<\/span><\/div>","protected":false},"author":1,"featured_media":60921,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"rank_math_lock_modified_date":false,"footnotes":""},"categories":[18,19],"tags":[188,334],"class_list":["post-60877","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tech03","category-tech04","tag-featured","tag-gpt-3","category-18","category-19","description-off"],"_links":{"self":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts\/60877","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/comments?post=60877"}],"version-history":[{"count":13,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts\/60877\/revisions"}],"predecessor-version":[{"id":60923,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts\/60877\/revisions\/60923"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/media\/60921"}],"wp:attachment":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/media?parent=60877"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/categories?post=60877"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/tags?post=60877"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}