{"id":39426,"date":"2022-12-24T22:23:19","date_gmt":"2022-12-24T14:23:19","guid":{"rendered":"https:\/\/swarma.org\/?p=39426"},"modified":"2022-12-24T22:23:19","modified_gmt":"2022-12-24T14:23:19","slug":"%e4%b8%87%e5%ad%97%e6%8b%86%e8%a7%a3%ef%bc%81%e8%bf%bd%e6%ba%afchatgpt%e5%90%84%e9%a1%b9%e8%83%bd%e5%8a%9b%e7%9a%84%e8%b5%b7%e6%ba%90","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=39426","title":{"rendered":"\u4e07\u5b57\u62c6\u89e3\uff01\u8ffd\u6eafChatGPT\u5404\u9879\u80fd\u529b\u7684\u8d77\u6e90"},"content":{"rendered":"
\n
\n
\n
\n

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

\n
\n
\n
\n
\n
\n

<\/section>\n<\/section>\n
\n
\n

\u5bfc\u8bed<\/strong><\/p>\n<\/section>\n<\/section>\n

\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n

\u4e00\u7bc7\u5341\u5206\u6df1\u5ea6\u5256\u6790GPT\u7cfb\u5217\u6a21\u578b\u7684\u6587\u7ae0\uff0c\u4f5c\u8005\u7ffb\u8bd1\u6210\u4e86\u4e2d\u6587\uff0c\u5728\u8fd9\u91cc\u5206\u4eab\u7ed9\u5927\u5bb6\u3002<\/span><\/strong><\/span><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/strong><\/section>\n<\/section>\n
\n

\u674erumor<\/span> <\/span>| <\/strong><\/span>\u6765\u6e90<\/strong><\/span><\/p>\n

<\/span><\/p>\n<\/section>\n

\n
<\/span><\/span><\/section>\n
\n
\n
\u4f5c\u8005\uff1a<\/span>\u7b26\u5c27<\/strong>, yao.fu@ed.ac.uk\uff0c<\/span>\u7231\u4e01\u5821\u5927\u5b66 (University of Edinburgh) \u535a\u58eb\u751f\uff0c\u672c\u79d1\u6bd5\u4e1a\u4e8e\u5317\u4eac\u5927\u5b66
\u4e0e <\/span>\u5f6d\u660a<\/strong>\uff0c<\/span>Tushar Khot<\/strong>
<\/sup>\u5728 \u827e\u4f26\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u9662 (Allen Institute for AI) \u5171\u540c\u5b8c\u6210\u82f1\u6587\u539f\u7a3f
\u4e0e \u5251\u6865\u5927\u5b66<\/span>\u90ed\u5fd7\u6c5f<\/strong> \u5171\u540c\u7ffb\u8bd1\u4e3a\u4e2d\u6587<\/span><\/span><\/section>\n<\/section>\n<\/blockquote>\n

<\/section>\n
\n
\n
\n
\u611f\u8c22 \u4e0a\u6d77\u4ea4\u901a\u5927\u5b66<\/span>\u4f55\u4fca\u8d24<\/strong>\uff0c\u52a0\u5dde\u5927\u5b66\u6d1b\u6749\u77f6\u5206\u6821<\/span>\u9c81\u76fc<\/strong>\uff0c\u8fbe\u7279\u8305\u65af\u5b66\u9662<\/span>\u5218\u777f\u535a<\/strong> \u5bf9\u521d\u7a3f\u7684\u8ba8\u8bba\u4e0e\u5efa\u8bae\u3002<\/span><\/section>\n
\u611f\u8c22 <\/span>Raj Ammanabrolu<\/strong> (Allen Institute for AI), <\/span>Peter Liu<\/strong> (Google Brain), <\/span>Brendan Dolan-Gavitt<\/strong> (New York University), <\/span>Denny Zhou<\/strong> (Google Brain) \u5bf9\u7ec8\u7a3f\u7684\u8ba8\u8bba\u548c\u5efa\u8bae\uff0c\u4ed6\u4eec\u7684\u5efa\u8bae\u6781\u5927\u7a0b\u5ea6\u4e0a\u589e\u52a0\u4e86\u672c\u6587\u7684\u5b8c\u6574\u5ea6\u3002<\/span><\/span><\/span><\/section>\n
\u82f1\u6587\u539f\u7248\uff1a<\/span><\/section>\n
https:\/\/franxyao.github.io\/blog.html<\/span><\/section>\n<\/section>\n<\/section>\n<\/blockquote>\n<\/section>\n

\u6700\u8fd1\uff0cOpenAI\u7684\u9884\u8bad\u7ec3\u6a21\u578bChatGPT\u7ed9\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u7814\u7a76\u4eba\u5458\u7559\u4e0b\u4e86\u6df1\u523b\u7684\u5370\u8c61\u548c\u542f\u53d1\u3002<\/span>\u6beb\u65e0\u7591\u95ee\uff0c\u5b83\u53c8\u5f3a\u53c8\u806a\u660e\uff0c\u4e14\u8ddf\u5b83\u8bf4\u8bdd\u5f88\u597d\u73a9\uff0c\u8fd8\u4f1a\u5199\u4ee3\u7801\u3002<\/span>\u5b83\u5728\u591a\u4e2a\u65b9\u9762\u7684\u80fd\u529b\u8fdc\u8fdc\u8d85\u8fc7\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406\u7814\u7a76\u8005\u4eec\u7684\u9884\u671f\u3002<\/span>\u4e8e\u662f\u6211\u4eec\u81ea\u7136\u5c31\u6709\u4e00\u4e2a\u95ee\u9898\uff1a<\/span>ChatGPT \u662f\u600e\u4e48\u53d8\u5f97\u8fd9\u4e48\u5f3a\u7684\uff1f<\/span>\u5b83\u7684\u5404\u79cd\u5f3a\u5927\u7684\u80fd\u529b\u5230\u5e95\u4ece\u4f55\u800c\u6765\uff1f<\/span>\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u8bd5\u56fe\u5256\u6790 ChatGPT \u7684\u7a81\u73b0\u80fd\u529b\uff08Emergent Ability\uff09\uff0c\u8ffd\u6eaf\u8fd9\u4e9b\u80fd\u529b\u7684\u6765\u6e90\uff0c\u5e0c\u671b\u80fd\u591f\u7ed9\u51fa\u4e00\u4e2a\u5168\u9762\u7684\u6280\u672f\u8def\u7ebf\u56fe\uff0c\u6765\u8bf4\u660e GPT-3.5 \u6a21\u578b\u7cfb\u5217\u4ee5\u53ca\u76f8\u5173\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\u662f\u5982\u4f55\u4e00\u6b65\u6b65\u8fdb\u5316\u6210\u76ee\u524d\u7684\u5f3a\u5927\u5f62\u6001\u3002<\/span>
<\/p>\n

\u6211\u4eec\u5e0c\u671b\u8fd9\u7bc7\u6587\u7ae0\u80fd\u591f\u4fc3\u8fdb\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u900f\u660e\u5ea6\uff0c\u6210\u4e3a\u5f00\u6e90\u793e\u533a\u5171\u540c\u52aa\u529b\u590d\u73b0 GPT-3.5 \u7684\u8def\u7ebf\u56fe\u3002<\/span><\/p>\n

\n

\u81f4\u56fd\u5185\u7684\u540c\u80de\u4eec\uff1a<\/span><\/p>\n

    \n
  • \n
    \u5728\u56fd\u9645\u5b66\u672f\u754c\u770b\u6765\uff0cChatGPT \/ GPT-3.5 \u662f\u4e00\u79cd\u5212\u65f6\u4ee3\u7684\u4ea7\u7269\uff0c\u5b83\u4e0e\u4e4b\u524d\u5e38\u89c1\u7684\u8bed\u8a00\u6a21\u578b (Bert\/ Bart\/ T5) \u7684\u533a\u522b\uff0c\u51e0\u4e4e\u662f\u5bfc\u5f39\u4e0e\u5f13\u7bad\u7684\u533a\u522b\uff0c\u4e00\u5b9a\u8981\u5f15\u8d77\u6700\u9ad8\u7a0b\u5ea6\u7684\u91cd\u89c6\u3002<\/span><\/section>\n<\/li>\n
  • \n
    \u5728\u6211\u8ddf\u56fd\u9645\u540c\u884c\u7684\u4ea4\u6d41\u4e2d\uff0c\u56fd\u9645\u4e0a\u7684\u4e3b\u6d41\u5b66\u672f\u673a\u6784 (\u5982\u65af\u5766\u798f\u5927\u5b66\uff0c\u4f2f\u514b\u5229\u52a0\u5dde\u5927\u5b66) \u548c\u4e3b\u6d41\u4e1a\u754c\u7814\u7a76\u9662\uff08\u5982\u8c37\u6b4c\u5927\u8111\uff0c\u5fae\u8f6f\u7814\u7a76\u9662\uff09\u90fd\u5df2\u7ecf\u5168\u9762\u62e5\u62b1\u5927\u6a21\u578b\u3002<\/span><\/section>\n<\/li>\n
  • \n
    \u5728\u5f53\u524d\u8fd9\u4e2a\u9636\u6bb5\uff0c\u56fd\u5185\u7684\u6280\u672f\u6c34\u51c6\uff0c\u5b66\u672f\u89c6\u91ce\uff0c\u6cbb\u5b66\u7406\u5ff5\u548c\u56fd\u9645\u524d\u6cbf\u7684\u5dee\u8ddd\u4f3c\u4e4e\u5e76\u6ca1\u6709\u51cf\u5c11\uff0c\u53cd\u800c\u6b63\u5728\u6269\u5927\uff0c\u5982\u679c\u73b0\u72b6\u6301\u7eed\u4e0b\u53bb\uff0c\u6781\u6709\u53ef\u80fd\u51fa\u73b0\u6280\u672f\u65ad\u4ee3\u3002<\/span><\/section>\n<\/li>\n
  • \n
    \u6b64\u8bda\u5371\u6025\u5b58\u4ea1\u4e4b\u79cb\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
    \n

    \u591a\u5e74\u4ee5\u540e\uff0c\u9762\u5bf9\u884c\u5211\u961f\uff0c\u5965\u96f7\u91cc\u4e9a\u8bfa\u00b7\u5e03\u6069\u8fea\u4e9a\u4e0a\u6821\u5c06\u4f1a\u56de\u60f3\u8d77\u7236\u4eb2\u5e26\u4ed6\u53bb\u89c1\u8bc6\u51b0\u5757\u7684\u90a3\u4e2a\u9065\u8fdc\u7684\u4e0b\u5348\u3002\u2014\u2014 \u300a\u767e\u5e74\u5b64\u72ec\u300b<\/em> \u52a0\u897f\u4e9a\u00b7\u9a6c\u5c14\u514b\u65af<\/span><\/p>\n<\/blockquote>\n


    <\/p>\n

    \n
    \n
    \n
    \n
    \n

    <\/section>\n

    <\/section>\n
    \n
    <\/svg><\/section>\n<\/section>\n<\/section>\n
    \n

    1. 2020\u7248\u521d\u4ee3GPT-3\u4e0e\u5927\u89c4\u6a21\u8bad\u7ec3<\/span><\/strong><\/strong><\/p>\n<\/section>\n

    \n

    <\/section>\n

    <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
     <\/span><\/section>\n

    \u521d\u4ee3GPT-3\u5c55\u793a\u4e86\u4e09\u4e2a\u91cd\u8981\u80fd\u529b\uff1a<\/span><\/p>\n

      \n
    • \n
      \u8bed\u8a00\u751f\u6210<\/strong>\uff1a\u9075\u5faa\u63d0\u793a\u8bcd\uff08prompt\uff09\uff0c\u7136\u540e\u751f\u6210\u8865\u5168\u63d0\u793a\u8bcd\u7684\u53e5\u5b50\u3002\u8fd9\u4e5f\u662f\u4eca\u5929\u4eba\u7c7b\u4e0e\u8bed\u8a00\u6a21\u578b\u6700\u666e\u904d\u7684\u4ea4\u4e92\u65b9\u5f0f\u3002<\/span><\/section>\n<\/li>\n
    • \n
      \u4e0a\u4e0b\u6587\u5b66\u4e60 (in-context learning)<\/strong>:  \u9075\u5faa\u7ed9\u5b9a\u4efb\u52a1\u7684\u51e0\u4e2a\u793a\u4f8b\uff0c\u7136\u540e\u4e3a\u65b0\u7684\u6d4b\u8bd5\u7528\u4f8b\u751f\u6210\u89e3\u51b3\u65b9\u6848\u3002\u5f88\u91cd\u8981\u7684\u4e00\u70b9\u662f\uff0cGPT-3\u867d\u7136\u662f\u4e2a\u8bed\u8a00\u6a21\u578b\uff0c\u4f46\u5b83\u7684\u8bba\u6587\u51e0\u4e4e\u6ca1\u6709\u8c08\u5230\u201c\u8bed\u8a00\u5efa\u6a21\u201d (language modeling) \u2014\u2014 \u4f5c\u8005\u5c06\u4ed6\u4eec\u5168\u90e8\u7684\u5199\u4f5c\u7cbe\u529b\u90fd\u6295\u5165\u5230\u4e86\u5bf9\u4e0a\u4e0b\u6587\u5b66\u4e60\u7684\u613f\u666f\u4e0a\uff0c\u8fd9\u624d\u662f GPT-3\u7684\u771f\u6b63\u91cd\u70b9\u3002<\/span><\/section>\n<\/li>\n
    • \n
      \u4e16\u754c\u77e5\u8bc6<\/strong>\uff1a\u5305\u62ec\u4e8b\u5b9e\u6027\u77e5\u8bc6 (factual knowledge) \u548c\u5e38\u8bc6 (commonsense)\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

      \u90a3\u4e48\u8fd9\u4e9b\u80fd\u529b\u4ece\u4f55\u800c\u6765\u5462\uff1f<\/span><\/p>\n

      \u57fa\u672c\u4e0a\uff0c\u4ee5\u4e0a\u4e09\u79cd\u80fd\u529b\u90fd\u6765\u81ea\u4e8e\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\uff1a\u5728\u67093000\u4ebf\u5355\u8bcd\u7684\u8bed\u6599\u4e0a\u9884\u8bad\u7ec3\u62e5\u67091750\u4ebf\u53c2\u6570\u7684\u6a21\u578b\uff08 \u8bad\u7ec3\u8bed\u6599\u768460%\u6765\u81ea\u4e8e 2016 – 2019 \u7684 C4 + 22% \u6765\u81ea\u4e8e WebText2 + 16% \u6765\u81ea\u4e8eBooks + 3%\u6765\u81ea\u4e8eWikipedia\uff09\u3002\u5176\u4e2d\uff1a<\/span><\/p>\n

        \n
      • \n
        \u8bed\u8a00\u751f\u6210<\/strong>\u7684\u80fd\u529b\u6765\u81ea\u4e8e\u8bed\u8a00\u5efa\u6a21\u7684\u8bad\u7ec3\u76ee\u6807<\/strong> (language modeling)\u3002<\/span><\/section>\n<\/li>\n
      • \n
        \u4e16\u754c\u77e5\u8bc6<\/strong>\u6765\u81ea 3000 \u4ebf\u5355\u8bcd\u7684\u8bad\u7ec3\u8bed\u6599\u5e93<\/strong>\uff08\u4e0d\u7136\u8fd8\u80fd\u662f\u54ea\u513f\u5462\uff09\u3002<\/span><\/section>\n<\/li>\n
      • \n
        \u6a21\u578b\u7684 1750 \u4ebf\u53c2\u6570<\/strong>\u662f\u4e3a\u4e86\u5b58\u50a8\u77e5\u8bc6<\/strong>\uff0cLiang et al. (2022) \u7684\u6587\u7ae0\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9\u3002\u4ed6\u4eec\u7684\u7ed3\u8bba\u662f\uff0c\u77e5\u8bc6\u5bc6\u96c6\u578b\u4efb\u52a1\u7684\u6027\u80fd\u4e0e\u6a21\u578b\u5927\u5c0f\u606f\u606f\u76f8\u5173\u3002<\/span><\/section>\n<\/li>\n
      • \n
        \u4e0a\u4e0b\u6587\u5b66\u4e60\u7684\u80fd\u529b\u6765\u6e90\u53ca\u4e3a\u4ec0\u4e48\u4e0a\u4e0b\u6587\u5b66\u4e60\u53ef\u4ee5\u6cdb\u5316\uff0c\u4ecd\u7136\u96be\u4ee5\u6eaf\u6e90\u3002<\/strong> \u76f4\u89c9\u4e0a\uff0c\u8fd9\u79cd\u80fd\u529b\u53ef\u80fd\u6765\u81ea\u4e8e\u540c\u4e00\u4e2a\u4efb\u52a1\u7684\u6570\u636e\u70b9\u5728\u8bad\u7ec3\u65f6\u6309\u987a\u5e8f\u6392\u5217\u5728\u540c\u4e00\u4e2a batch \u4e2d\u3002\u7136\u800c\uff0c\u5f88\u5c11\u6709\u4eba\u7814\u7a76\u4e3a\u4ec0\u4e48\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u4f1a\u4fc3\u4f7f\u4e0a\u4e0b\u6587\u5b66\u4e60\uff0c\u4ee5\u53ca\u4e3a\u4ec0\u4e48\u4e0a\u4e0b\u6587\u5b66\u4e60\u7684\u884c\u4e3a\u4e0e\u5fae\u8c03 (fine-tuning) \u5982\u6b64\u4e0d\u540c\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

        \u4ee4\u4eba\u597d\u5947\u7684\u662f\uff0c\u521d\u4ee3\u7684GPT-3\u6709\u591a\u5f3a\u3002<\/strong>\u5176\u5b9e\u6bd4\u8f83\u96be\u786e\u5b9a\u521d\u4ee3 GPT-3\uff08\u5728 OpenAI API \u4e2d\u88ab\u79f0\u4e3a<\/span>davinci<\/span><\/code>\uff09\u5230\u5e95\u662f\u201c\u5f3a\u201d\u8fd8\u662f\u201c\u5f31\u201d\u3002\u4e00\u65b9\u9762\uff0c\u5b83\u5408\u7406\u5730\u56de\u5e94\u4e86\u67d0\u4e9b\u7279\u5b9a\u7684\u67e5\u8be2\uff0c\u5e76\u5728\u8bb8\u591a\u6570\u636e\u96c6\u4e2d\u8fbe\u5230\u4e86\u8fd8\u4e0d\u9519\u7684\u6027\u80fd\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u5b83\u5728\u8bb8\u591a\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u8fd8\u4e0d\u5982 T5 \u8fd9\u6837\u7684\u5c0f\u6a21\u578b<\/strong>\uff08\u53c2\u89c1\u5176\u539f\u59cb\u8bba\u6587\uff09\u3002\u5728\u4eca\u5929\uff082022 \u5e74 12 \u6708\uff09ChatGPT \u7684\u6807\u51c6\u4e0b\uff0c\u5f88\u96be\u8bf4\u521d\u4ee3\u7684 GPT-3 \u662f\u201c\u667a\u80fd\u7684\u201d\u3002Meta \u5f00\u6e90\u7684 OPT \u6a21\u578b\u8bd5\u56fe\u590d\u73b0\u521d\u4ee3 GPT-3\uff0c\u4f46\u5b83\u7684\u80fd\u529b\u4e0e\u5f53\u4eca\u7684\u6807\u51c6\u4e5f\u5f62\u6210\u4e86\u5c16\u9510\u7684\u5bf9\u6bd4\u3002\u8bb8\u591a\u6d4b\u8bd5\u8fc7 OPT \u7684\u4eba\u4e5f\u8ba4\u4e3a\u4e0e\u73b0\u5728\u7684<\/span>text-davinci-002<\/span><\/code>\u76f8\u6bd4\uff0c\u8be5\u6a21\u578b\u786e\u5b9e \u201c\u4e0d\u548b\u5730\u201d\u3002\u5c3d\u7ba1\u5982\u6b64\uff0cOPT \u53ef\u80fd\u662f\u521d\u4ee3 GPT-3 \u7684\u4e00\u4e2a\u8db3\u591f\u597d\u7684\u5f00\u6e90\u7684\u8fd1\u4f3c\u6a21\u578b\u4e86\uff08\u6839\u636e OPT \u8bba\u6587\u548c\u65af\u5766\u798f\u5927\u5b66\u7684 HELM \u8bc4\u4f30\uff09\u3002<\/span><\/p>\n

        \u867d\u7136\u521d\u4ee3\u7684 GPT-3 \u53ef\u80fd\u8868\u9762\u4e0a\u770b\u8d77\u6765\u5f88\u5f31\uff0c\u4f46\u540e\u6765\u7684\u5b9e\u9a8c\u8bc1\u660e\uff0c\u521d\u4ee3 GPT-3 \u6709\u7740\u975e\u5e38\u5f3a\u7684\u6f5c\u529b\u3002\u8fd9\u4e9b\u6f5c\u529b\u540e\u6765\u88ab\u4ee3\u7801\u8bad\u7ec3\u3001\u6307\u4ee4\u5fae\u8c03 (instruction tuning) \u548c\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60 (reinforcement learning with human feedback, RLHF) \u89e3\u9501\uff0c\u6700\u7ec8\u4f53\u5c55\u793a\u51fa\u6781\u4e3a\u5f3a\u5927\u7684\u7a81\u73b0\u80fd\u529b\u3002<\/span><\/section>\n


        <\/p>\n

        \n
        \n
        \n
        \n
        \n

        <\/section>\n

        <\/section>\n
        \n
        <\/svg><\/section>\n<\/section>\n<\/section>\n
        \n

        2. \u4ece2020\u7248GPT-3\u52302022\u7248ChatGPT<\/span><\/strong><\/strong><\/p>\n<\/section>\n

        \n

        <\/section>\n

        <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

         <\/span><\/p>\n

        \u4ece\u6700\u521d\u7684 GPT-3 \u5f00\u59cb\uff0c\u4e3a\u4e86\u5c55\u793a OpenAI \u662f\u5982\u4f55\u53d1\u5c55\u5230ChatGPT\u7684\uff0c\u6211\u4eec\u770b\u4e00\u4e0b GPT-3.5 \u7684\u8fdb\u5316\u6811\uff1a<\/span><\/p>\n

        <\/figure>\n

        <\/section>\n

        \u5728 2020 \u5e74 7 \u6708<\/strong>\uff0cOpenAI \u53d1\u5e03\u4e86\u6a21\u578b\u7d22\u5f15\u4e3a\u7684 <\/span>davinci<\/span><\/code> \u7684\u521d\u4ee3 GPT-3 \u8bba\u6587\uff0c\u4ece\u6b64\u5b83\u5c31\u5f00\u59cb\u4e0d\u65ad\u8fdb\u5316\u3002\u5728 2021 \u5e74 7 \u6708<\/strong>\uff0cCodex \u7684\u8bba\u6587\u53d1\u5e03\uff0c\u5176\u4e2d\u521d\u59cb\u7684 Codex \u662f\u6839\u636e\uff08\u53ef\u80fd\u662f\u5185\u90e8\u7684\uff09120 \u4ebf\u53c2\u6570\u7684 GPT-3 \u53d8\u4f53\u8fdb\u884c\u5fae\u8c03\u7684\u3002\u540e\u6765\u8fd9\u4e2a 120 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u6f14\u53d8\u6210 OpenAI API \u4e2d\u7684<\/span>code-cushman-001<\/span><\/code>\u3002\u5728 2022 \u5e74 3 \u6708<\/strong>\uff0cOpenAI \u53d1\u5e03\u4e86\u6307\u4ee4\u5fae\u8c03 (instruction tuning) \u7684\u8bba\u6587\uff0c\u5176\u76d1\u7763\u5fae\u8c03 (supervised instruction tuning) \u7684\u90e8\u5206\u5bf9\u5e94\u4e86<\/span>davinci-instruct-beta<\/span><\/code>\u548c<\/span>text-davinci-001<\/span><\/code>\u3002\u5728 2022 \u5e74 4 \u6708\u81f3 7 \u6708\u7684<\/strong>\uff0cOpenAI \u5f00\u59cb\u5bf9<\/span>code-davinci-002<\/span><\/code>\u6a21\u578b\u8fdb\u884c Beta \u6d4b\u8bd5\uff0c\u4e5f\u79f0\u5176\u4e3a Codex\u3002\u7136\u540e<\/span>code-davinci-002<\/span><\/code>\u3001<\/span>text-davinci-003<\/span><\/code>\u548c<\/span>ChatGPT<\/span><\/code> \u90fd\u662f\u4ece<\/span>code-davinci-002<\/span><\/code>\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03\u5f97\u5230\u7684\u3002\u8be6\u7ec6\u4fe1\u606f\u8bf7\u53c2\u9605 OpenAI\u7684\u6a21\u578b\u7d22\u5f15\u6587\u6863\u3002<\/span><\/p>\n

        \u5c3d\u7ba1 Codex \u542c\u7740\u50cf\u662f\u4e00\u4e2a\u53ea\u7ba1\u4ee3\u7801\u7684\u6a21\u578b\uff0c\u4f46<\/span>code-davinci-002<\/span><\/code>\u53ef\u80fd\u662f\u6700\u5f3a\u5927\u7684\u9488\u5bf9\u81ea\u7136\u8bed\u8a00<\/strong>\u7684GPT-3.5 \u53d8\u4f53\uff08\u4f18\u4e8e <\/span>text-davinci-002<\/span><\/code>\u548c <\/span>-003<\/span><\/code>\uff09\u3002<\/span>code-davinci-002<\/span><\/code>\u5f88\u53ef\u80fd\u5728\u6587\u672c\u548c\u4ee3\u7801\u4e0a\u90fd\u7ecf\u8fc7\u8bad\u7ec3\uff0c\u7136\u540e\u6839\u636e\u6307\u4ee4\u8fdb\u884c\u8c03\u6574\uff08\u5c06\u5728\u4e0b\u9762\u89e3\u91ca\uff09\u3002\u7136\u540e2022 \u5e74 5-6 \u6708<\/strong>\u53d1\u5e03\u7684<\/span>text-davinci-002<\/span><\/code>\u662f\u4e00\u4e2a\u57fa\u4e8e<\/span>code-davinci-002<\/span><\/code>\u7684\u6709\u76d1\u7763\u6307\u4ee4\u5fae\u8c03 (supervised instruction tuned) \u6a21\u578b\u3002\u5728<\/span>text-davinci-002<\/span><\/code>\u4e0a\u9762\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03<\/strong>\u5f88\u53ef\u80fd\u964d\u4f4e<\/strong>\u4e86\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60<\/strong>\u80fd\u529b\uff0c\u4f46\u662f\u589e\u5f3a\u4e86<\/strong>\u6a21\u578b\u7684\u96f6\u6837\u672c\u80fd\u529b<\/strong>\uff08\u5c06\u5728\u4e0b\u9762\u89e3\u91ca\uff09\u3002\u7136\u540e\u662f<\/span>text-davinci-003<\/span><\/code>\u548c <\/span>ChatGPT<\/span><\/code>\uff0c\u5b83\u4eec\u90fd\u5728 2022 \u5e74 11 \u6708<\/strong>\u53d1\u5e03\uff0c\u662f\u4f7f\u7528\u7684\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\u7684\u7248\u672c\u6307\u4ee4\u5fae\u8c03 (instruction tuning with reinforcement learning from human feedback) \u6a21\u578b\u7684\u4e24\u79cd\u4e0d\u540c\u53d8\u4f53\u3002<\/span>text-davinci-003<\/span><\/code> \u6062\u590d\u4e86\uff08\u4f46\u4ecd\u7136\u6bd4<\/span>code-davinci-002<\/span><\/code>\u5dee\uff09\u4e00\u4e9b\u5728<\/span>text-davinci-002<\/span><\/code> \u4e2d\u4e22\u5931\u7684\u90e8\u5206\u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd<\/strong>\u529b\uff08\u5927\u6982\u662f\u56e0\u4e3a\u5b83\u5728\u5fae\u8c03\u7684\u65f6\u5019\u6df7\u5165\u4e86\u8bed\u8a00\u5efa\u6a21\uff09 \u5e76\u8fdb\u4e00\u6b65\u6539\u8fdb\u4e86\u96f6\u6837\u672c\u80fd\u529b\uff08\u5f97\u76ca\u4e8eRLHF\uff09\u3002\u53e6\u4e00\u65b9\u9762\uff0cChatGPT \u4f3c\u4e4e\u727a\u7272\u4e86\u51e0\u4e4e\u6240\u6709\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60\u7684\u80fd\u529b<\/strong>\u6765\u6362\u53d6<\/strong>\u5efa\u6a21\u5bf9\u8bdd\u5386\u53f2\u7684\u80fd\u529b\u3002<\/span><\/p>\n

        \u603b\u7684\u6765\u8bf4\uff0c\u5728 2020 – 2021 \u5e74\u671f\u95f4\uff0c\u5728<\/span>code-davinci-002<\/span><\/code>\u4e4b\u524d\uff0cOpenAI \u5df2\u7ecf\u6295\u5165\u4e86\u5927\u91cf\u7684\u7cbe\u529b\u901a\u8fc7\u4ee3\u7801\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u6765\u589e\u5f3aGPT-3\u3002\u5f53\u4ed6\u4eec\u5b8c\u6210<\/span>code-davinci-002<\/span><\/code>\u65f6\uff0c\u6240\u6709\u7684\u80fd\u529b\u90fd\u5df2\u7ecf\u5b58\u5728\u4e86\u3002\u5f88\u53ef\u80fd\u540e\u7eed\u7684\u6307\u4ee4\u5fae\u8c03\uff0c\u65e0\u8bba\u662f\u901a\u8fc7\u6709\u76d1\u7763\u7684\u7248\u672c\u8fd8\u662f\u5f3a\u5316\u5b66\u4e60\u7684\u7248\u672c\uff0c\u90fd\u4f1a\u505a\u4ee5\u4e0b\u4e8b\u60c5\uff08\u7a0d\u540e\u4f1a\u8be6\u7ec6\u8bf4\u660e\uff09\uff1a<\/span><\/p>\n

          \n
        • \n
          \u6307\u4ee4\u5fae\u8c03\u4e0d\u4f1a\u4e3a\u6a21\u578b\u6ce8\u5165\u65b0\u7684\u80fd\u529b<\/strong> \u2014\u2014 \u6240\u6709\u7684\u80fd\u529b\u90fd\u5df2\u7ecf\u5b58\u5728\u4e86\u3002\u6307\u4ee4\u5fae\u8c03\u7684\u4f5c\u7528\u662f\u89e3\u9501 \/ \u6fc0\u53d1\u8fd9\u4e9b\u80fd\u529b<\/strong>\u3002\u8fd9\u4e3b\u8981\u662f\u56e0\u4e3a\u6307\u4ee4\u5fae\u8c03\u7684\u6570\u636e\u91cf\u6bd4\u9884\u8bad\u7ec3\u6570\u636e\u91cf\u5c11\u51e0\u4e2a\u6570\u91cf\u7ea7\uff08\u57fa\u7840\u7684\u80fd\u529b\u662f\u901a\u8fc7\u9884\u8bad\u7ec3\u6ce8\u5165\u7684\uff09\u3002<\/span><\/section>\n<\/li>\n
        • \n
          \u6307\u4ee4\u5fae\u8c03\u5c06 GPT-3.5 \u7684\u5206\u5316\u5230\u4e0d\u540c\u7684\u6280\u80fd\u6811\u3002<\/strong>\u6709\u4e9b\u66f4\u64c5\u957f\u4e0a\u4e0b\u6587\u5b66\u4e60**\uff0c\u5982<\/span>text-davinci-003<\/span><\/code>\uff0c\u6709\u4e9b\u66f4\u64c5\u957f\u5bf9\u8bdd\uff0c\u5982<\/span>ChatGPT<\/span><\/code>\u3002<\/span><\/section>\n<\/li>\n
        • \n
          \u6307\u4ee4\u5fae\u8c03\u901a\u8fc7\u727a\u7272\u6027\u80fd\u6362\u53d6\u4e0e\u4eba\u7c7b\u7684\u5bf9\u9f50\uff08alignment\uff09<\/strong>\u3002OpenAI \u7684\u4f5c\u8005\u5728\u4ed6\u4eec\u7684\u6307\u4ee4\u5fae\u8c03\u8bba\u6587\u4e2d\u79f0\u5176\u4e3a \u201c\u5bf9\u9f50\u7a0e\u201d (alignment tax)\u3002\u8bb8\u591a\u8bba\u6587\u90fd\u62a5\u9053\u4e86<\/span>code-davinci-002<\/span><\/code>\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5b9e\u73b0\u4e86\u6700\u4f73\u6027\u80fd\uff08\u4f46\u6a21\u578b\u4e0d\u4e00\u5b9a\u7b26\u5408\u4eba\u7c7b\u671f\u671b\uff09\u3002\u5728<\/span>code-davinci-002<\/span><\/code>\u4e0a\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03\u540e\uff0c\u6a21\u578b\u53ef\u4ee5\u751f\u6210\u66f4\u52a0\u7b26\u5408\u4eba\u7c7b\u671f\u5f85\u7684\u53cd\u9988\uff08\u6216\u8005\u8bf4\u6a21\u578b\u4e0e\u4eba\u7c7b\u5bf9\u9f50\uff09\uff0c\u4f8b\u5982\uff1a\u96f6\u6837\u672c\u95ee\u7b54\u3001\u751f\u6210\u5b89\u5168\u548c\u516c\u6b63\u7684\u5bf9\u8bdd\u56de\u590d\u3001\u62d2\u7edd\u8d85\u51fa\u6a21\u578b\u5b83\u77e5\u8bc6\u8303\u56f4\u7684\u95ee\u9898\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n


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

          \n
          \n
          \n
          \n
          \n

          <\/section>\n

          <\/section>\n
          \n
          <\/svg><\/section>\n<\/section>\n<\/section>\n
          \n

          3. Code-Davinci-002\u548c Text-Davinci-<\/span><\/strong><\/strong><\/span><\/strong><\/strong>002\uff0c\u5728\u4ee3\u7801\u4e0a\u8bad\u7ec3\uff0c\u5728\u6307\u4ee4\u4e0a\u5fae\u8c03<\/span><\/strong><\/strong><\/p>\n<\/section>\n

          \n

          <\/section>\n

          <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
           <\/span><\/section>\n

          \u5728<\/span>code-davinci-002<\/span><\/code>\u548c<\/span>text-davinci-002<\/span><\/code>\u4e4b\u524d\uff0c\u6709\u4e24\u4e2a\u4e2d\u95f4\u6a21\u578b\uff0c\u5206\u522b\u662f davinci-instruct-beta \u548c text-davinci-001\u3002\u4e24\u8005\u5728\u5f88\u591a\u65b9\u9762\u90fd\u6bd4\u4e0a\u8ff0\u7684\u4e24\u4e2a-002\u6a21\u578b\u5dee\uff08\u4f8b\u5982\uff0ctext-davinci-001 \u94fe\u5f0f\u601d\u7ef4\u63a8\u7406\u80fd\u529b\u4e0d\u5f3a\uff09\u3002\u6240\u4ee5\u6211\u4eec\u5728\u672c\u8282\u4e2d\u91cd\u70b9\u4ecb\u7ecd -002 \u578b\u53f7\u3002<\/span><\/p>\n

          3.1 \u590d\u6742\u63a8\u7406\u80fd\u529b\u7684\u6765\u6e90\u548c\u6cdb\u5316\u5230\u65b0\u4efb\u52a1\u7684\u80fd\u529b<\/strong><\/span><\/h2>\n

          \u6211\u4eec\u5173\u6ce8<\/span>code-davinci-002<\/span><\/code>\u548c<\/span>text-davinci-002<\/span><\/code>\uff0c\u8fd9\u4e24\u5144\u5f1f\u662f\u7b2c\u4e00\u7248\u7684 GPT3.5 \u6a21\u578b\uff0c\u4e00\u4e2a\u7528\u4e8e\u4ee3\u7801\uff0c\u53e6\u4e00\u4e2a\u7528\u4e8e\u6587\u672c\u3002\u5b83\u4eec\u8868\u73b0\u51fa\u4e86\u4e09\u79cd\u91cd\u8981\u80fd\u529b\u4e0e\u521d\u4ee3 GPT-3 \u4e0d\u540c\u7684\u80fd\u529b\uff1a<\/span><\/p>\n

            \n
          • \n
            \u54cd\u5e94\u4eba\u7c7b\u6307\u4ee4<\/strong>\uff1a\u4ee5\u524d\uff0cGPT-3 \u7684\u8f93\u51fa\u4e3b\u8981\u8bad\u7ec3\u96c6\u4e2d\u5e38\u89c1\u7684\u53e5\u5b50\u3002\u73b0\u5728\u7684\u6a21\u578b\u4f1a\u9488\u5bf9\u6307\u4ee4 \/ \u63d0\u793a\u8bcd\u751f\u6210\u66f4\u5408\u7406\u7684\u7b54\u6848\uff08\u800c\u4e0d\u662f\u76f8\u5173\u4f46\u65e0\u7528\u7684\u53e5\u5b50\uff09\u3002<\/span><\/section>\n<\/li>\n
          • \n
            \u6cdb\u5316\u5230\u6ca1\u6709\u89c1\u8fc7\u7684\u4efb\u52a1<\/strong>\uff1a\u5f53\u7528\u4e8e\u8c03\u6574\u6a21\u578b\u7684\u6307\u4ee4\u6570\u91cf\u8d85\u8fc7\u4e00\u5b9a\u7684\u89c4\u6a21\u65f6\uff0c\u6a21\u578b\u5c31\u53ef\u4ee5\u81ea\u52a8\u5728\u4ece\u6ca1\u89c1\u8fc7\u7684\u65b0\u6307\u4ee4\u4e0a\u4e5f\u80fd\u751f\u6210\u6709\u6548\u7684\u56de\u7b54\u3002\u8fd9\u79cd\u80fd\u529b\u5bf9\u4e8e\u4e0a\u7ebf\u90e8\u7f72\u81f3\u5173\u91cd\u8981<\/strong>\uff0c\u56e0\u4e3a\u7528\u6237\u603b\u4f1a\u63d0\u65b0\u7684\u95ee\u9898\uff0c\u6a21\u578b\u5f97\u7b54\u5f97\u51fa\u6765\u624d\u884c\u3002<\/span><\/section>\n<\/li>\n
          • \n
            \u4ee3\u7801\u751f\u6210\u548c\u4ee3\u7801\u7406\u89e3<\/strong>\uff1a\u8fd9\u4e2a\u80fd\u529b\u5f88\u663e\u7136\uff0c\u56e0\u4e3a\u6a21\u578b\u7528\u4ee3\u7801\u8bad\u7ec3\u8fc7\u3002<\/span><\/section>\n<\/li>\n
          • \n
            \u5229\u7528\u601d\u7ef4\u94fe (chain-of-thought) \u8fdb\u884c\u590d\u6742\u63a8\u7406<\/strong>\uff1a\u521d\u4ee3 GPT3 \u7684\u6a21\u578b\u601d\u7ef4\u94fe\u63a8\u7406\u7684\u80fd\u529b\u5f88\u5f31\u751a\u81f3\u6ca1\u6709\u3002code-davinci-002 \u548c text-davinci-002 \u662f\u4e24\u4e2a\u62e5\u6709\u8db3\u591f\u5f3a\u7684\u601d\u7ef4\u94fe\u63a8\u7406\u80fd\u529b\u7684\u6a21\u578b\u3002<\/strong><\/span><\/section>\n<\/li>\n
              \n
            • \n
              \u601d\u7ef4\u94fe\u63a8\u7406\u4e4b\u6240\u4ee5\u91cd\u8981\uff0c\u662f\u56e0\u4e3a\u601d\u7ef4\u94fe\u53ef\u80fd\u662f\u89e3\u9501\u7a81\u73b0\u80fd\u529b\u548c\u8d85\u8d8a\u7f29\u653e\u6cd5\u5219 (scaling laws) \u7684\u5173\u952e\u3002\u8bf7\u53c2\u9605\u4e0a\u4e00\u7bc7\u535a\u6587\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n

              \u8fd9\u4e9b\u80fd\u529b\u4ece\u4f55\u800c\u6765\uff1f<\/span><\/p>\n

              \u4e0e\u4e4b\u524d\u7684\u6a21\u578b\u76f8\u6bd4\uff0c\u4e24\u4e2a\u4e3b\u8981\u533a\u522b\u662f\u6307\u4ee4\u5fae\u8c03<\/strong>\u548c\u4ee3\u7801\u8bad\u7ec3<\/strong>\u3002\u5177\u4f53\u6765\u8bf4<\/span><\/p>\n

                \n
              • \n
                \u80fd\u591f\u54cd\u5e94\u4eba\u7c7b\u6307\u4ee4<\/strong>\u7684\u80fd\u529b\u662f\u6307\u4ee4\u5fae\u8c03<\/strong>\u7684\u76f4\u63a5\u4ea7\u7269\u3002<\/span><\/section>\n<\/li>\n
              • \n
                \u5bf9\u6ca1\u6709\u89c1\u8fc7\u7684\u6307\u4ee4\u505a\u51fa\u53cd\u9988<\/strong>\u7684\u6cdb\u5316\u80fd\u529b\u662f\u5728\u6307\u4ee4\u6570\u91cf\u8d85\u8fc7\u4e00\u5b9a\u7a0b\u5ea6\u4e4b\u540e\u81ea\u52a8\u51fa\u73b0\u7684<\/strong>\uff0cT0\u3001Flan \u548c FlanPaLM \u8bba\u6587\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9<\/span><\/section>\n<\/li>\n
              • \n
                \u4f7f\u7528\u601d\u7ef4\u94fe<\/strong>\u8fdb\u884c\u590d\u6742\u63a8\u7406<\/strong>\u7684\u80fd\u529b\u5f88\u53ef\u80fd\u662f\u4ee3\u7801\u8bad\u7ec3<\/strong>\u7684\u4e00\u4e2a\u795e\u5947\u7684\u526f\u4ea7\u7269<\/strong>\u3002\u5bf9\u6b64\uff0c\u6211\u4eec\u6709\u4ee5\u4e0b\u7684\u4e8b\u5b9e\u4f5c\u4e3a\u4e00\u4e9b\u652f\u6301\uff1a<\/span><\/section>\n<\/li>\n
                  \n
                • \n
                  \u6700\u521d\u7684 GPT-3 \u6ca1\u6709\u63a5\u53d7\u8fc7\u4ee3\u7801\u8bad\u7ec3\uff0c\u5b83\u4e0d\u80fd\u505a\u601d\u7ef4\u94fe<\/strong>\u3002<\/span><\/section>\n<\/li>\n
                • \n
                  text-davinci-001 \u6a21\u578b\uff0c\u867d\u7136\u7ecf\u8fc7\u4e86\u6307\u4ee4\u5fae\u8c03\uff0c\u4f46\u7b2c\u4e00\u7248\u601d\u7ef4\u94fe\u8bba\u6587\u62a5\u544a\u8bf4\uff0c\u5b83\u7684\u5b83\u601d\u7ef4\u94fe\u63a8\u7406\u7684\u80fd\u529b\u975e\u5e38\u5f31 \u2014\u2014 \u6240\u4ee5\u6307\u4ee4\u5fae\u8c03\u53ef\u80fd\u4e0d\u662f\u601d\u7ef4\u94fe\u5b58\u5728\u7684\u539f\u56e0\uff0c\u4ee3\u7801\u8bad\u7ec3\u624d\u662f\u6a21\u578b\u80fd\u505a\u601d\u7ef4\u94fe\u63a8\u7406\u7684\u6700\u53ef\u80fd\u539f\u56e0\u3002<\/strong><\/span><\/section>\n<\/li>\n
                • \n
                  PaLM \u6709 5% \u7684\u4ee3\u7801\u8bad\u7ec3\u6570\u636e\uff0c\u53ef\u4ee5\u505a\u601d\u7ef4\u94fe\u3002<\/span><\/section>\n<\/li>\n
                • \n
                  Codex\u8bba\u6587\u4e2d\u7684\u4ee3\u7801\u6570\u636e\u91cf\u4e3a 159G \uff0c\u5927\u7ea6\u662f\u521d\u4ee3 GPT-3 5700 \u4ebf\u8bad\u7ec3\u6570\u636e\u768428%\u3002code-davinci-002 \u53ca\u5176\u540e\u7eed\u53d8\u4f53\u53ef\u4ee5\u505a\u601d\u7ef4\u94fe\u63a8\u7406\u3002<\/span><\/section>\n<\/li>\n
                • \n
                  \u5728 HELM \u6d4b\u8bd5\u4e2d\uff0cLiang et al. (2022) \u5bf9\u4e0d\u540c\u6a21\u578b\u8fdb\u884c\u4e86\u5927\u89c4\u6a21\u8bc4\u4f30\u3002\u4ed6\u4eec\u53d1\u73b0\u4e86\u9488\u5bf9\u4ee3\u7801\u8bad\u7ec3\u7684\u6a21\u578b\u5177\u6709\u5f88\u5f3a\u7684\u8bed\u8a00\u63a8\u7406\u80fd\u529b\uff0c\u5305\u62ec 120\u4ebf\u53c2\u6570\u7684code-cushman-001.\u3002<\/span><\/section>\n<\/li>\n
                • \n
                  \u6211\u4eec\u5728 AI2 \u7684\u5de5\u4f5c\u4e5f\u8868\u660e\uff0c\u5f53\u914d\u5907\u590d\u6742\u7684\u601d\u7ef4\u94fe\u65f6\uff0ccode-davinci-002 \u5728 GSM8K \u7b49\u91cd\u8981\u6570\u5b66\u57fa\u51c6\u4e0a\u662f\u76ee\u524d\u8868\u73b0\u6700\u597d\u7684\u6a21\u578b<\/span><\/section>\n<\/li>\n
                • \n
                  \u76f4\u89c9\u6765\u8bf4\uff0c\u9762\u5411\u8fc7\u7a0b\u7684\u7f16\u7a0b (procedure-oriented programming)<\/strong> \u8ddf\u4eba\u7c7b\u9010\u6b65\u89e3\u51b3\u4efb\u52a1<\/strong>\u7684\u8fc7\u7a0b\u5f88\u7c7b\u4f3c\uff0c\u9762\u5411\u5bf9\u8c61\u7f16\u7a0b (object-oriented programming)<\/strong> \u8ddf\u4eba\u7c7b\u5c06\u590d\u6742\u4efb\u52a1\u5206\u89e3\u4e3a\u591a\u4e2a\u7b80\u5355\u4efb\u52a1<\/strong>\u7684\u8fc7\u7a0b\u5f88\u7c7b\u4f3c\u3002<\/span><\/section>\n<\/li>\n
                • \n
                  \u4ee5\u4e0a\u6240\u6709\u89c2\u5bdf\u7ed3\u679c\u90fd\u662f\u4ee3\u7801\u4e0e\u63a8\u7406\u80fd\u529b \/ \u601d\u7ef4\u94fe\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u3002\u4ee3\u7801\u548c\u63a8\u7406\u80fd\u529b \/ \u601d\u7ef4\u94fe\u4e4b\u95f4\u7684\u8fd9\u79cd\u76f8\u5173\u6027\u5bf9\u7814\u7a76\u793e\u533a\u6765\u8bf4\u662f\u4e00\u4e2a\u975e\u5e38\u6709\u8da3\u7684\u95ee\u9898\uff0c\u4f46\u76ee\u524d\u4ecd\u672a\u5f97\u5230\u5f88\u597d\u7684\u7406\u89e3\u3002\u7136\u800c\uff0c\u4ecd\u7136\u6ca1\u6709\u786e\u51ff\u7684\u8bc1\u636e\u8868\u660e\u4ee3\u7801\u8bad\u7ec3\u5c31\u662fCoT\u548c\u590d\u6742\u63a8\u7406\u7684\u539f\u56e0\u3002<\/strong> \u601d\u7ef4\u94fe\u7684\u6765\u6e90\u4ecd\u7136\u662f\u4e00\u4e2a\u5f00\u653e\u6027\u7684\u7814\u7a76\u95ee\u9898\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
                • \n
                  \u6b64\u5916\uff0c \u4ee3\u7801\u8bad\u7ec3<\/strong>\u53e6\u4e00\u4e2a\u53ef\u80fd\u7684\u526f\u4ea7\u54c1\u662f\u957f\u8ddd\u79bb\u4f9d\u8d56<\/strong>\uff0c\u6b63\u5982Peter Liu\u6240\u6307\u51fa\uff1a\u201c\u8bed\u8a00\u4e2d\u7684\u4e0b\u4e2a\u8bcd\u8bed\u9884\u6d4b\u901a\u5e38\u662f\u975e\u5e38\u5c40\u90e8\u7684\uff0c\u800c\u4ee3\u7801\u901a\u5e38\u9700\u8981\u66f4\u957f\u7684\u4f9d\u8d56\u5173\u7cfb\u6765\u505a\u4e00\u4e9b\u4e8b\u60c5\uff0c\u6bd4\u5982\u524d\u540e\u62ec\u53f7\u7684\u5339\u914d\u6216\u5f15\u7528\u8fdc\u5904\u7684\u51fd\u6570\u5b9a\u4e49\u201d\u3002\u8fd9\u91cc\u6211\u60f3\u8fdb\u4e00\u6b65\u8865\u5145\u7684\u662f\uff1a\u7531\u4e8e\u9762\u5411\u5bf9\u8c61\u7f16\u7a0b\u4e2d\u7684\u7c7b\u7ee7\u627f\uff0c\u4ee3\u7801\u4e5f\u53ef\u80fd\u6709\u52a9\u4e8e\u6a21\u578b\u5efa\u7acb\u7f16\u7801\u5c42\u6b21\u7ed3\u6784\u7684\u80fd\u529b\u3002\u6211\u4eec\u5c06\u5bf9\u8fd9\u4e00\u5047\u8bbe\u7684\u68c0\u9a8c\u7559\u7ed9\u672a\u6765\u7684\u5de5\u4f5c\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

                  \u53e6\u5916\u8fd8\u8981\u6ce8\u610f\u4e00\u4e9b\u7ec6\u8282\u5dee\u5f02\uff1a<\/span><\/p>\n

                    \n
                  • \n
                    text-davinci-002 \u4e0e code-davinci-002<\/strong><\/span><\/section>\n<\/li>\n
                      \n
                    • \n
                      Code-davinci-002 \u662f\u57fa\u7840\u6a21\u578b\uff0ctext-davinci-002 \u662f\u6307\u4ee4\u5fae\u8c03 code-davinci-002 \u7684\u4ea7\u7269\uff08\u89c1 OpenAI \u7684\u6587\u6863\uff09\u3002\u5b83\u5728\u4ee5\u4e0b\u6570\u636e\u4e0a\u4f5c\u4e86\u5fae\u8c03\uff1a\uff08\u4e00\uff09\u4eba\u5de5\u6807\u6ce8\u7684\u6307\u4ee4\u548c\u671f\u5f85\u7684\u8f93\u51fa\uff1b\uff08\u4e8c\uff09\u7531\u4eba\u5de5\u6807\u6ce8\u8005\u9009\u62e9\u7684\u6a21\u578b\u8f93\u51fa\u3002<\/span><\/section>\n<\/li>\n
                    • \n
                      \u5f53\u6709\u4e0a\u4e0b\u6587\u793a\u4f8b (in-context example) \u7684\u65f6\u5019\uff0c Code-davinci-002 \u66f4\u64c5\u957f\u4e0a\u4e0b\u6587\u5b66\u4e60\uff1b\u5f53\u6ca1\u6709\u4e0a\u4e0b\u6587\u793a\u4f8b \/ \u96f6\u6837\u672c\u7684\u65f6\u5019\uff0c text-davinci-002 \u5728\u96f6\u6837\u672c\u4efb\u52a1\u5b8c\u6210\u65b9\u9762\u8868\u73b0\u66f4\u597d\u3002\u4ece\u8fd9\u4e2a\u610f\u4e49\u4e0a\u8bf4\uff0ctext-davinci-002 \u66f4\u7b26\u5408\u4eba\u7c7b\u7684\u671f\u5f85\uff08\u56e0\u4e3a\u5bf9\u4e00\u4e2a\u4efb\u52a1\u5199\u4e0a\u4e0b\u6587\u793a\u4f8b\u53ef\u80fd\u4f1a\u6bd4\u8f83\u9ebb\u70e6\uff09\u3002<\/span><\/section>\n<\/li>\n
                    • \n
                      OpenAI \u4e0d\u592a\u53ef\u80fd\u6545\u610f\u727a\u7272\u4e86\u4e0a\u4e0b\u6587\u5b66\u4e60\u7684\u80fd\u529b\u6362\u53d6\u96f6\u6837\u672c\u80fd\u529b \u2014\u2014 \u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd\u529b\u7684\u964d\u4f4e\u66f4\u591a\u662f\u6307\u4ee4\u5b66\u4e60\u7684\u4e00\u4e2a\u526f\u4f5c\u7528\uff0cOpenAI \u7ba1\u8fd9\u53eb\u5bf9\u9f50\u7a0e\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
                    • \n
                      001 \u6a21\u578b\uff08code-cushman-001 \u548c text-davinci-001\uff09v.s. 002 \u6a21\u578b\uff08code-davinci-002 \u548c text-davinci-002\uff09<\/strong><\/span><\/section>\n<\/li>\n
                        \n
                      • \n
                        001 \u6a21\u578b\u4e3b\u8981\u662f\u4e3a\u4e86\u505a\u7eaf\u4ee3\u7801 \/ \u7eaf\u6587\u672c\u4efb\u52a1\uff1b002 \u6a21\u578b\u5219\u6df1\u5ea6\u878d\u5408\u4e86\u4ee3\u7801\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\uff0c\u4ee3\u7801\u548c\u6587\u672c\u90fd\u884c\u3002<\/span><\/section>\n<\/li>\n
                      • \n
                        Code-davinci-002 \u53ef\u80fd\u662f\u7b2c\u4e00\u4e2a\u6df1\u5ea6\u878d\u5408\u4e86\u4ee3\u7801\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u7684\u6a21\u578b\u3002\u8bc1\u636e\u6709\uff1acode-cushman-001 \u53ef\u4ee5\u8fdb\u884c\u63a8\u7406\u4f46\u5728\u7eaf\u6587\u672c\u4e0a\u8868\u73b0\u4e0d\u4f73\uff0ctext-davinci-001 \u5728\u7eaf\u6587\u672c\u4e0a\u8868\u73b0\u4e0d\u9519\u4f46\u5728\u63a8\u7406\u4e0a\u4e0d\u5927\u884c\u3002code-davinci-002 \u5219\u53ef\u4ee5\u540c\u65f6\u505a\u5230\u8fd9\u4e24\u70b9\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n

                        3.2 \u8fd9\u4e9b\u80fd\u529b\u662f\u5728\u9884\u8bad\u7ec3\u4e4b\u540e\u5df2\u7ecf\u5b58\u5728\u8fd8\u662f\u5728\u4e4b\u540e\u901a\u8fc7\u5fae\u8c03\u6ce8\u5165\uff1f<\/strong><\/span><\/h2>\n

                        \u5728\u8fd9\u4e2a\u9636\u6bb5\uff0c\u6211\u4eec\u5df2\u7ecf\u786e\u5b9a\u4e86\u6307\u4ee4\u5fae\u8c03\u548c\u4ee3\u7801\u8bad\u7ec3\u7684\u5173\u952e\u4f5c\u7528\u3002\u4e00\u4e2a\u91cd\u8981\u7684\u95ee\u9898\u662f\u5982\u4f55\u8fdb\u4e00\u6b65\u5206\u6790\u4ee3\u7801\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u7684\u5f71\u54cd\uff1f\u5177\u4f53\u6765\u8bf4\uff1a\u4e0a\u8ff0\u4e09\u79cd\u80fd\u529b\u662f\u5426\u5df2\u7ecf\u5b58\u5728\u4e8e\u521d\u4ee3\u7684GPT-3<\/strong>\u4e2d\uff0c\u53ea\u662f\u901a\u8fc7\u6307\u4ee4\u548c\u4ee3\u7801\u8bad\u7ec3\u89e6\u53d1 \/ \u89e3\u9501<\/strong>\uff1f\u6216\u8005\u8fd9\u4e9b\u80fd\u529b\u5728\u521d\u4ee3\u7684 GPT-3 \u4e2d\u5e76\u4e0d\u5b58\u5728<\/strong>\uff0c\u662f\u901a\u8fc7\u6307\u4ee4\u548c\u4ee3\u7801\u8bad\u7ec3\u6ce8\u5165\uff1f<\/strong>\u5982\u679c\u7b54\u6848\u5df2\u7ecf\u5728\u521d\u4ee3\u7684 GPT-3 \u4e2d\uff0c\u90a3\u4e48\u8fd9\u4e9b\u80fd\u529b\u4e5f\u5e94\u8be5\u5728 OPT \u4e2d\u3002\u56e0\u6b64\uff0c\u8981\u590d\u73b0\u8fd9\u4e9b\u80fd\u529b\uff0c\u6216\u8bb8\u53ef\u4ee5\u76f4\u63a5\u901a\u8fc7\u6307\u4ee4\u548c\u4ee3\u7801\u8c03\u6574 OPT\u3002<\/strong> \u4f46\u662f\uff0ccode-davinci-002 \u4e5f\u53ef\u80fd\u4e0d\u662f\u57fa\u4e8e\u6700\u521d\u7684 GPT-3 davinci\uff0c\u800c\u662f\u57fa\u4e8e\u6bd4\u521d\u4ee3 GPT-3 \u66f4\u5927\u7684\u6a21\u578b\u3002\u5982\u679c\u662f\u8fd9\u79cd\u60c5\u51b5\uff0c\u53ef\u80fd\u5c31\u6ca1\u529e\u6cd5\u901a\u8fc7\u8c03\u6574 OPT \u6765\u590d\u73b0\u4e86\u3002\u7814\u7a76\u793e\u533a\u9700\u8981\u8fdb\u4e00\u6b65\u5f04\u6e05\u695a OpenAI \u8bad\u7ec3\u4e86\u4ec0\u4e48\u6837\u7684\u6a21\u578b\u4f5c\u4e3a code-davinci-002 \u7684\u57fa\u7840\u6a21\u578b\u3002<\/span><\/p>\n

                        \u6211\u4eec\u6709\u4ee5\u4e0b\u7684\u5047\u8bbe\u548c\u8bc1\u636e\uff1a<\/span><\/p>\n

                          \n
                        • \n
                          code-davinci-002\u7684\u57fa\u7840\u6a21\u578b\u53ef\u80fd\u4e0d\u662f\u521d\u4ee3GPT-3 davinci \u6a21\u578b<\/strong>\u3002\u4ee5\u4e0b\u662f\u8bc1\u636e\uff1a<\/span><\/section>\n<\/li>\n
                            \n
                          • \n
                            \u521d\u4ee3\u7684GPT-3\u5728\u6570\u636e\u96c6 C4 2016 – 2019 \u4e0a\u8bad\u7ec3\uff0c\u800c code-davinci-002 \u8bad\u7ec3\u96c6\u5219\u5728\u5ef6\u957f\u52302021\u5e74\u624d\u7ed3\u675f\u3002\u56e0\u6b64 code-davinci-002 \u6709\u53ef\u80fd\u5728 C4 \u7684 2019-2021 \u7248\u672c\u4e0a\u8bad\u7ec3\u3002<\/span><\/section>\n<\/li>\n
                          • \n
                            \u521d\u4ee3\u7684 GPT-3 \u6709\u4e00\u4e2a\u5927\u5c0f\u4e3a 2048<\/strong> \u4e2a\u8bcd\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002code-davinci-002 \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u5219\u4e3a 8192<\/strong>\u3002GPT \u7cfb\u5217\u4f7f\u7528\u7edd\u5bf9\u4f4d\u7f6e\u5d4c\u5165 (absolute positional embedding)\uff0c\u76f4\u63a5\u5bf9\u7edd\u5bf9\u4f4d\u7f6e\u5d4c\u5165\u8fdb\u884c\u5916\u63a8\u800c\u4e0d\u7ecf\u8fc7\u8bad\u7ec3\u662f\u6bd4\u8f83\u96be\u7684\uff0c\u5e76\u4e14\u4f1a\u4e25\u91cd\u635f\u5bb3\u6a21\u578b\u7684\u6027\u80fd\uff08\u53c2\u8003 Press et al., 2022\uff09\u3002\u5982\u679c code-davinci-002 \u662f\u57fa\u4e8e\u521d\u4ee3GPT-3\uff0c\u90a3OpenAI \u662f\u5982\u4f55\u6269\u5c55\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\uff1f<\/span><\/section>\n<\/li>\n<\/ul>\n
                          • \n
                            \u53e6\u4e00\u65b9\u9762\uff0c\u65e0\u8bba\u57fa\u7840\u6a21\u578b\u662f\u521d\u4ee3\u7684 GPT-3 \u8fd8\u662f\u540e\u6765\u8bad\u7ec3\u7684\u6a21\u578b\uff0c \u9075\u5faa\u6307\u4ee4\u548c\u96f6\u6837\u672c\u6cdb\u5316\u7684\u80fd\u529b\u90fd\u53ef\u80fd\u5df2\u7ecf\u5b58\u5728\u4e8e\u57fa\u7840\u6a21\u578b<\/strong>\u4e2d\uff0c\u540e\u6765\u624d\u901a\u8fc7\u6307\u4ee4\u5fae\u8c03\u6765\u89e3\u9501<\/strong> \uff08\u800c\u4e0d\u662f\u6ce8\u5165\uff09<\/strong><\/span><\/section>\n<\/li>\n
                              \n
                            • \n
                              \u8fd9\u4e3b\u8981\u662f\u56e0\u4e3a OpenAI \u7684\u8bba\u6587\u62a5\u544a\u7684\u6307\u4ee4\u6570\u636e\u91cf\u5927\u5c0f\u53ea\u6709 77K\uff0c\u6bd4\u9884\u8bad\u7ec3\u6570\u636e\u5c11\u4e86\u51e0\u4e2a\u6570\u91cf\u7ea7\u3002<\/span><\/section>\n<\/li>\n
                            • \n
                              \u5176\u4ed6\u6307\u4ee4\u5fae\u8c03\u8bba\u6587\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u6570\u636e\u96c6\u5927\u5c0f\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u5bf9\u6bd4\uff0c\u4f8b\u5982 Chung et al. (2022) \u7684\u5de5\u4f5c\u4e2d\uff0c Flan-PaLM \u7684\u6307\u4ee4\u5fae\u8c03\u4ec5\u4e3a\u9884\u8bad\u7ec3\u8ba1\u7b97\u7684 0.4%\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u6307\u4ee4\u6570\u636e\u4f1a\u663e\u8457\u5c11\u4e8e\u9884\u8bad\u7ec3\u6570\u636e\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
                            • \n
                              \u7136\u800c \uff0c\u6a21\u578b\u7684\u590d\u6742\u63a8\u7406\u80fd\u529b\u53ef\u80fd\u662f\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u901a\u8fc7\u4ee3\u7801\u6570\u636e\u6ce8\u5165<\/strong><\/span><\/section>\n<\/li>\n
                                \n
                              • \n
                                \u4ee3\u7801\u6570\u636e\u96c6\u7684\u89c4\u6a21\u4e0e\u4e0a\u8ff0\u6307\u4ee4\u5fae\u8c03\u7684\u60c5\u51b5\u4e0d\u540c\u3002\u8fd9\u91cc\u7684\u4ee3\u7801\u6570\u636e\u91cf\u8db3\u591f\u5927\uff0c\u53ef\u4ee5\u5360\u636e\u8bad\u7ec3\u6570\u636e\u7684\u91cd\u8981\u90e8\u5206\uff08\u4f8b\u5982\uff0cPaLM \u6709 8% \u7684\u4ee3\u7801\u8bad\u7ec3\u6570\u636e\uff09<\/span><\/section>\n<\/li>\n
                              • \n
                                \u5982\u4e0a\u6240\u8ff0\uff0c\u5728 code-davinci-002 \u4e4b\u524d\u7684\u6a21\u578b text-davinci-001 \u5927\u6982\u6ca1\u6709\u5728\u4ee3\u7801\u6570\u636e\u4e0a\u9762\u5fae\u8c03\u8fc7\uff0c\u6240\u4ee5\u5b83\u7684\u63a8\u7406 \/ \u601d\u7ef4\u94fe\u80fd\u529b\u662f\u975e\u5e38\u5dee\u7684\uff0c\u6b63\u5982\u7b2c\u4e00\u7248\u601d\u7ef4\u94fe\u8bba\u6587\u4e2d\u6240\u62a5\u544a\u7684\u90a3\u6837\uff0c\u6709\u65f6\u751a\u81f3\u6bd4\u53c2\u6570\u91cf\u66f4\u5c0f\u7684 code-cushman-001 \u8fd8\u5dee\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
                              • \n
                                \u533a\u5206\u4ee3\u7801\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u6548\u679c\u7684\u6700\u597d\u65b9\u6cd5<\/strong>\u53ef\u80fd\u662f\u6bd4\u8f83 code-cushman-001\u3001T5 \u548c FlanT5<\/strong><\/span><\/section>\n<\/li>\n
                                  \n
                                • \n
                                  \u56e0\u4e3a\u5b83\u4eec\u5177\u6709\u76f8\u4f3c\u7684\u6a21\u578b\u5927\u5c0f\uff08110\u4ebf \u548c 120\u4ebf\uff09\uff0c\u76f8\u4f3c\u7684\u8bad\u7ec3\u6570\u636e\u96c6 (C4)\uff0c\u5b83\u4eec\u6700\u5927\u7684\u533a\u522b\u5c31\u662f\u6709\u6ca1\u6709\u5728\u4ee3\u7801\u4e0a\u8bad\u7ec3\u8fc7 \/ \u6709\u6ca1\u6709\u505a\u8fc7\u6307\u4ee4\u5fae\u8c03\u3002<\/span><\/section>\n<\/li>\n
                                • \n
                                  \u76ee\u524d\u8fd8\u6ca1\u6709\u8fd9\u6837\u7684\u6bd4\u8f83\u3002\u6211\u4eec\u628a\u8fd9\u4e2a\u7559\u7ed9\u672a\u6765\u7684\u7814\u7a76\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n


                                  <\/p>\n

                                  \n
                                  \n
                                  \n
                                  \n
                                  \n

                                  <\/section>\n

                                  <\/section>\n
                                  \n
                                  <\/svg><\/section>\n<\/section>\n<\/section>\n
                                  \n

                                  4. text-davinci-003 \u548c ChatGPT\uff0c<\/span><\/strong><\/strong><\/p>\n

                                  \u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60<\/span><\/strong><\/strong><\/p>\n

                                  (Reinforcement Learning from Human Feedback, RLHF) \u7684\u5a01\u529b<\/span><\/strong><\/strong><\/p>\n<\/section>\n

                                  \n

                                  <\/section>\n

                                  <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
                                   <\/span><\/section>\n

                                  \u5728\u5f53\u524d\u9636\u6bb5\uff082022 \u5e74 12 \u6708\uff09\uff0c text-davinci-002\u3001text-davinci-003 \u548c ChatGPT\u4e4b\u95f4\u51e0\u4e4e\u6ca1\u6709\u4e25\u683c\u7684\u7edf\u8ba1\u4e0a\u7684\u6bd4\u8f83<\/strong> \uff0c\u4e3b\u8981\u662f\u56e0\u4e3a<\/span><\/p>\n

                                    \n
                                  • \n
                                    text-davinci-003 \u548c ChatGPT \u5728\u64b0\u5199\u672c\u6587\u65f6\u624d\u53d1\u5e03\u4e0d\u5230\u4e00\u4e2a\u6708\u3002<\/span><\/section>\n<\/li>\n
                                  • \n
                                    ChatGPT \u4e0d\u80fd\u901a\u8fc7 OpenAI API \u88ab\u8c03\u7528\uff0c\u6240\u4ee5\u60f3\u8981\u5728\u6807\u51c6\u57fa\u51c6\u4e0a\u6d4b\u8bd5\u5b83\u5f88\u9ebb\u70e6\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

                                    \u6240\u4ee5\u5728\u8fd9\u4e9b\u6a21\u578b\u4e4b\u95f4\u7684\u6bd4\u8f83\u66f4\u591a\u662f\u57fa\u4e8e\u7814\u7a76\u793e\u533a\u7684\u96c6\u4f53\u7ecf\u9a8c<\/strong> \uff08\u7edf\u8ba1\u4e0a\u4e0d\u662f\u5f88\u4e25\u683c\uff09\u3002\u4e0d\u8fc7\uff0c\u6211\u4eec\u76f8\u4fe1\u521d\u6b65\u7684\u63cf\u8ff0\u6027\u6bd4\u8f83\u4ecd\u7136\u53ef\u4ee5\u63ed\u793a\u6a21\u578b\u7684\u673a\u5236\u3002<\/span><\/p>\n

                                    \u6211\u4eec\u9996\u5148\u6ce8\u610f\u5230\u4ee5\u4e0b text-davinci-002\uff0ctext-davinci-003 \u548c ChatGPT \u4e4b\u95f4\u7684\u6bd4\u8f83\uff1a<\/span><\/p>\n

                                      \n
                                    • \n
                                      \u6240\u6709\u4e09\u4e2a\u6a21\u578b\u90fd\u7ecf\u8fc7\u6307\u4ee4\u5fae\u8c03<\/strong>\u3002<\/span><\/section>\n<\/li>\n
                                    • \n
                                      text-davinci-002<\/strong> \u662f\u4e00\u4e2a\u7ecf\u8fc7\u76d1\u7763\u5b66\u4e60\u6307\u4ee4\u5fae\u8c03<\/strong> (supervised instruction tuning) \u7684\u6a21\u578b<\/span><\/section>\n<\/li>\n
                                    • \n
                                      text-davinci-003 \u548c ChatGPT<\/strong> \u662f\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\u7684\u6307\u4ee4\u5fae\u8c03<\/strong> (Instruction tuning with Reinforcement Learning from Human Feedback RLHF)\u3002\u8fd9\u662f\u5b83\u4eec\u4e4b\u95f4\u6700\u663e\u7740\u7684\u533a\u522b\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

                                      \u8fd9\u610f\u5473\u7740\u5927\u591a\u6570\u65b0\u6a21\u578b\u7684\u884c\u4e3a\u90fd\u662f RLHF \u7684\u4ea7\u7269<\/strong>\u3002<\/span><\/p>\n

                                      \u90a3\u4e48\u8ba9\u6211\u4eec\u770b\u770b RLHF \u89e6\u53d1\u7684\u80fd\u529b\uff1a<\/span><\/p>\n

                                        \n
                                      • \n
                                        \u7fd4\u5b9e\u7684\u56de\u5e94<\/strong>\uff1atext-davinci-003 \u7684\u751f\u6210\u901a\u5e38\u6bd4 text-davinci-002\u957f\u3002ChatGPT \u7684\u56de\u5e94\u5219\u66f4\u52a0\u5197\u957f\uff0c\u4ee5\u81f3\u4e8e\u7528\u6237\u5fc5\u987b\u660e\u786e\u8981\u6c42\u201c\u7528\u4e00\u53e5\u8bdd\u56de\u7b54\u6211\u201d\uff0c\u624d\u80fd\u5f97\u5230\u66f4\u52a0\u7b80\u6d01\u7684\u56de\u7b54\u3002\u8fd9\u662f RLHF \u7684\u76f4\u63a5\u4ea7\u7269\u3002<\/span><\/section>\n<\/li>\n
                                      • \n
                                        \u516c\u6b63\u7684\u56de\u5e94<\/strong>\uff1aChatGPT \u901a\u5e38\u5bf9\u6d89\u53ca\u591a\u4e2a\u5b9e\u4f53\u5229\u76ca\u7684\u4e8b\u4ef6\uff08\u4f8b\u5982\u653f\u6cbb\u4e8b\u4ef6\uff09\u7ed9\u51fa\u975e\u5e38\u5e73\u8861\u7684\u56de\u7b54\u3002\u8fd9\u4e5f\u662fRLHF\u7684\u4ea7\u7269\u3002<\/span><\/section>\n<\/li>\n
                                      • \n
                                        \u62d2\u7edd\u4e0d\u5f53\u95ee\u9898<\/strong>\uff1a\u8fd9\u662f\u5185\u5bb9\u8fc7\u6ee4\u5668\u548c\u7531 RLHF \u89e6\u53d1\u7684\u6a21\u578b\u81ea\u8eab\u80fd\u529b\u7684\u7ed3\u5408\uff0c\u8fc7\u6ee4\u5668\u8fc7\u6ee4\u6389\u4e00\u90e8\u5206\uff0c\u7136\u540e\u6a21\u578b\u518d\u62d2\u7edd\u4e00\u90e8\u5206\u3002<\/span><\/section>\n<\/li>\n
                                      • \n
                                        \u62d2\u7edd\u5176\u77e5\u8bc6\u8303\u56f4\u4e4b\u5916\u7684\u95ee\u9898<\/strong>\uff1a\u4f8b\u5982\uff0c\u62d2\u7edd\u57282021 \u5e74 6 \u6708\u4e4b\u540e\u53d1\u751f\u7684\u65b0\u4e8b\u4ef6\uff08\u56e0\u4e3a\u5b83\u6ca1\u5728\u8fd9\u4e4b\u540e\u7684\u6570\u636e\u4e0a\u8bad\u7ec3\u8fc7\uff09\u3002\u8fd9\u662f RLHF \u6700\u795e\u5947\u7684\u90e8\u5206\uff0c\u56e0\u4e3a\u5b83\u4f7f\u6a21\u578b\u80fd\u591f\u9690\u5f0f\u5730\u533a\u5206\u54ea\u4e9b\u95ee\u9898\u5728\u5176\u77e5\u8bc6\u8303\u56f4\u5185\uff0c\u54ea\u4e9b\u95ee\u9898\u4e0d\u5728\u5176\u77e5\u8bc6\u8303\u56f4\u5185\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

                                        \u6709\u4e24\u4ef6\u4e8b\u60c5\u503c\u5f97\u6ce8\u610f\uff1a<\/span><\/p>\n

                                          \n
                                        • \n
                                          \u6240\u6709\u7684\u80fd\u529b\u90fd\u662f\u6a21\u578b\u672c\u6765\u5c31\u6709\u7684\uff0c \u800c\u4e0d\u662f\u901a\u8fc7RLHF \u6ce8\u5165\u7684<\/strong>\u3002RLHF \u7684\u4f5c\u7528\u662f\u89e6\u53d1 \/ \u89e3\u9501\u7a81\u73b0\u80fd\u529b<\/strong>\u3002\u8fd9\u4e2a\u8bba\u70b9\u4e3b\u8981\u6765\u81ea\u4e8e\u6570\u636e\u91cf\u5927\u5c0f\u7684\u6bd4\u8f83\uff1a\u56e0\u4e3a\u4e0e\u9884\u8bad\u7ec3\u7684\u6570\u636e\u91cf\u76f8\u6bd4\uff0cRLHF \u5360\u7528\u7684\u8ba1\u7b97\u91cf \/ \u6570\u636e\u91cf\u8981\u5c11\u5f97\u591a\u3002<\/span><\/section>\n<\/li>\n
                                        • \n
                                          \u6a21\u578b\u77e5\u9053\u5b83\u4e0d\u77e5\u9053\u4ec0\u4e48\u4e0d\u662f\u901a\u8fc7\u7f16\u5199\u89c4\u5219\u6765\u5b9e\u73b0\u7684\uff0c<\/strong> \u800c\u662f\u901a\u8fc7RLHF\u89e3\u9501\u7684\u3002\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u4ee4\u4eba\u60ca\u8bb6\u7684\u53d1\u73b0\uff0c\u56e0\u4e3a RLHF \u7684\u6700\u521d\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u751f\u6210\u590d\u5408\u4eba\u7c7b\u671f\u671b\u7684\u56de\u7b54\uff0c\u8fd9\u66f4\u591a\u662f\u8ba9\u6a21\u578b\u751f\u6210\u5b89\u5168\u7684\u53e5\u5b50\uff0c\u800c\u4e0d\u662f\u8ba9\u6a21\u578b\u77e5\u9053\u5b83\u4e0d\u77e5\u9053\u7684\u5185\u5bb9\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

                                          \u5e55\u540e\u53d1\u751f\u7684\u4e8b\u60c5\u53ef\u80fd\u662f\uff1a<\/span><\/p>\n

                                            \n
                                          • \n
                                            ChatGPT: \u901a\u8fc7\u727a\u7272\u4e0a\u4e0b\u6587\u5b66\u4e60<\/strong>\u7684\u80fd\u529b\u6362\u53d6\u5efa\u6a21\u5bf9\u8bdd\u5386\u53f2<\/strong>\u7684\u80fd\u529b\u3002\u8fd9\u662f\u4e00\u4e2a\u57fa\u4e8e\u7ecf\u9a8c\u7684\u89c2\u6d4b\u7ed3\u679c\uff0c\u56e0\u4e3a ChatGPT \u4f3c\u4e4e\u4e0d\u50cf text-davinci-003 \u90a3\u6837\u53d7\u5230\u4e0a\u4e0b\u6587\u6f14\u793a\u7684\u5f3a\u70c8\u5f71\u54cd\u3002<\/span><\/section>\n<\/li>\n
                                          • \n
                                            text-davinci-003\uff1a\u6062\u590d\u4e86<\/strong> text-davinci-002 \u6240\u727a\u7272\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd\u529b<\/strong>\uff0c \u63d0\u9ad8\u96f6\u6837\u672c\u7684\u80fd\u529b<\/strong>\u3002\u6211\u4eec\u4e0d\u786e\u5b9a\u8fd9\u662f\u5426\u4e5f\u662f RLHF \u6216\u5176\u4ed6\u4e1c\u897f\u7684\u526f\u4ea7\u54c1\u3002<\/span> \u6839\u636einstructGPT\u7684\u8bba\u6587\uff0c\u8fd9\u662f\u6765\u81ea\u4e8e\u5f3a\u5316\u5b66\u4e60\u8c03\u6574\u9636\u6bb5\u6df7\u5165\u4e86\u8bed\u8a00\u5efa\u6a21\u7684\u76ee\u6807\uff08\u800c\u4e0d\u662f RLHF \u672c\u8eab\uff09\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n


                                            <\/p>\n

                                            \n
                                            \n
                                            \n
                                            \n
                                            \n

                                            <\/section>\n

                                            <\/section>\n
                                            \n
                                            <\/svg><\/section>\n<\/section>\n<\/section>\n
                                            \n

                                            5. \u603b\u7ed3\u5f53\u524d\u9636\u6bb5GPT-3.5\u7684\u8fdb\u5316\u5386\u7a0b<\/span><\/strong><\/strong><\/p>\n<\/section>\n

                                            \n

                                            <\/section>\n

                                            <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
                                             <\/span><\/section>\n

                                            \u5230\u76ee\u524d\u4e3a\u6b62\uff0c\u6211\u4eec\u5df2\u7ecf\u4ed4\u7ec6\u68c0\u67e5\u4e86\u6cbf\u7740\u8fdb\u5316\u6811\u51fa\u73b0\u7684\u6240\u6709\u80fd\u529b\uff0c\u4e0b\u8868\u603b\u7ed3\u4e86\u6f14\u5316\u8def\u5f84\uff1a<\/span><\/p>\n

                                            <\/figure>\n

                                            <\/section>\n

                                            \u6211\u4eec\u53ef\u4ee5\u5f97\u51fa\u7ed3\u8bba\uff1a<\/span><\/p>\n

                                              \n
                                            • \n
                                              \u8bed\u8a00\u751f\u6210\u80fd\u529b + \u57fa\u7840\u4e16\u754c\u77e5\u8bc6 + \u4e0a\u4e0b\u6587\u5b66\u4e60\u90fd\u662f\u6765\u81ea\u4e8e\u9884\u8bad\u7ec3\uff08<\/span>davinci<\/span><\/code>\uff09<\/span><\/section>\n<\/li>\n
                                            • \n
                                              \u5b58\u50a8\u5927\u91cf\u77e5\u8bc6\u7684\u80fd\u529b\u6765\u81ea 1750 \u4ebf\u7684\u53c2\u6570\u91cf\u3002<\/span><\/section>\n<\/li>\n
                                            • \n
                                              \u9075\u5faa\u6307\u4ee4\u548c\u6cdb\u5316\u5230\u65b0\u4efb\u52a1\u7684\u80fd\u529b\u6765\u81ea\u4e8e\u6269\u5927\u6307\u4ee4\u5b66\u4e60\u4e2d\u6307\u4ee4\u7684\u6570\u91cf\uff08<\/span>Davinci-instruct-beta<\/span><\/code>)<\/span><\/section>\n<\/li>\n
                                            • \n
                                              \u6267\u884c\u590d\u6742\u63a8\u7406\u7684\u80fd\u529b\u5f88\u53ef\u80fd\u6765\u81ea\u4e8e\u4ee3\u7801\u8bad\u7ec3\uff08<\/span>code-davinci-002<\/span><\/code>\uff09<\/span><\/section>\n<\/li>\n
                                            • \n
                                              \u751f\u6210\u4e2d\u7acb\u3001\u5ba2\u89c2\u7684\u80fd\u529b\u3001\u5b89\u5168\u548c\u7fd4\u5b9e\u7684\u7b54\u6848\u6765\u81ea\u4e0e\u4eba\u7c7b\u7684\u5bf9\u9f50\u3002\u5177\u4f53\u6765\u8bf4\uff1a<\/span><\/section>\n<\/li>\n
                                                \n
                                              • \n
                                                \u5982\u679c\u662f\u76d1\u7763\u5b66\u4e60\u7248\uff0c\u5f97\u5230\u7684\u6a21\u578b\u662f<\/span>text-davinci-002<\/span><\/code><\/section>\n<\/li>\n
                                              • \n
                                                \u5982\u679c\u662f\u5f3a\u5316\u5b66\u4e60\u7248 (RLHF) \uff0c\u5f97\u5230\u7684\u6a21\u578b\u662f<\/span>text-davinci-003<\/span><\/code><\/section>\n<\/li>\n
                                              • \n
                                                \u65e0\u8bba\u662f\u6709\u76d1\u7763\u8fd8\u662f RLHF \uff0c\u6a21\u578b\u5728\u5f88\u591a\u4efb\u52a1\u7684\u6027\u80fd\u90fd\u65e0\u6cd5\u8d85\u8fc7 code-davinci-002 \uff0c\u8fd9\u79cd\u56e0\u4e3a\u5bf9\u9f50\u800c\u9020\u6210\u6027\u80fd\u8870\u9000\u7684\u73b0\u8c61\u53eb\u505a\u5bf9\u9f50\u7a0e\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
                                              • \n
                                                \u5bf9\u8bdd\u80fd\u529b\u4e5f\u6765\u81ea\u4e8e RLHF\uff08<\/span>ChatGPT<\/span><\/code>\uff09\uff0c\u5177\u4f53\u6765\u8bf4\u5b83\u727a\u7272\u4e86\u4e0a\u4e0b\u6587\u5b66\u4e60\u7684\u80fd\u529b\uff0c\u6765\u6362\u53d6\uff1a<\/span><\/section>\n<\/li>\n
                                                  \n
                                                • \n
                                                  \u5efa\u6a21\u5bf9\u8bdd\u5386\u53f2<\/span><\/section>\n<\/li>\n
                                                • \n
                                                  \u589e\u52a0\u5bf9\u8bdd\u4fe1\u606f\u91cf<\/span><\/section>\n<\/li>\n
                                                • \n
                                                  \u62d2\u7edd\u6a21\u578b\u77e5\u8bc6\u8303\u56f4\u4e4b\u5916\u7684\u95ee\u9898<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n


                                                  <\/p>\n

                                                  \n
                                                  \n
                                                  \n
                                                  \n
                                                  \n

                                                  <\/section>\n

                                                  <\/section>\n
                                                  \n
                                                  <\/svg><\/section>\n<\/section>\n<\/section>\n
                                                  \n

                                                  6. GPT-3.5\u76ee\u524d\u4e0d\u80fd\u505a\u4ec0\u4e48<\/span><\/strong><\/strong><\/p>\n<\/section>\n

                                                  \n

                                                  <\/section>\n

                                                  <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
                                                   <\/span><\/section>\n

                                                  \u867d\u7136GPT-3.5\u662f\u81ea\u7136\u8bed\u8a00\u5904\u7406\u7814\u7a76\u4e2d\u7684\u91cd\u8981\u4e00\u6b65\uff0c\u4f46\u5b83\u5e76\u6ca1\u6709\u5b8c\u5168\u5305\u542b\u8bb8\u591a\u7814\u7a76\u4eba\u5458\uff08\u5305\u62ec AI2\uff09\u8bbe\u60f3\u7684\u6240\u6709\u7406\u60f3\u5c5e\u6027\u3002\u4ee5\u4e0b\u662fGPT-3.5\u4e0d\u5177\u5907\u7684\u67d0\u4e9b\u91cd\u8981\u5c5e\u6027\uff1a<\/span><\/p>\n

                                                    \n
                                                  • \n
                                                    \u5b9e\u65f6\u6539\u5199\u6a21\u578b\u7684\u4fe1\u5ff5<\/strong>\uff1a\u5f53\u6a21\u578b\u8868\u8fbe\u5bf9\u67d0\u4e8b\u7684\u4fe1\u5ff5\u65f6\uff0c\u5982\u679c\u8be5\u4fe1\u5ff5\u662f\u9519\u8bef\u7684\uff0c\u6211\u4eec\u53ef\u80fd\u5f88\u96be\u7ea0\u6b63\u5b83\uff1a<\/span><\/section>\n<\/li>\n
                                                      \n
                                                    • \n
                                                      \u6211\u6700\u8fd1\u9047\u5230\u7684\u4e00\u4e2a\u4f8b\u5b50\u662f\uff1aChatGPT \u575a\u6301\u8ba4\u4e3a 3599 \u662f\u4e00\u4e2a\u8d28\u6570\uff0c\u5c3d\u7ba1\u5b83\u627f\u8ba4 3599 = 59 * 61\u3002\u53e6\u5916\uff0c\u8bf7\u53c2\u9605Reddit\u4e0a\u5173\u4e8e\u6e38\u5f97\u6700\u5feb\u7684\u6d77\u6d0b\u54fa\u4e73\u52a8\u7269\u7684\u4f8b\u5b50\u3002<\/span><\/section>\n<\/li>\n
                                                    • \n
                                                      \u7136\u800c\uff0c\u6a21\u578b\u4fe1\u5ff5\u7684\u5f3a\u5ea6\u4f3c\u4e4e\u5b58\u5728\u4e0d\u540c\u7684\u5c42\u6b21\u3002\u4e00\u4e2a\u4f8b\u5b50\u662f\u5373\u4f7f\u6211\u544a\u8bc9\u5b83\u8fbe\u65af\u00b7\u7ef4\u8fbe\uff08\u661f\u7403\u5927\u6218\u7535\u5f71\u4e2d\u7684\u4eba\u7269\uff09\u8d62\u5f97\u4e862020\u5e74\u5927\u9009\uff0c\u6a21\u578b\u4f9d\u65e7\u4f1a\u8ba4\u4e3a\u7f8e\u56fd\u73b0\u4efb\u603b\u7edf\u662f\u62dc\u767b\u3002\u4f46\u662f\u5982\u679c\u6211\u5c06\u9009\u4e3e\u5e74\u4efd\u6539\u4e3a 2024 \u5e74\uff0c\u5b83\u5c31\u4f1a\u8ba4\u4e3a\u603b\u7edf\u662f\u8fbe\u65af\u00b7\u7ef4\u8fbe\u662f 2026 \u5e74\u7684\u603b\u7edf\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
                                                    • \n
                                                      \u5f62\u5f0f\u63a8\u7406<\/strong>\uff1aGPT-3.5\u7cfb\u5217\u4e0d\u80fd\u5728\u6570\u5b66\u6216\u4e00\u9636\u903b\u8f91\u7b49\u5f62\u5f0f\u4e25\u683c\u7684\u7cfb\u7edf\u4e2d\u8fdb\u884c\u63a8\u7406\uff1a<\/span><\/section>\n<\/li>\n
                                                        \n
                                                      • \n
                                                        \u4e00\u4e2a\u4f8b\u5b50\u662f\u4e25\u683c\u7684\u6570\u5b66\u8bc1\u660e\uff0c\u8981\u6c42\u4e2d\u95f4\u6b65\u9aa4\u4e2d\u4e0d\u80fd\u8df3\uff0c\u4e0d\u80fd\u6a21\u7cca\uff0c\u4e0d\u80fd\u9519\u3002<\/span><\/section>\n<\/li>\n
                                                      • \n
                                                        \u4f46\u8fd9\u79cd\u4e25\u683c\u63a8\u7406\u5230\u5e95\u662f\u5e94\u8be5\u8ba9\u8bed\u8a00\u6a21\u578b\u505a\u8fd8\u662f\u8ba9\u7b26\u53f7\u7cfb\u7edf\u505a\u8fd8\u6709\u5f85\u8ba8\u8bba\u3002\u4e00\u4e2a\u4f8b\u5b50\u662f\uff0c\u4e0e\u5176\u52aa\u529b\u8ba9 GPT \u505a\u4e09\u4f4d\u6570\u52a0\u6cd5\uff0c\u4e0d\u5982\u76f4\u63a5\u8c03 Python\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n
                                                          \n
                                                        • \n
                                                          \u751f\u6210\u5982\u4f55\u505a\u8c46\u8150\u8111\u7684\u65b9\u6cd5\u3002\u505a\u8c46\u8150\u8111\u7684\u65f6\u5019\uff0c\u4e2d\u95f4\u5f88\u591a\u6b65\u9aa4\u6a21\u7cca\u4e00\u70b9\u662f\u53ef\u4ee5\u63a5\u53d7\u7684\uff0c\u6bd4\u5982\u5230\u5e95\u662f\u505a\u54b8\u7684\u8fd8\u662f\u505a\u751c\u7684\u3002\u53ea\u8981\u6574\u4f53\u6b65\u9aa4\u5927\u81f4\u6b63\u786e\uff0c\u505a\u51fa\u6765\u7684\u8c46\u8150\u8111\u513f\u5c31\u80fd\u5403\u3002<\/span><\/section>\n<\/li>\n
                                                        • \n
                                                          \u6570\u5b66\u5b9a\u7406\u7684\u8bc1\u660e\u601d\u8def\u3002\u8bc1\u660e\u601d\u8def\u662f\u7528\u8bed\u8a00\u8868\u8fbe\u7684\u975e\u6b63\u5f0f\u7684\u9010\u6b65\u89e3\u6cd5\uff0c\u5176\u4e2d\u6bcf\u4e00\u6b65\u7684\u4e25\u683c\u63a8\u5bfc\u53ef\u4ee5\u4e0d\u7528\u592a\u5177\u4f53\u3002\u8bc1\u660e\u601d\u8def\u7ecf\u5e38\u88ab\u7528\u5230\u6570\u5b66\u6559\u5b66\uff1a\u53ea\u8981\u8001\u5e08\u7ed9\u4e00\u4e2a\u5927\u81f4\u6b63\u786e\u7684\u6574\u4f53\u6b65\u9aa4\uff0c\u5b66\u751f\u5c31\u53ef\u4ee5\u5927\u6982\u660e\u767d\u3002\u7136\u540e\u8001\u5e08\u628a\u5177\u4f53\u7684\u8bc1\u660e\u7ec6\u8282\u4f5c\u4e3a\u4f5c\u4e1a\u5e03\u7f6e\u7ed9\u5b66\u751f\uff0c\u7b54\u6848\u7565\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n