{"id":41040,"date":"2023-03-04T22:17:05","date_gmt":"2023-03-04T14:17:05","guid":{"rendered":"https:\/\/swarma.org\/?p=41040"},"modified":"2023-03-08T09:54:54","modified_gmt":"2023-03-08T01:54:54","slug":"%e7%90%86%e8%a7%a3%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b-10%e7%af%87%e8%ae%ba%e6%96%87%e7%9a%84%e7%ae%80%e6%98%8e%e6%b8%85%e5%8d%95","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=41040","title":{"rendered":"\u7406\u89e3\u5927\u8bed\u8a00\u6a21\u578b\u2014\u201410\u7bc7\u8bba\u6587\u7684\u7b80\u660e\u6e05\u5355"},"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

\u7406\u89e3\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4e00\u4efd\u6700\u77ed\u9605\u8bfb\u6e05\u5355\uff0c\u5e94\u8be5\u5305\u542b\u54ea\u4e9b\u8bba\u6587\uff1f\u8fd1\u65e5\u4e00\u4f4d\u535a\u4e3b\u68b3\u7406\u4e8610\u7bc7\u91cc\u7a0b\u7891\u5f0f\u7684\u8bba\u6587\u3002<\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/p>\n

\u00a0<\/span><\/strong><\/span><\/strong><\/span><\/strong><\/p>\n<\/section>\n

\n

\u5173\u952e\u8bcd\uff1a\u8bed\u8a00\u6a21\u578b\uff0c\u673a\u5668\u5b66\u4e60<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

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

Sebastian Raschka<\/span>\u00a0<\/span>| \u4f5c\u8005<\/strong><\/span><\/span><\/p>\n

\u8303\u601d\u96e8<\/span>\u00a0<\/span>| \u8bd1\u8005<\/span><\/strong><\/span><\/p>\n

\u9093\u4e00\u96ea<\/span>\u00a0<\/span>| <\/strong><\/span>\u7f16\u8f91<\/strong><\/p>\n

\n

\n<\/section>\n

\u5927\u8bed\u8a00\u6a21\u578b<\/span>\uff08Large language models\uff0cLLMs\uff09<\/span>\u5982\u4eca\u5df2\u6210\u4e3a\u516c\u4f17\u5173\u6ce8\u7684\u7126\u70b9\u3002\u5728\u77ed\u77ed\u4e94\u5e74\u65f6\u95f4\u5185\uff0c\u57fa\u4e8e Transformer\u7ed3\u6784\u7684\u5927\u8bed\u8a00\u6a21\u578b\u51e0\u4e4e\u5f7b\u5e95\u6539\u53d8\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\u7684\u7814\u7a76\u3002\u6b64\u5916\uff0c\u8fd9\u4e9b\u6a21\u578b\u4e5f\u5f00\u59cb\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u548c\u8ba1\u7b97\u751f\u7269\u5b66\u7b49\u9886\u57df\u4ea7\u751f\u9769\u547d\u6027\u5f71\u54cd\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\u8003\u8651\u5230\u5927\u8bed\u8a00\u6a21\u578b\u7684\u51fa\u73b0\u5bf9\u6bcf\u4e2a\u4eba\u7684\u7814\u7a76\u4ea7\u751f\u4e86\u5982\u6b64\u5927\u7684\u5f71\u54cd\uff0c\u6211\u60f3\u4e3a\u6b63\u5728\u5165\u95e8\u673a\u5668\u5b66\u4e60\u7684\u5b66\u754c\u548c\u4e1a\u754c\u670b\u53cb\u5217\u51fa\u4e00\u4efd\u9605\u8bfb\u6e05\u5355\u3002\u5728\u4e0b\u6587\u4e2d\uff0c\u6211\u5c06\u4f1a\u7ed9\u51fa\u4e00\u4efd\u4ec5\u5305\u542b\u5b66\u672f\u8bba\u6587\u7684\u9605\u8bfb\u6e05\u5355\uff0c\u5e76\u6309\u7167\u8bba\u6587\u53d1\u8868\u7684\u65f6\u95f4\u987a\u5e8f\u4f9d\u6b21\u4ecb\u7ecd\u3002\u5f53\u7136\uff0c\u8fd8\u6709\u5f88\u591a\u5176\u4ed6\u6709\u7528\u7684\u8d44\u6e90\u4f9b\u5404\u4f4d\u8bfb\u8005\u53c2\u8003\uff0c\u4f8b\u5982\uff1a<\/span><\/section>\n
\u00a0\u2022\u00a0Jay Alammar \u89e3\u91ca Transformer \u6a21\u578b\u7684\u6587\u7ae0<\/span><\/section>\n

http:\/\/jalammar.github.io\/illustrated-transformer\/<\/span><\/p>\n

\u00a0\u2022\u00a0Lilian Weng \u66f4\u504f\u6280\u672f\u6027\u7684\u535a\u5ba2\u6587\u7ae0<\/span><\/section>\n

https:\/\/lilianweng.github.io\/posts\/2020-04-07-the-transformer-family\/<\/span><\/p>\n

\u00a0\u2022\u00a0Xavier Amatriain \u5173\u4e8e Transformer \u7cfb\u5217\u6a21\u578b\u7684\u79d1\u666e\u6587\u7ae0<\/span><\/section>\n

https:\/\/amatriain.net\/blog\/transformer-models-an-introduction-and-catalog-2d1e9039f376\/<\/span><\/p>\n

\u00a0\u2022\u00a0Andrej Karpathy\uff08\u7279\u65af\u62c9\u524dAI\u603b\u76d1\uff09\u8bad\u7ec3\u4e2d\u578bGPT\u7684\u5f00\u6e90\u4ee3\u7801\u5e93\uff08nanoGPT\uff09<\/span><\/section>\n

https:\/\/github.com\/karpathy\/nanoGPT<\/span><\/p>\n

\u00a0\u2022\u00a0\u4ee5\u53ca\u7531\u6211\u81ea\u5df1\u6388\u8bfe\u7684\u7cfb\u5217\u8bb2\u5ea7\u89c6\u9891\u548c\u4e66\u7c4d\u7ae0\u8282<\/span><\/section>\n
https:\/\/sebastianraschka.com\/blog\/2021\/dl-course.html#l19-self-attention-and-transformer-networks<\/span><\/section>\n
https:\/\/github.com\/rasbt\/machine-learning-book\/tree\/main\/ch16<\/span><\/section>\n

<\/h2>\n
<\/section>\n
\u00a0<\/span><\/section>\n
\n
\n
\n
\n
<\/section>\n
<\/section>\n
\n
<\/section>\n<\/section>\n<\/section>\n
\n

\u7406\u89e3\u5927\u8bed\u8a00\u6a21\u578b\u7684\u7ed3\u6784\u548c\u4efb\u52a1<\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
<\/section>\n
\u5982\u679c\u4f60\u662f\u521a\u521a\u63a5\u89e6 Transformer \u7cfb\u5217\u6a21\u578b\u6216\u8005\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u90a3\u4e48\u6211\u4eec\u6700\u597d\u4ece\u5934\u5f00\u59cb\u4ecb\u7ecd\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\uff081\uff09\u7ed3\u5408\u5bf9\u9f50\u548c\u7ffb\u8bd1\u7684\u795e\u7ecf\u7f51\u7edc\u673a\u5668\u7ffb\u8bd1\u6a21\u578b<\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aNeural Machine Translation by Jointly Learning to Align and Translate (2014)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aBahdanau, Cho, Bengio<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1409.0473<\/span><\/section>\n<\/blockquote>\n
\u5982\u679c\u65f6\u95f4\u5145\u8db3\uff0c\u6211\u5efa\u8bae\u4ece\u8fd9\u7bc7\u8bba\u6587\u5f00\u59cb\u8bfb\u8d77\u3002\u8fd9\u7bc7\u6587\u7ae0\u5f15\u5165\u4e86\u4e00\u79cd\u6ce8\u610f\u529b\u673a\u5236<\/span>\uff08attention mechanism\uff09<\/span>\uff0c\u7528\u4e8e\u63d0\u5347\u9012\u5f52\u795e\u7ecf\u7f51\u7edc<\/span>\uff08RNN\uff09<\/span>\u7684\u957f\u5e8f\u5217\u5efa\u6a21\u80fd\u529b\u3002\u8fd9\u4f7f\u5f97 RNN \u80fd\u591f\u66f4\u51c6\u786e\u5730\u7ffb\u8bd1\u66f4\u957f\u7684\u53e5\u5b50\u2014\u2014\u8fd9\u4e5f\u662f\u540e\u6765\u5f00\u53d1\u51fa\u539f\u59cb Transformer \u6a21\u578b\u7684\u52a8\u673a\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n
\uff082\uff09\u6ce8\u610f\u529b\u673a\u5236<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aAttention Is All You Need (2017)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aVaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1706.03762<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\u8fd9\u7bc7\u8bba\u6587\u4ecb\u7ecd\u4e86\u539f\u59cb Transformer \u6a21\u578b\u7684\u7ed3\u6784\u3002\u8be5\u6a21\u578b\u7531\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u4e24\u90e8\u5206\u7ec4\u6210\uff0c\u8fd9\u4e24\u4e2a\u90e8\u5206\u5728\u540e\u7eed\u6a21\u578b\u4e2d\u5206\u79bb\u6210\u4e24\u4e2a\u72ec\u7acb\u7684\u6a21\u5757\u3002\u6b64\u5916\uff0c\u8be5\u8bba\u6587\u8fd8\u5f15\u5165\u4e86\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u673a\u5236<\/span>\uff08Scaled Dot Product Attention Mechanism\uff09<\/span>\u3001\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236<\/span>\uff08Multi-head Attention Blocks\uff09<\/span>\u548c\u4f4d\u7f6e\u7f16\u7801<\/span>\uff08Positional Input Encoding\uff09<\/span>\u7b49\u6982\u5ff5\uff0c\u8fd9\u4e9b\u6982\u5ff5\u4ecd\u7136\u662f\u73b0\u4ee3 Transformer \u7cfb\u5217\u6a21\u578b\u7684\u57fa\u7840\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n
\uff083\uff09BERT: \u8bed\u8a00\u7406\u89e3\u7684\u6df1\u5ea6\u53cc\u5411 Transformer \u9884\u8bad\u7ec3<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aDevlin, Chang, Lee, and Toutanova<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1810.04805<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\u5728\u539f\u59cb\u7684 Transformer \u6a21\u578b\u4e4b\u540e\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7814\u7a76\u5f00\u59cb\u5411\u4e24\u4e2a\u65b9\u5411\u5206\u5316\uff1a\u57fa\u4e8e\u7f16\u7801\u5668\u7ed3\u6784\u7684 Transformer \u6a21\u578b\u7528\u4e8e\u9884\u6d4b\u5efa\u6a21\u4efb\u52a1\uff0c\u4f8b\u5982\u6587\u672c\u5206\u7c7b\uff1b\u800c\u57fa\u4e8e\u89e3\u7801\u5668\u7ed3\u6784\u7684 Transformer \u6a21\u578b\u7528\u4e8e\u751f\u6210\u5efa\u6a21\u4efb\u52a1\uff0c\u4f8b\u5982\u7ffb\u8bd1\u3001\u6458\u8981\u548c\u5176\u4ed6\u5f62\u5f0f\u7684\u6587\u672c\u5185\u5bb9\u751f\u6210\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n
\u4e0a\u8ff0 BERT \u8bba\u6587\u5f15\u5165\u4e86\u63a9\u7801\u8bed\u8a00\u5efa\u6a21<\/span>\uff08Masked-language Modeling\uff09<\/span>\u6982\u5ff5\uff0c\u9664\u6b64\u4ee5\u5916\uff0c\u4e0b\u53e5\u9884\u6d4b<\/span>\uff08Next-Sentence Prediction\uff09<\/span>\u5219\u662f\u4e00\u79cd\u5f88\u91cd\u8981\u7684\u57fa\u4e8e\u89e3\u7801\u5668\u7ed3\u6784\u7684\u9884\u8bad\u7ec3\u4efb\u52a1\u3002\u5982\u679c\u4f60\u5bf9\u8fd9\u4e00\u7814\u7a76\u5206\u652f\u611f\u5174\u8da3\uff0c\u6211\u5efa\u8bae\u4f60\u7ee7\u7eed\u7814\u7a76 RoBERTa \u6a21\u578b\uff0c\u5b83\u901a\u8fc7\u53d6\u6d88\u4e0b\u53e5\u9884\u6d4b\u4efb\u52a1\u7b80\u5316\u4e86\u9884\u8bad\u7ec3\u76ee\u6807\u3002<\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aRoBERTa: A Robustly Optimized BERT Pretraining Approach<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1907.11692<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\uff084\uff09\u901a\u8fc7\u751f\u6210\u9884\u8bad\u7ec3\u6539\u8fdb\u8bed\u8a00\u7406\u89e3<\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aImproving Language Understanding by Generative Pre-Training (2018)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aRadford and Narasimhan<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/www.semanticscholar.org\/paper\/Improving-Language-Understanding-by-Generative-Radford-Narasimhan\/cd18800a0fe0b668a1cc19f2ec95b5003d0a5035<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\u8fd9\u7bc7\u5173\u4e8e\u751f\u6210\u5f0f\u9884\u8bad\u7ec3\u6a21\u578b<\/span>\uff08Generative Pre-Training\uff0cGPT\uff09<\/span>\u7684\u6587\u7ae0\u4ecb\u7ecd\u4e86\u57fa\u4e8e\u89e3\u7801\u5668\u7684\u6a21\u578b\u7ed3\u6784\u548c\u4e0b\u53e5\u9884\u6d4b\u7684\u9884\u8bad\u7ec3\u4efb\u52a1\u3002BERT \u53ef\u4ee5\u88ab\u8ba4\u4e3a\u662f\u53cc\u5411 Transformer \u6a21\u578b\uff0c\u56e0\u4e3a\u5176\u91c7\u7528\u4e86\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\uff0cGPT \u5219\u662f\u5355\u5411\u81ea\u56de\u5f52\u6a21\u578b\u3002GPT \u7684\u5d4c\u5165\u4e5f\u53ef\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1\uff0cGPT \u53ef\u88ab\u79f0\u4e3a\u73b0\u5728\u6700\u5177\u5f71\u54cd\u529b\u7684\u5404\u7c7b\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6838\u5fc3\uff0c\u4f8b\u5982 ChatGPT\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n
\u5982\u679c\u4f60\u5bf9\u8fd9\u4e2a\u7814\u7a76\u65b9\u5411\u611f\u5174\u8da3\uff0c\u6211\u5efa\u8bae\u4f60\u7ee7\u7eed\u9605\u8bfb GPT-2 \u548c GPT-3 \u7684\u76f8\u5173\u8bba\u6587\u3002\u8fd9\u4e24\u7bc7\u8bba\u6587\u8868\u660e\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u591f\u8fdb\u884c\u96f6\u6837\u672c\u5b66\u4e60\u548c\u5c11\u91cf\u6837\u672c\u5b66\u4e60\uff0c\u5e76\u5f3a\u8c03\u4e86\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u6d8c\u73b0\u80fd\u529b<\/span>\uff08Emergent Ability\uff09<\/span>\u3002GPT-3 \u4e5f\u662f\u8bad\u7ec3\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\u7684\u57fa\u51c6\u6a21\u578b\uff0c\u4f8b\u5982 ChatGPT\u3002\u6211\u4eec\u5c06\u5728\u540e\u7eed\u5355\u72ec\u4ecb\u7ecd\u4e4b\u540e\u5b55\u80b2\u51fa ChatGPT \u7684\u6a21\u578b\u2014\u2014 InstructGPT\u3002<\/span><\/section>\n
\n
GPT-2 \u8bba\u6587<\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aLanguage Models are Unsupervised Multitask Learners<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/www.semanticscholar.org\/paper\/Language-Models-are-Unsupervised-Multitask-Learners-Radford-Wu\/9405cc0d6169988371b2755e573cc28650d14dfe<\/span><\/section>\n
<\/section>\n<\/section>\n<\/blockquote>\n
\n
GPT-3 \u8bba\u6587<\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aLanguage Models are Few-Shot Learners<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2005.14165<\/span><\/section>\n<\/section>\n
<\/section>\n<\/blockquote>\n
<\/section>\n
\uff085\uff09BART\uff1a\u7528\u4e8e\u81ea\u7136\u8bed\u8a00\u751f\u6210\u3001\u7ffb\u8bd1\u548c\u7406\u89e3\u7684\u964d\u566a\u5e8f\u5217\u5bf9\u5e8f\u5217\u9884\u8bad\u7ec3<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aBART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension (2019)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aLewis, Liu, Goyal, Ghazvininejad, Mohamed, Levy, Stoyanov, and Zettlemoyer<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1910.13461<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\u5982\u524d\u6240\u8ff0\uff0c\u7c7b\u4f3cBERT\u8fd9\u6837\u57fa\u4e8e\u7f16\u7801\u5668\u7ed3\u6784\u7684\u5927\u8bed\u8a00\u6a21\u578b\u901a\u5e38\u7528\u4e8e\u9884\u6d4b\u5efa\u6a21\u4efb\u52a1\u3002\u800c\u7c7b\u4f3c GPT \u8fd9\u6837\u57fa\u4e8e\u89e3\u7801\u5668\u7ed3\u6784\u7684\u5927\u8bed\u8a00\u6a21\u578b\u5219\u66f4\u9002\u5408\u6587\u672c\u751f\u6210\u4efb\u52a1\u3002\u4e3a\u4e86\u517c\u987e\u4e24\u7c7b\u6a21\u578b\u7684\u4f18\u70b9\uff0c\u8fd9\u7bc7 BART \u8bba\u6587\u7ed3\u5408\u4e86\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u4e24\u4e2a\u90e8\u5206<\/span>\uff08\u7c7b\u4f3c\u4e8e\u672c\u6e05\u5355\u4e2d\u63d0\u5230\u7684\u7b2c\u4e8c\u7bc7\u8bba\u6587\uff09<\/span>\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n

\n

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

\u89c4\u6a21\u6cd5\u5219\u548c\u6a21\u578b\u6548\u7387\u63d0\u5347<\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
<\/section>\n
\u5982\u679c\u4f60\u60f3\u4e86\u89e3\u66f4\u591a\u6709\u5173\u63d0\u5347 Transformer \u6a21\u578b\u6548\u7387\u7684\u65b9\u6cd5\uff0c\u6211\u63a8\u8350\u9605\u8bfb2020\u5e74\u7684 Efficient Transformers: A Survey<\/em> \u4e00\u6587\uff0c\u7136\u540e\u518d\u9605\u8bfb2023\u5e74\u7684 A Survey on Efficient Training of Transformers<\/em>\u3002\u6b64\u5916\uff0c\u4ee5\u4e0b\u662f\u6211\u8ba4\u4e3a\u6709\u8da3\u4e14\u503c\u5f97\u6df1\u5165\u7814\u7a76\u7684\u4e24\u7bc7\u6587\u7ae0\u3002<\/span><\/section>\n
<\/section>\n
\uff086\uff09Flash Attention\uff1a\u5177\u6709 IO \u611f\u77e5\u529f\u80fd\u7684\u5feb\u901f\u9ad8\u6548\u7684\u7cbe\u786e\u6ce8\u610f\u529b\u673a\u5236<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aFlash Attention: Fast and Memory-Efficient Exact Attention with IO-Awareness (2022)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aDao, Fu, Ermon, Rudra, and R\u00e9<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2205.14135<\/span><\/section>\n<\/blockquote>\n
\u5c3d\u7ba1\u5927\u591a\u6570 Transformer \u8bba\u6587\u5e76\u4e0d\u8003\u8651\u7528\u5176\u4ed6\u673a\u5236\u66ff\u6362\u7f29\u653e\u70b9\u79ef\u673a\u5236\u6765\u5b9e\u73b0\u81ea\u6ce8\u610f\u529b<\/span>\uff08Self-Attention\uff09<\/span>\uff0c\u4f46 Flash Attention \u662f\u6211\u6700\u8fd1\u770b\u5230\u7684\u6700\u5e38\u88ab\u5f15\u7528\u7684\u65b0\u673a\u5236\u4e4b\u4e00\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n
\uff087\uff09Cramming\uff1a\u5229\u7528\u5355\u4e2a GPU \u5355\u65e5\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aCramming: Training a Language Model on a Single GPU in One Day (2022)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aGeiping and Goldstein<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2212.14034<\/span><\/section>\n<\/blockquote>\n
\u5728\u8fd9\u7bc7\u8bba\u6587\u4e2d\uff0c\u7814\u7a76\u8005\u5728\u5355\u4e2aGPU\u4e0a\u5bf9\u63a9\u7801\u8bed\u8a00\u6a21\u578b\/\u57fa\u4e8e\u7f16\u7801\u5668\u7684\u5927\u8bed\u8a00\u6a21\u578b<\/span>\uff08\u8fd9\u91cc\u6307BERT\uff09<\/span>\u8fdb\u884c\u4e8624\u5c0f\u65f6\u7684\u8bad\u7ec3\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c2018\u5e74\u7684BERT\u8bba\u6587\u4f7f\u7528\u4e8616\u4e2aTPU\u8fdb\u884c\u4e86\u56db\u5929\u7684\u8bad\u7ec3\u3002\u6709\u8da3\u7684\u662f\uff0c\u867d\u7136\u8f83\u5c0f\u7684\u6a21\u578b\u5177\u6709\u66f4\u9ad8\u7684\u541e\u5410\u91cf<\/span>\uff08throughput\uff09<\/span>\uff0c\u4f46\u5b83\u4eec\u5b66\u4e60\u7684\u6548\u7387\u4e5f\u66f4\u4f4e\u3002\u56e0\u6b64\uff0c\u66f4\u5927\u7684\u6a21\u578b\u4e0d\u4e00\u5b9a\u9700\u8981\u82b1\u8d39\u66f4\u591a\u7684\u8bad\u7ec3\u65f6\u95f4\u5c31\u80fd\u8fbe\u5230\u7279\u5b9a\u7684\u9884\u6d4b\u6027\u80fd\u9608\u503c\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n
\uff088\uff09\u8bad\u7ec3\u8ba1\u7b97\u6700\u4f18\u7684\u5927\u8bed\u8a00\u6a21\u578b<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aTraining Compute-Optimal Large Language Models (2022)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aHoffmann, Borgeaud, Mensch, Buchatskaya, Cai, Rutherford, de Las Casas, Hendricks, Welbl, Clark, Hennigan, Noland, Millican, van den Driessche, Damoc, Guy, Osindero, Simonyan, Elsen, Rae, Vinyals, and Sifre<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2203.15556<\/span><\/section>\n<\/blockquote>\n
\u8fd9\u7bc7\u8bba\u6587\u4ecb\u7ecd\u4e86 Chinchilla \u6a21\u578b\uff0c\u5b83\u5305\u542b 700 \u4ebf\u4e2a\u53c2\u6570\u3002\u5728\u751f\u6210\u5efa\u6a21\u4efb\u52a1\u4e2d\uff0c\u8be5\u6a21\u578b\u8868\u73b0\u4f18\u4e8e\u5305\u542b 1750 \u4ebf\u4e2a\u53c2\u6570\u7684 GPT-3 \u6a21\u578b\u3002\u7136\u800c\uff0c\u8fd9\u7bc7\u8bba\u6587\u7684\u4e3b\u8981\u8bba\u70b9\u662f\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\u5b58\u5728\u201c\u8bad\u7ec3\u4e0d\u8db3\u201d\u7684\u95ee\u9898\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\u8be5\u8bba\u6587\u5b9a\u4e49\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u7684\u7ebf\u6027\u7f29\u653e\u5f8b<\/span>\uff08Linear Scaling Law\uff09<\/span>\u3002\u4f8b\u5982\uff0c\u867d\u7136 Chinchilla \u7684\u53c2\u6570\u91cf\u53ea\u6709 GPT-3 \u7684\u4e00\u534a\uff0c\u4f46\u5b83\u7684\u8868\u73b0\u5374\u4f18\u4e8e GPT-3\u3002\u56e0\u4e3a\u5b83\u662f\u5728 1.4 \u4e07\u4ebf\u4e2a\u8bcd\u8bed\u6807\u8bb0<\/span>\uff08\u800c\u4e0d\u4ec5\u4ec5\u662f 3000 \u4ebf\u4e2a\uff09<\/span>\u4e0a\u8fdb\u884c\u8bad\u7ec3\u7684\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u8bcd\u8bed\u6807\u8bb0\u7684\u6570\u91cf\u4e0e\u6a21\u578b\u5927\u5c0f\u4e00\u6837\u91cd\u8981\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
\u00a0<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\n
\n
\n
\n
<\/section>\n
<\/section>\n
\n
<\/section>\n<\/section>\n<\/section>\n
\n

\u5bf9\u9f50\u2014\u2014\u5f15\u5bfc\u5927\u8bed\u8a00\u6a21\u578b\u5b8c\u6210\u8bad\u7ec3\u76ee\u6807<\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
<\/section>\n
\u8fd1\u5e74\u6765\uff0c\u6211\u4eec\u770b\u5230\u8bb8\u591a\u8868\u73b0\u51fa\u8272\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u5b83\u4eec\u53ef\u4ee5\u751f\u6210\u975e\u5e38\u8d34\u8fd1\u771f\u5b9e\u7684\u6587\u672c\u5185\u5bb9<\/span>\uff08\u4f8b\u5982GPT-3\u548cChinchilla\uff09<\/span>\u3002\u5c31\u5e38\u7528\u7684\u9884\u8bad\u7ec3\u8303\u5f0f\u800c\u8a00\uff0c\u6211\u4eec\u4f3c\u4e4e\u5df2\u7ecf\u8fbe\u5230\u4e86\u4e00\u4e2a\u4e0a\u9650\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\u4e3a\u4e86\u4f7f\u8bed\u8a00\u6a21\u578b\u53d1\u6325\u66f4\u5927\u4f5c\u7528\uff0c\u51cf\u5c11\u9519\u8bef\u4fe1\u606f\u548c\u6709\u5bb3\u8bed\u8a00\u7684\u751f\u6210\uff0c\u7814\u7a76\u8005\u8bbe\u8ba1\u4e86\u65b0\u7684\u8bad\u7ec3\u8303\u5f0f\u6765\u5bf9\u9884\u8bad\u7ec3\u7684\u57fa\u7840\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\uff089\uff09\u4f7f\u7528\u4eba\u7c7b\u53cd\u9988\u8fdb\u884c\u5fae\u8c03\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5728\u5404\u79cd\u4efb\u52a1\u4e0a\u4e0e\u7528\u6237\u610f\u56fe\u4fdd\u6301\u4e00\u81f4<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aTraining Language Models to Follow Instructions with Human Feedback (2022)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aOuyang, Wu, Jiang, Almeida, Wainwright, Mishkin, Zhang, Agarwal, Slama, Ray, Schulman, Hilton, Kelton, Miller, Simens, Askell, Welinder, Christiano, Leike, and Lowe<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2203.02155<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\u5728\u8fd9\u7bc7\u88ab\u79f0\u4e3a InstructGPT \u7684\u8bba\u6587\u4e2d\uff0c\u7814\u7a76\u8005\u4f7f\u7528\u4e86\u4e00\u79cd\u5229\u7528\u4eba\u7c7b\u53cd\u9988\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u7684\u673a\u5236<\/span>\uff08Reinforcement Learning with Human Feedback \uff0cRLHF\uff09<\/span>\u3002\u4ed6\u4eec\u57fa\u4e8e\u9884\u8bad\u7ec3\u7684 GPT-3 \u6a21\u578b\uff0c\u4f7f\u7528\u76d1\u7763\u5b66\u4e60\u5bf9\u4eba\u7c7b\u751f\u6210\u7684\u63d0\u793a\u54cd\u5e94\u5bf9<\/span>\uff08 Prompt-Response Pairs\uff09<\/span>\u5bf9\u6a21\u578b\u8fdb\u884c\u8fdb\u4e00\u6b65\u5fae\u8c03<\/span>\uff08\u6b65\u9aa41\uff09<\/span>\u3002\u63a5\u4e0b\u6765\uff0c\u4ed6\u4eec\u8981\u6c42\u4eba\u7c7b\u5bf9\u6a21\u578b\u7684\u8f93\u51fa\u8fdb\u884c\u6392\u5e8f\u4ee5\u8bad\u7ec3\u5956\u52b1\u6a21\u578b<\/span>\uff08\u6b65\u9aa42\uff09<\/span>\u3002\u6700\u540e\uff0c\u4ed6\u4eec\u4f7f\u7528\u5956\u52b1\u6a21\u578b\uff0c\u901a\u8fc7\u8fd1\u7aef\u7b56\u7565\u4f18\u5316<\/span>\uff08Proximal Policy Optimization\uff09\uff08\u6b65\u9aa43\uff09<\/span>\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff0c\u6765\u8c03\u6574\u7ecf\u8fc7\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u7684GPT-3\u6a21\u578b\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\u987a\u4fbf\u8bf4\u4e00\u53e5\uff0c\u8fd9\u7bc7\u8bba\u6587\u4e5f\u88ab\u79f0\u4e3a\u9610\u8ff0 ChatGPT \u601d\u60f3\u5185\u6838\u7684\u8bba\u6587\u2014\u2014\u6839\u636e\u6700\u8fd1\u7684\u8bf4\u6cd5\uff0cChatGPT\u662fInstructGPT\u7684\u6269\u5c55\u7248\u672c\uff0c\u5b83\u5b9e\u73b0\u4e86\u5728\u66f4\u5927\u7684\u6570\u636e\u96c6\u4e0a\u7684\u5fae\u8c03\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
<\/section>\n
\uff0810\uff09Constitutional AI\uff1aAI \u53cd\u9988\u8fbe\u6210\u65e0\u5bb3\u6027<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aConstitutional AI: Harmlessness from AI Feedback (2022)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aYuntao, Saurav, Sandipan, Amanda, Jackson, Jones, Chen, Anna, Mirhoseini, McKinnon, Chen, Olsson, Olah, Hernandez, Drain, Ganguli, Li, Tran-Johnson, Perez, Kerr, Mueller, Ladish, Landau, Ndousse, Lukosuite, Lovitt, Sellitto, Elhage, Schiefer, Mercado, DasSarma, Lasenby, Larson, Ringer, Johnston, Kravec, El Showk, Fort, Lanham, Telleen-Lawton, Conerly, Henighan, Hume, Bowman, Hatfield-Dodds, Mann, Amodei, Joseph, McCandlish, Brown, Kaplan<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2212.08073<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\u5728\u8fd9\u7bc7\u8bba\u6587\u4e2d\uff0c\u7814\u7a76\u8005\u5c06\u5bf9\u9f50\u601d\u60f3<\/span>\uff08Alignment\uff09<\/span>\u53c8\u5f80\u524d\u63a8\u4e86\u4e00\u6b65\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u8bad\u7ec3\u673a\u5236\uff0c\u7528\u4e8e\u521b\u5efa\u4e00\u4e2a\u201c\u65e0\u5bb3\u201d\u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u3002\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u81ea\u6211\u8bad\u7ec3\u673a\u5236\uff0c\u8fd9\u79cd\u673a\u5236\u57fa\u4e8e\u4e00\u7cfb\u5217\u89c4\u5219<\/span>\uff08\u7531\u4eba\u7c7b\u63d0\u4f9b\uff09<\/span>\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u4eba\u7c7b\u76d1\u7763\u3002\u8fd9\u7bc7\u6587\u7ae0\u4e0e\u524d\u6587\u63d0\u5230\u7684 InstructGPT \u4e00\u6837\u91c7\u7528\u4e86\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n

<\/h2>\n
\u00a0<\/span><\/section>\n
\n
\n
\n
\n
<\/section>\n
<\/section>\n
\n
<\/section>\n<\/section>\n<\/section>\n
\n

\u9644\u52a0\u9605\u8bfb\u6e05\u5355\uff1a<\/strong><\/strong><\/p>\n

\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09<\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
<\/section>\n
\u5c3d\u7ba1 RLHF \u53ef\u80fd\u65e0\u6cd5\u89e3\u51b3\u5927\u8bed\u8a00\u6a21\u578b\u5f53\u524d\u9762\u4e34\u7684\u6240\u6709\u95ee\u9898\uff0c\u4f46\u5b83\u4ecd\u7136\u88ab\u8ba4\u4e3a\u662f\u6700\u4f73\u7684\u9009\u62e9\uff0c\u7279\u522b\u662f\u4e0e\u4e0a\u4e00\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\u76f8\u6bd4\u3002\u672a\u6765\uff0c\u6211\u4eec\u4f1a\u6709\u5f88\u5927\u5e0c\u671b\u770b\u5230\u66f4\u591a\u521b\u9020\u6027\u7684\u65b9\u6cd5\u5c06 RLHF \u5e94\u7528\u5230\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5176\u4ed6\u9886\u57df\u3002<\/span><\/section>\n
<\/section>\n
\u4e0a\u9762\u7684\u4e24\u7bc7\u8bba\u6587\uff0cInstructGPT \u548c Constitutional AI\uff0c\u90fd\u4f7f\u7528\u4e86 RLHF\u3002\u6bcb\u5eb8\u7f6e\u7591\uff0cRLHF \u4ee3\u8868\u4e86\u672a\u6765\u7684\u7814\u7a76\u8d8b\u52bf\u3002\u5982\u679c\u4f60\u60f3\u4e86\u89e3 RLHF\uff0c\u672c\u8282\u63d0\u4f9b\u4e86\u989d\u5916\u7684\u53c2\u8003\u8d44\u6599\u3002<\/span>(\u4ece\u6280\u672f\u5c42\u9762\u8bb2\uff0cConstitutional AI \u8fd9\u7bc7\u8bba\u6587\u4f7f\u7528\u4e86 AI \u800c\u4e0d\u662f\u4eba\u7c7b\u53cd\u9988\uff0c\u4f46\u5b83\u4f7f\u7528\u4e86\u7c7b\u4f3c RL \u7684\u6982\u5ff5\u3002)<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\uff0811\uff09\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7684\u5f02\u6b65\u65b9\u6cd5<\/span><\/strong><\/span><\/section>\n
\u00a0<\/span><\/section>\n
\u8fd9\u7bc7\u6587\u7ae0\u5f15\u5165\u4e86\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\uff0c\u4ee5\u66ff\u4ee3\u4f5c\u4e3a\u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u7684\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684 Q-learning \u65b9\u6cd5\u3002<\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aAsynchronous Methods for Deep Reinforcement Learning (2016)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aMnih, Badia, Mirza, Graves, Lillicrap, Harley, Silver, and Kavukcuoglu<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1602.01783<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\uff0812\uff09\u8fd1\u7aef\u7b56\u7565\u4f18\u5316<\/span><\/strong><\/span><\/section>\n
\u00a0<\/span><\/section>\n
\u8fd9\u7bc7\u6587\u7ae0\u63d0\u51fa\u4e86\u4e00\u79cd\u6539\u8fdb\u7684\u8fd1\u7aef\u7b56\u7565\u5f3a\u5316\u5b66\u4e60\u8fc7\u7a0b\uff0c\u8be5\u8fc7\u7a0b\u6bd4\u4e0a\u9762\u7684\u666e\u901a\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\u66f4\u5177\u6570\u636e\u9ad8\u6548\u6027\u548c\u53ef\u6269\u5c55\u6027\u3002<\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aProximal Policy Optimization Algorithms (2017)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aSchulman, Wolski, Dhariwal, Radford, Klimov<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1707.06347<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\uff0813\uff09\u501f\u52a9\u4eba\u7c7b\u53cd\u9988\u5b66\u4e60\u603b\u7ed3<\/span><\/strong><\/span><\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aLearning to Summarize from Human Feedback (2022)<\/span><\/section>\n
\u8bba\u6587\u4f5c\u8005\uff1aStiennon, Ouyang, Wu, Ziegler, Lowe, Voss, Radford, Amodei, Christiano<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2009.01325<\/span><\/section>\n<\/blockquote>\n
<\/section>\n

\u8fd9\u7bc7\u6587\u7ae0\u4ecb\u7ecd\u4e86\u5b9e\u73b0 RLHF \u7684\u4e09\u6b65\u8fc7\u7a0b\uff1a<\/span><\/p>\n

1. \u9884\u8bad\u7ec3 GPT-3 \u6a21\u578b<\/span><\/p>\n

2. \u5bf9\u6a21\u578b\u8fdb\u884c\u76d1\u7763\u5fae\u8c03<\/span><\/p>\n

3. \u4ee5\u6709\u76d1\u7763\u65b9\u6cd5\u8bad\u7ec3\u5956\u52b1\u6a21\u578b\uff0c\u7136\u540e\u4f7f\u7528\u8be5\u5956\u52b1\u6a21\u578b\u53e0\u52a0\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\u65b9\u6cd5\u4ee5\u8bad\u7ec3\u51fa\u5fae\u8c03\u6a21\u578b\u3002<\/span><\/p>\n

<\/section>\n
\u8fd9\u7bc7\u6587\u7ae0\u8fd8\u8868\u660e\uff0c\u4e0e\u5e38\u89c4\u7684\u76d1\u7763\u5b66\u4e60\u76f8\u6bd4\uff0c\u4f7f\u7528\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\u7684\u5f3a\u5316\u5b66\u4e60\u53ef\u4ee5\u8bad\u7ec3\u51fa\u66f4\u597d\u7684\u6a21\u578b\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
<\/section>\n
\u00a0<\/span><\/section>\n
\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u524d\u6587\u4e2d\u8ba8\u8bba\u7684 InstructGPT \u8bba\u6587\uff089\uff09\u5bf9 RLHF \u4f7f\u7528\u4e86\u7c7b\u4f3c\u7684\u4e09\u6b65\u8fc7\u7a0b\uff0c\u4f46\u5b83\u4e0d\u662f\u603b\u7ed3\u6587\u672c\uff0c\u800c\u662f\u4fa7\u91cd\u4e8e\u6839\u636e\u4eba\u5de5\u6307\u4ee4\u751f\u6210\u6587\u672c\u3002\u6b64\u5916\uff0c\u5b83\u8fd8\u4f7f\u7528\u6253\u6807\u7b7e\u7684\u65b9\u6cd5\u5bf9\u6a21\u578b\u8f93\u51fa\u8fdb\u884c\u4ece\u597d\u5230\u574f\u7684\u6392\u5e8f<\/span>\uff08\u800c\u4e0d\u4ec5\u4ec5\u662f\u4eba\u7c7b\u548c\u4eba\u5de5\u667a\u80fd\u751f\u6210\u7684\u6587\u672c\u4e4b\u95f4\u7684\u6bd4\u8f83\uff09<\/span>\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\n
\n
\n
\n
<\/section>\n
<\/section>\n
\n
<\/section>\n<\/section>\n<\/section>\n
\n

\u7ed3\u8bba\u4ee5\u53ca\u5ef6\u4f38\u9605\u8bfb<\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
<\/section>\n
\u5728\u672c\u6587\u4e2d\uff0c\u6211\u5c1d\u8bd5\u63d0\u4f9b\u4e00\u4efd\u7b80\u6d01\u3001\u6709\u7528\u7684\u9605\u8bfb\u6e05\u5355\uff0c\u4e8e\u662f\u6311\u9009\u4e8610\u7bc7\u6838\u5fc3\u6587\u7ae0<\/span>\uff08\u9644\u52a0\u4e09\u7bc7RLHF\u7684\u6587\u7ae0\uff09<\/span>\u6765\u5e2e\u52a9\u5404\u4f4d\u4e86\u89e3\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bbe\u8ba1\u3001\u7ed3\u6784\u6f14\u53d8\u4ee5\u53ca\u6a21\u578b\u9650\u5236\u3002<\/span><\/section>\n
\u00a0<\/span><\/section>\n
\u60f3\u8981\u4e86\u89e3\u66f4\u591a\u5185\u5bb9\uff0c\u6211\u5efa\u8bae\u7ee7\u7eed\u9605\u8bfb\u4e0a\u8ff0\u8bba\u6587\u7684\u53c2\u8003\u6587\u732e\u3002\u6216\u8005\uff0c\u53c2\u8003\u4ee5\u4e0b\u8d44\u6e90\u4f5c\u4e3a\u5ef6\u4f38\u9605\u8bfb\uff1a<\/span><\/section>\n
<\/section>\n
GPT \u7684\u5f00\u6e90\u66ff\u4ee3\u54c1<\/span><\/section>\n