{"id":11297,"date":"2022-12-29T21:20:07","date_gmt":"2022-12-29T13:20:07","guid":{"rendered":"http:\/\/139.9.1.231\/?p=11297"},"modified":"2022-12-30T17:20:16","modified_gmt":"2022-12-30T09:20:16","slug":"gp123","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/12\/29\/gp123\/","title":{"rendered":"GPT\u7cfb\u5217\u8bba\u6587\uff1a\u751f\u6210\u5f0f\u9884\u8bad\u7ec3\u4e0e\u96f6\u6837\u672c\u5b66\u4e60"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"284\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-70-1024x284.png\" alt=\"\" class=\"wp-image-11380\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-70-1024x284.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-70-300x83.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-70-768x213.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-70.png 1147w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n\n\n<p class=\"has-text-align-center\">\u672c\u6587\u7684\u4e3b\u8981\u53c2\u8003\u662f\u674e\u6c90\u8001\u5e08\u5173\u4e8e GPT \u7cfb\u5217\u7684\u89e3\u8bfb\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/www.bilibili.com\/video\/BV1AF411b7xQ\/?spm_id_from=333.788&amp;vd_source=dab57cea5e1b38b49ad994543c0c61bf\" target=\"_blank\"><strong>\u8bba\u6587\u7cbe\u8bfb<\/strong><\/a><\/p>\n\n\n\n<p>GPT1:&nbsp;<a rel=\"noreferrer noopener\" href=\"https:\/\/cdn.openai.com\/research-covers\/language-unsupervised\/language_understanding_paper.pdf\" target=\"_blank\">Improving Language Understanding by Generative Pre-Training<\/a>&nbsp;(Generative Pre-Train Model \u5c31\u662fGPT\u6a21\u578b\u7684\u540d\u5b57\u7531\u6765\uff09<\/p>\n\n\n\n<p>GPT2:&nbsp;<a rel=\"noreferrer noopener\" href=\"https:\/\/cdn.openai.com\/better-language-models\/language_models_are_unsupervised_multitask_learners.pdf\" target=\"_blank\">Language Models are Unsupervised Multitask Learners<\/a><\/p>\n\n\n\n<p>GPT3:&nbsp;<a rel=\"noreferrer noopener\" href=\"https:\/\/arxiv.org\/pdf\/2005.14165.pdf\" target=\"_blank\">Language Models are Few-Shot Learners<\/a><\/p>\n\n\n\n<p class=\"has-text-align-center has-bright-blue-background-color has-background\">GPT3\u5f00\u53d1\u7684demo:  <a href=\"https:\/\/gpt3demo.com\/\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/gpt3demo.com\/<\/a><\/p>\n\n\n\n<p class=\"has-text-align-center\" id=\"6b3f\"><a href=\"https:\/\/towardsdatascience.com\/gpt-3-demos-use-cases-implications-77f86e540dc1\" target=\"_blank\" rel=\"noreferrer noopener\">GPT-3: Demos, Use-cases, Implications<\/a><\/p>\n\n\n\n<p id=\"2b50\">More concretely:<\/p>\n\n\n\n<ul><li>Language model performance scales as a\u00a0<a rel=\"noreferrer noopener\" href=\"https:\/\/lambdalabs.com\/blog\/demystifying-gpt-3\/#5\" target=\"_blank\"><strong>power-law<\/strong><\/a>\u00a0of model size, dataset size, and the amount of computation.<\/li><li>A language model trained on enough data can solve NLP tasks that it has never encountered. In other words, GPT-3 studies the model as a general solution for many downstream jobs\u00a0<strong>without fine-tuning<\/strong>.<\/li><\/ul>\n\n\n\n<h2>\u5173\u4e8eBERT\u548cGPT<\/h2>\n\n\n\n<p>Transformer\/BERT\/GPT \u65f6\u95f4\u7ebf\uff1aTransformer &#8212;&gt; GPT &#8212;&gt; BERT &#8212;-&gt; GPT2 &#8212;&gt; GPT3\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u53bb\u67e5 GPT \u7cfb\u5217\u548c BERT \u7684\u5f15\u7528\u6570\u91cf\uff0c\u4f1a\u53d1\u73b0 BERT \u4e00\u7bc7\u7684\u5f15\u7528\u6bd4 GPT \u7cfb\u5217\u4e09\u7bc7\u52a0\u8d77\u6765\u8fd8\u591a\u51e0\u500d\uff0c\u56e0\u6b64 BERT \u5728\u5b66\u754c\u5f71\u54cd\u529b\u66f4\u5927\u662f\u6bcb\u5eb8\u7f6e\u7591\u7684\u3002\u4f46\u8fd9\u5e76\u4e0d\u80fd\u8bf4\u660e BERT \u7684\u9884\u8bad\u7ec3\u4efb\u52a1\u5c31\u6bd4 GPT \u7684\u66f4 \u201c\u597d\u201d\u3002\u9996\u5148\uff0cGPT \u65e9\u4e8e BERT \u63d0\u51fa\u5728\u65e0\u76d1\u7763\u7684\u8bed\u6599\u4e0a\u8fdb\u884c\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\uff0cBERT \u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u4e5f\u662f\u53d7\u5230 GPT \u7684\u542f\u53d1\u3002\u5176\u6b21\uff0cGPT \u7684\u9884\u8bad\u7ec3\u4efb\u52a1\u662f\u6807\u51c6\u7684\u8bed\u8a00\u6a21\u578b\uff08Language Model\uff09\uff0c\u5373\u81ea\u56de\u5f52\u5f0f\uff08auto-regressive\uff09\u5730\u9884\u6d4b\u53e5\u5b50\u4e2d\u7684\u4e0b\u4e00\u4e2a\u5355\u8bcd\uff0c\u76f8\u6bd4\u4e8e BERT \u201c\u5b8c\u5f62\u586b\u7a7a\u201d \u5f0f\u7684\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u65e0\u7591\u8981\u96be\u4e0a\u8bb8\u591a\u3002\u8fd9\u4f7f\u5f97 GPT \u5fc5\u987b\u6a21\u578b\u591f\u5927\u3001\u6570\u636e\u591f\u591a\u624d\u80fd\u8bad\u7ec3\u8d77\u6765\uff0c\u5f97\u5230\u6bd4\u8f83\u597d\u7684\u7ed3\u679c\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 BERT \u53ea\u6709\u4e00\u7bc7\u8bba\u6587\uff0c\u800c GPT \u8fd8\u6709 GPT-2\u3001GPT-3\uff0c\u901a\u8fc7\u4e0d\u65ad\u6269\u589e\u6a21\u578b\u548c\u6570\u636e\u7684\u89c4\u6a21\uff0c\u6700\u7ec8\u4f7f\u5f97 GPT-3 \u6709\u5982\u6b64\u60ca\u8273\u7684\u6548\u679c\u3002\u5bf9\u4e8e\u8fd9\u6837\u56f0\u96be\u7684\u9884\u8bad\u7ec3\u4efb\u52a1\u548c\u5de8\u5927\u7684\u8d44\u6e90\u9700\u6c42\uff0c\u4e00\u822c\u7684\u516c\u53f8\u6216\u4e2a\u4eba\u6839\u672c\u73a9\u4e0d\u8f6c\u3002\u800c BERT \u7531\u4e8e\u4efb\u52a1\u96be\u5ea6\u8f83\u5c0f\uff0c\u76f8\u5bf9\u5e76\u4e0d\u9700\u8981\u90a3\u4e48\u591a\u8d44\u6e90\u5c31\u53ef\u4ee5\u8fdb\u884c\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\uff0c\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 BERT \u7684\u540e\u7eed\u5de5\u4f5c\u90a3\u4e48\u591a\uff08\u4f53\u73b0\u5728\u5f15\u7528\u91cf\u4e0a\uff09\u3002\u4f46\u662f\uff0c\u8bed\u8a00\u6a21\u578b\u9884\u6d4b\u4e0b\u4e00\u4e2a\u5355\u8bcd\u7684\u751f\u6210\u5f0f\u4efb\u52a1\uff0c\u4f7f\u5f97 GPT \u7684\u4e0a\u9650\u6781\u9ad8\uff0cGPT-3 \u901a\u8fc7\u5199\u51fa\u8db3\u4ee5\u4ee5\u5047\u4e71\u771f\u7684\u6587\u7ae0\uff0c\u6210\u4e3a\u4e86\u6700\u706b\u51fa\u5708\u7684 NLP \u6a21\u578b\u3002\u53e6\u5916\uff0c\u7531\u4e8e\u8bed\u8a00\u6a21\u578b\u751f\u6210\u5f0f\u4efb\u52a1\u7684\u7075\u6d3b\u6027\u548c\u5de8\u5927\u7684\u9884\u8bad\u7ec3\u89c4\u6a21<strong>\uff0cGPT \u751a\u81f3\u53ef\u4ee5\u4e0d\u9700\u8981\uff08\u66f4\u65b0\u6a21\u578b\u53c2\u6570\u7684\uff09\u5fae\u8c03\uff0c\u800c\u662f\u901a\u8fc7\u6587\u672c prompt \u63d0\u793a\uff0c\u5c31\u53ef\u4ee5\u76f4\u63a5\u5904\u7406\u4e0b\u6e38\u4efb\u52a1\u3002<\/strong><\/p>\n\n\n\n<p>\u53e6\u5916\uff0c\u5f88\u591a\u4eba\u559c\u6b22\u4ece\u4ece\u6a21\u578b\u7ed3\u6784\u4e0a\u6765\u5c06 BERT \u548c GPT \u8fdb\u884c\u533a\u5206\uff1aBERT \u4f7f\u7528\u4e86 Transformer \u7684\u7f16\u7801\u5668\uff0c\u9002\u5408\u4e8e\u5224\u522b\u5f0f\u4efb\u52a1\uff1bGPT \u4f7f\u7528\u4e86 Transformer \u7684\u89e3\u7801\u5668\uff0c\u9002\u5408\u4e8e\u751f\u6210\u5f0f\u4efb\u52a1\u3002\u7136\u800c\uff0c\u4f7f\u7528\u4ec0\u4e48\u6837\u7684\u6a21\u578b\u7ed3\u6784\u5e76\u4e0d\u662f BERT \u548c GPT \u7684\u672c\u8d28\u533a\u522b\u3002\u4e8c\u8005\u7684\u672c\u8d28\u533a\u522b\u5728\u4e8e\u9009\u7528\u4e86\u4ec0\u4e48\u6837\u7684\u9884\u8bad\u7ec3\u76ee\u6807\u51fd\u6570\uff0c\u9009\u7528 Transformer \u7684\u7f16\u7801\u5668\u6216\u89e3\u7801\u5668\u53ea\u662f\u5728\u786e\u5b9a\u4e86\u76ee\u6807\u51fd\u6570\u4e4b\u540e\u7684\u5fc5\u7136\u9009\u62e9\u3002GPT \u9009\u7528\u7684\u662f\u6807\u51c6\u8bed\u8a00\u6a21\u578b\u7684\u76ee\u6807\u51fd\u6570\uff0c\u9884\u6d4b\u53e5\u5b50\u4e2d\u7684\u4e0b\u4e00\u4e2a\u5355\u8bcd\uff0c\u6b64\u65f6\u6a21\u578b\u5e94\u8be5\u53ea\u80fd\u770b\u5230\u5f53\u524d\u8bcd\u548c\u5b83\u4e4b\u524d\u7684\u8bcd\uff0c\u6240\u4ee5\u5fc5\u987b\u5c06\u5f53\u524d\u8bcd\u540e\u9762\u7684\u8bcd\u5168\u90e8 mask \u6389\uff0c\u6545\u800c\u9009\u7528\u5e26\u6709 masked self-attention \u7684 Transfomer \u89e3\u7801\u5668\uff1b\u800c BERT \u662f\u8bbe\u8ba1\u4e86\u4e00\u79cd \u201c\u5b8c\u5f62\u586b\u7a7a\u201d \u5f0f\u7684\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u6839\u636e\u5f53\u524d\u8bcd\u524d\u540e\u7684\u5185\u5bb9\u8fd8\u539f\u5f53\u524d\u8bcd\uff0c\u6b64\u65f6\u6a21\u578b\u5e94\u8be5\u53ef\u4ee5\u770b\u5230\u6574\u4e2a\u5e8f\u5217\u7684\u6240\u6709\u5355\u8bcd\uff08\u5f53\u524d\u8bcd\u5df2\u88ab\u66ff\u6362\u4e3a\u7279\u6b8a token\uff09\uff0c\u6545\u800c\u9009\u7528\u4e86 Transformer \u7684\u7f16\u7801\u5668\u3002\u5f53\u7136\uff0c\u5728\u8ba8\u8bba BERT \u4e0e GPT \u65f6\uff0c\u5c06\u5b83\u4eec\u5404\u81ea\u9009\u7528\u7684\u67b6\u6784\u4f5c\u4e3a\u76f4\u89c2\u7684\u533a\u5206\u65b9\u5f0f\u4e5f\u662f\u6ca1\u6709\u95ee\u9898\u7684\u3002<\/p>\n\n\n\n<h2>GPT-1<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>Paper\uff1a<a href=\"https:\/\/s3-us-west-2.amazonaws.com\/openai-assets\/research-covers\/language-unsupervised\/language_understanding_paper.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Improving Language Understanding by Generative Pre-Training<\/a><\/p><\/blockquote>\n\n\n\n<h3>\u524d\u8a00<\/h3>\n\n\n\n<p>    <strong>GPT \u9996\u5148\u63d0\u51fa\u4e86\u5728\u65e0\u76d1\u7763\u7684\u5927\u89c4\u6a21\u9884\u6599\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u518d\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u8fdb\u884c\u5fae\u8c03\u7684\u8bad\u7ec3\u8303\u5f0f\u3002<\/strong>\u81f3\u4e8e\u4e3a\u4ec0\u4e48\u4f7f\u7528 Transformer \u6a21\u578b\uff0c\u800c\u975e RNN\uff0c\u4f5c\u8005\u6307\u51fa\uff1aTransformer \u6a21\u578b\u6709\u66f4\u7ed3\u6784\u5316\u7684\u8bb0\u5fc6\uff08more structured memory\uff09\uff0c\u80fd\u591f\u66f4\u597d\u5730\u5904\u7406\u6587\u672c\u4e2d\u7684\u957f\u8ddd\u79bb\uff08long-term\uff09\u4f9d\u8d56\u5173\u7cfb\uff0c\u4ece\u800c\u80fd\u66f4\u597d\u5730\u62bd\u53d6\u51fa\u53e5\u5b50\u5c42\u9762\u548c\u6bb5\u843d\u5c42\u9762\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u56e0\u6b64\u5728\u8fc1\u79fb\u5b66\u4e60\u4e2d\uff0cTranformer \u5b66\u4e60\u5230\u7684\u7279\u5f81\u66f4\u52a0\u7a33\u5065\u3002\u5728\u8fc1\u79fb\u5b66\u4e60\u65f6\uff0cGPT \u8bbe\u8ba1\u4e86\u5404\u79cd\u4efb\u52a1\u76f8\u5173\uff08task-specific\uff09\u7684\u8f93\u5165\u8868\u793a\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u8fd9\u91cc\u6240\u8c13\u7684\u66f4\u7ed3\u6784\u5316\u7684\u8bb0\u5fc6\u3001\u957f\u8ddd\u79bb\u6587\u672c\u4fe1\u606f\u7684\u8bba\u8ff0\uff0c\u7b14\u8005\u662f\u8fd9\u6837\u7406\u89e3\u7684\uff1aRNN\uff08\u5982 LSTM\uff09 \u9700\u8981\u4e00\u6b65\u4e00\u6b65\u5730\u5904\u7406\u5e8f\u5217\u5185\u5bb9\uff0c\u5982\u679c\u5e8f\u5217\u8ddd\u79bb\u8fc7\u957f\uff0c\u53ef\u80fd\u8d70\u5230\u540e\u9762\u65f6\uff0c\u524d\u9762\u7684\u4fe1\u606f\u4f1a\u6709\u6240\u4e22\u5931\uff1b\u800c\u5728 Transformer \u4e2d\uff0c\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u662f\u5b8c\u5168\u5e76\u884c\u7684\uff0c\u5e8f\u5217\u7684\u4f4d\u7f6e\u4fe1\u606f\u662f\u901a\u8fc7\u4f4d\u7f6e\u5d4c\u5165\u6765\u7f16\u7801\u7684\uff0c\u5c31\u4e0d\u4f1a\u6709\u8fd9\u4e2a\u95ee\u9898\uff0c\u5373\u674e\u5b8f\u6bc5\u8001\u5e08\u6240\u8bf4\u7684\uff1a\u201c\u5929\u6daf\u82e5\u6bd4\u90bb\u201d\u3002<\/p><\/blockquote>\n\n\n\n<h3>\u65b9\u6cd5<\/h3>\n\n\n\n<p>\u539f\u6587\u65b9\u6cd5\u90e8\u5206\u5206\u4e3a\u4e09\u4e2a\u5c0f\u8282\uff0c\u5206\u522b\u4ecb\u7ecd\u5982\u4f55\u5728\u65e0\u6807\u6ce8\u7684\u6570\u636e\u4e0a\u8fdb\u884c\u81ea\u76d1\u7763\u9884\u8bad\u7ec3\u3001\u600e\u6837\u8fdb\u884c\u5fae\u8c03\u3001\u600e\u6837\u5bf9\u4e8e\u4e0d\u540c\u7684 NLP \u4e0b\u6e38\u4efb\u52a1\u6784\u9020\u8f93\u5165\u3002<\/p>\n\n\n\n<h4>\u9884\u8bad\u7ec3\uff1a<\/h4>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"567\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-71-1024x567.png\" alt=\"\" class=\"wp-image-11391\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-71-1024x567.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-71-300x166.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-71-768x426.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-71.png 1200w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h4>\u5fae\u8c03<\/h4>\n\n\n\n<p>\u65b9\u6cd5\u90e8\u5206\u7684\u7b2c\u4e8c\u5c0f\u8282\u4ecb\u7ecd\u4e86\u5982\u4f55\u5728\u9884\u8bad\u7ec3\u5b8c\u6210\u4e4b\u540e\uff0c\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n\n\n\n<p>\u5047\u8bbe\u6709\u5e26\u6807\u7b7e\u6570\u636e\u96c6C \uff0c\u5176\u4e2d\u6bcf\u4e2a\u6837\u672c\u662f\u4e00\u4e2a\u7531\u4e00\u7cfb\u5217\u5355\u8bcd\u7ec4\u6210\u7684\u53e5\u5b50\u548c\u6807\u7b7e y \u7ec4\u6210\u3002\u5c06\u53e5\u5b50\u8f93\u5165\u5230 GPT \u6a21\u578b\u4e2d\uff0c\u53d6\u6700\u540e\u4e00\u4e2a transformer block \u6700\u540e\u4e00\u4e2a\u5355\u8bcd\u7684\u8f93\u51fa\u7279\u5f81\uff0c\u5c06\u5b83\u9001\u5165\u5230\u7ebf\u6027\u5c42\u4e2d\u8fdb\u884c\u9884\u6d4b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"137\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-72-1024x137.png\" alt=\"\" class=\"wp-image-11394\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-72-1024x137.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-72-300x40.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-72-768x103.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-72.png 1136w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3>\u4e0d\u540c\u4efb\u52a1\u7684\u8f93\u51fa\u6784\u9020<\/h3>\n\n\n\n<p>\u4ecb\u7ecd\u5b8c\u5982\u4f55\u5fae\u8c03\u4e4b\u540e\uff0c\u63a5\u4e0b\u6765\u5c31\u8981\u4ecb\u7ecd\u5982\u4f55\u5c06 NLP \u4e2d\u4e0d\u540c\u7684\u4e0b\u6e38\u4efb\u52a1\u7684\u8f93\u5165\u8868\u793a\u6210\u7b2c\u4e8c\u5c0f\u8282\u4e2d\u53e5\u5b50+\u6807\u7b7e  \u7684\u5f62\u5f0f\u3002\u5982\u4e0b\u56fe\u53f3\u4fa7\u6240\u793a\uff0c\u56fe\u4e2d\u5c55\u793a\u4e86\u51e0\u79cd\u4e0d\u540c\u7c7b\u578b\u7684 NLP \u4e0b\u6e38\u4efb\u52a1\u9002\u914d GPT \u9884\u8bad\u7ec3\u6a21\u578b\u7684\u8f93\u5165\u6784\u9020\u65b9\u6cd5\uff1a<\/p>\n\n\n\n<p><strong>\u5206\u7c7b\u4efb\u52a1<\/strong><\/p>\n\n\n\n<ul><li>\u4efb\u52a1\u7b80\u4ecb\uff1a\u4efb\u52a1\u7ed9\u5b9a\u4e00\u6bb5\u6587\u672c\uff0c\u8f93\u51fa\u5206\u7c7b\u7ed3\u679c\u3002\u4f8b\u5982\uff1a\u60c5\u611f\u5206\u7c7b\u3002<\/li><li>\u6784\u9020\u65b9\u6cd5\uff1a\u5c06\u7ed9\u5b9a\u7684\u6587\u672c\u9996\u5c3e\u5404\u52a0\u4e0a\u4e00\u4e2a token Start\/Extract\uff0c\u7136\u540e\u9001\u5230 GPT \u9884\u8bad\u7ec3\u6a21\u578b\u4e2d\uff0c\u5c06\u8f93\u51fa\u7279\u5f81\u63a5\u4e00\u4e2a\u7ebf\u6027\u5c42\u8fdb\u884c\u5206\u7c7b\u3002\u5206\u7c7b\u4efb\u52a1\u4e0e\u4e4b\u524d\u5fae\u8c03\u5c0f\u8282\u4ecb\u7ecd\u7684\u505a\u6cd5\u662f\u5b8c\u5168\u4e00\u81f4\u7684\uff0c<\/li><\/ul>\n\n\n\n<p><strong>\u8574\u542b\u4efb\u52a1\uff08\u975e\u5bf9\u79f0\u6027\u53e5\u5b50\u5173\u7cfb\u4efb\u52a1\uff09<\/strong><\/p>\n\n\n\n<ul><li>\u4efb\u52a1\u7b80\u4ecb\uff1a\u7ed9\u5b9a\u4e24\u6bb5\u6587\u672c\uff0c\u5224\u65ad\u524d\u8005\u5bf9\u540e\u8005\u5173\u7cfb\u3002\u4f8b\u5982\uff1a\u8574\u542b\u4efb\u52a1\uff0c\u5224\u65ad\u7b2c\u4e00\u53e5\u5bf9\u7b2c\u4e8c\u53e5\u7684\u5173\u7cfb\u662f\u8574\u542b\/\u4e0d\u8574\u542b\/\u65e0\u5173\u3002<\/li><li>\u6784\u9020\u65b9\u6cd5\uff1a\u5c06\u4e24\u4e2a\u53e5\u5b50\u4e2d\u95f4\u6dfb\u52a0\u4e00\u4e2a\u5206\u5272 token Delim\uff0c\u7136\u540e\u5c06\u6574\u4e2a\u6587\u672c\u7684\u9996\u5c3e\u518d\u52a0\u4e0a Start\/Extract\uff0c\u9001\u5165 GPT \u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u5c06\u8f93\u51fa\u7279\u5f81\u9001\u5165\u7ebf\u6027\u5c42\u5206\u7c7b\u3002<\/li><\/ul>\n\n\n\n<p><strong>\u76f8\u4f3c\u5ea6\u4efb\u52a1\uff08\u5bf9\u79f0\u6027\u53e5\u5b50\u5173\u7cfb\u4efb\u52a1\uff09<\/strong><\/p>\n\n\n\n<ul><li>\u4efb\u52a1\u7b80\u4ecb\uff1a\u7ed9\u5b9a\u4e24\u6bb5\u6587\u672c\uff0c\u5224\u65ad\u4e8c\u8005\u5173\u7cfb\u3002\u4f8b\u5982\uff1a\u76f8\u4f3c\u5ea6\u4efb\u52a1\uff0c\u5224\u65ad\u4e24\u4e2a\u53e5\u5b50\u662f\u5426\u76f8\u4f3c\u3002<\/li><li>\u6784\u9020\u65b9\u6cd5\uff1a\u5c06\u4e24\u4e2a\u53e5\u5b50\u5206\u522b\u4f5c\u4e3a\u524d\u53e5\u6216\u540e\u53e5\uff0c\u6784\u9020\u4e24\u4e2a\u5b8c\u6574\u6587\u672c\uff0c\u5404\u81ea\u9001\u5165 GPT \u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u63d0\u53d6\u51fa\u7279\u5f81\u5e76\u8fdb\u884c\u878d\u5408\uff0c\u518d\u9001\u5165\u7ebf\u6027\u5c42\u5206\u7c7b\u3002<\/li><\/ul>\n\n\n\n<p><strong>\u591a\u9009\u4efb\u52a1<\/strong><\/p>\n\n\n\n<ul><li>\u4efb\u52a1\u7b80\u4ecb\uff1a\u7ed9\u5b9a\u4e00\u6bb5\u6587\u672c\u548c\u591a\u4e2a\u7b54\u6848\uff0c\u5224\u65ad\u54ea\u4e2a\u6b63\u786e\u3002<\/li><li>\u6784\u9020\u65b9\u6cd5\uff1a\u5c06\u7ed9\u5b9a\u6587\u672c\u548c N \u4e2a\u7b54\u6848\u7ed3\u5408\uff0c\u6784\u9020 N \u4e2a\u5b8c\u6574\u6587\u672c\uff0c\u5404\u81ea\u9001\u5165 GPT \u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u63d0\u53d6\u51fa\u7279\u5f81\u5e76\u9001\u5165\u7ebf\u6027\u5c42\uff0c\u53d6\u7f6e\u4fe1\u5ea6\u6700\u5927\u8005\u3002<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"449\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-74-1024x449.png\" alt=\"\" class=\"wp-image-11420\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-74-1024x449.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-74-300x132.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-74-768x337.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/12\/image-74.png 1391w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>GPT\u6a21\u578b\u7ed3\u6784\uff08\u5de6\uff09\u4e0e\u5fae\u8c03\u4e0b\u6e38\u4efb\u52a1\u8f93\u5165\u6784\u9020\u65b9\u5f0f\uff08\u53f3\uff09<\/p>\n\n\n\n<p>\u6ce8\u610f\u56fe\u4e2d\u7684\u5f00\u59cb\u7b26\uff08Start\uff09\u3001\u5206\u9694\u7b26\uff08Delim\uff09\u3001\u7ed3\u675f\u7b26\uff08Extract\uff09\u4e0d\u662f\u8fd9\u51e0\u4e2a\u5355\u8bcd\u672c\u8eab\uff0c\u800c\u662f\u4e09\u4e2a\u7279\u6b8a\u7684\u7b26\u53f7\u3002<\/p>\n\n\n\n<h2>GPT-2<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>Paper\uff1a<a href=\"https:\/\/d4mucfpksywv.cloudfront.net\/better-language-models\/language_models_are_unsupervised_multitask_learners.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Language Models are Unsupervised Multitask Learners<\/a><\/p><\/blockquote>\n\n\n\n<h3>\u524d\u8a00<\/h3>\n\n\n\n<p>GPT \u51fa\u73b0\u540e\u4e0d\u4e45\uff0cBERT \u5c31\u63d0\u51fa\u4e86\u3002\u901a\u8fc7\u65b0\u578b\u7684 MLM\uff08Masked Language Model\uff09\u4efb\u52a1\u548c\u66f4\u5927\u7684\u6a21\u578b\u3001\u66f4\u5927\u7684\u8bad\u7ec3\u6570\u636e\u91cf\uff0c\u5728\u591a\u9879\u6307\u6807\u4e0a\u8d85\u8d8a\u4e86 GPT\u3002GPT \u7684\u4f5c\u8005\u60f3\u8981\u518d\u6b21\u53cd\u8d85\uff0c\u9996\u5148\u80af\u5b9a\u8981\u6269\u5927\u6a21\u578b\u548c\u6570\u636e\u7684\u89c4\u6a21\u3002\u4f46\u662f\u9664\u6b64\u4e4b\u5916\uff0cGPT-2 \u8fd8\u6709\u4e00\u4e2a\u60ca\u4eba\u7684\u8bbe\u5b9a\uff1azero-shot\uff08\u96f6\u6837\u672c\uff09\u3002\u987e\u540d\u601d\u4e49\uff0c\u5728 zero-shot \u8bbe\u5b9a\u4e0b\uff0c\u6a21\u578b\u5728\u9884\u8bad\u7ec3\u5b8c\u6210\u4e4b\u540e\u4e0d\u9700\u8981\u4efb\u4f55\u4e0b\u6e38\u4efb\u52a1\u7684\u6807\u6ce8\u6570\u636e\u6765\u8fdb\u884c\u5fae\u8c03\uff0c\u800c\u662f\u76f4\u63a5\u8fdb\u884c\u9884\u6d4b\u3002<\/p>\n\n\n\n<p>GPT-2 \u5728\u7814\u7a76\u601d\u8def\u4e0a\u5e26\u7ed9\u6211\u4eec\u7684\u542f\u793a\u662f\uff1a\u6709\u65f6\u5019\u505a\u7814\u7a76\u4e0d\u4e00\u5b9a\u8981\u5728\u4e00\u4e2a\u65e2\u5b9a\u6307\u6807\u4e0a\u6b7b\u78d5\u3002\u5728\u65b9\u6cd5\u6ca1\u6709\u5927\u521b\u65b0\u7684\u60c5\u51b5\u4e0b\uff0c\u901a\u8fc7 \u201c\u5927\u529b\u51fa\u5947\u8ff9\u201d \uff0c\u5373\u4f7f\u80fd\u591f\u6bd4\u4e4b\u524d\u65b9\u6cd5\u6709\u6240\u63d0\u5347\uff0c\u6587\u7ae0\u4e5f\u4f1a\u663e\u5f97\u6709\u4e9b\u65e0\u804a\u3002\u8fd9\u65f6\u53ef\u4ee5\u601d\u8003\u4e00\u4e9b\u8bbe\u5b9a\u4e0a\u7684\u521b\u65b0\uff0c\u5982\u672c\u6587\u7684 zero-shot\uff0c\u8fd9\u65f6\u5373\u4f7f\u6307\u6807\u4e0a\u63d0\u5347\u4e0d\u591a\u751a\u81f3\u6301\u5e73\uff0c\u4e5f\u4f1a\u6709\u66f4\u6709\u65b0\u610f\u3001\u66f4\u6709\u8da3\u3002<\/p>\n\n\n\n<h3>\u65b9\u6cd5<\/h3>\n\n\n\n<p>     GPT2\u8fd8\u662f\u505a\u8bed\u8a00\u6a21\u578b\uff0c\u4f46\u662f\u5728\u505a\u5230\u4e0b\u6e38\u4efb\u52a1\u7684\u65f6\u5019\uff0c\u4f1a\u7528\u4e00\u4e2a\u53eb\u505azero-shot\u7684\u8bbe\u5b9a\uff0czero-shot\u662f\u8bf4\uff0c\u5728\u505a\u5230\u4e0b\u6e38\u4efb\u52a1\u7684\u65f6\u5019\uff0c\u4e0d\u9700\u8981\u4e0b\u6e38\u4efb\u52a1\u7684\u4efb\u4f55\u6807\u6ce8\u4fe1\u606f\uff0c\u90a3\u4e48\u4e5f\u4e0d\u9700\u8981\u53bb\u91cd\u65b0\u8bad\u7ec3\u5df2\u7ecf\u9884\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u3002\u8fd9\u6837\u5b50\u7684\u597d\u5904\u662f\u6211\u53ea\u8981\u8bad\u7ec3\u597d\u4e00\u4e2a\u6a21\u578b\uff0c\u5728\u4efb\u4f55\u5730\u65b9\u90fd\u53ef\u4ee5\u7528\u3002<br>     \u5982\u679c\u4f5c\u8005\u5c31\u662f\u5728GPT1\u7684\u57fa\u7840\u4e0a\u7528\u4e00\u4e2a\u66f4\u5927\u7684\u6570\u636e\u96c6\u8bad\u7ec3\u4e00\u4e2a\u66f4\u5927\u7684\u6a21\u578b\uff0c\u8bf4\u6211\u7684\u7ed3\u679c\u6bd4Bert\u597d\u4e00\u4e9b\uff0c\u53ef\u80fd\u4e5f\u5c31\u597d\u90a3\u4e48\u4e00\u70b9\u70b9\uff0c\u4e0d\u662f\u597d\u90a3\u4e48\u591a\u7684\u60c5\u51b5\u4e0b\uff0c\u5927\u5bb6\u4f1a\u89c9\u5f97gpt2\u8fd9\u7bc7\u6587\u7ae0\u5c31\u6ca1\u4ec0\u4e48\u610f\u601d\u4e86\uff0c\u5de5\u7a0b\u5473\u7279\u522b\u91cd\u3002\u90a3\u4e48\u6211\u6362\u4e00\u4e2a\u89d2\u5ea6\uff0c\u9009\u62e9\u4e00\u4e2a\u66f4\u96be\u7684\u95ee\u9898\uff0c\u6211\u8bf4\u505azero-shot\u3002\u867d\u7136\u7ed3\u679c\u53ef\u80fd\u6ca1\u90a3\u4e48\u5389\u5bb3\u4e86\uff0c\u6ca1\u90a3\u4e48\u6709\u4f18\u52bf\uff0c\u4f46\u662f\u65b0\u610f\u5ea6\u4e00\u4e0b\u5c31\u6765\u4e86\u3002<\/p>\n\n\n\n<p>     GPT-2 \u7684\u6a21\u578b\u8ddf GPT-1 \u4e00\u6837\uff0c\u8fd9\u91cc\u5c31\u4e0d\u518d\u8fc7\u591a\u4ecb\u7ecd\u3002\u672c\u8282\u4e3b\u8981\u6765\u8bf4\u4e00\u4e0b zero-shot \u8981\u600e\u4e48\u505a\u3002<\/p>\n\n\n\n<p class=\"has-bright-blue-background-color has-background\"><strong>\u5728 GPT-1 \u4e2d\uff0c\u6a21\u578b\u9884\u8bad\u7ec3\u5b8c\u6210\u4e4b\u540e\u4f1a\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u5fae\u8c03\uff0c\u5728\u6784\u9020\u4e0d\u540c\u4efb\u52a1\u7684\u5bf9\u5e94\u8f93\u5165\u65f6\uff0c\u6211\u4eec\u4f1a\u5f15\u5165\u5f00\u59cb\u7b26\uff08Start\uff09\u3001\u5206\u9694\u7b26\uff08Delim\uff09\u3001\u7ed3\u675f\u7b26\uff08Extract\uff09\u3002\u867d\u7136\u6a21\u578b\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u4ece\u672a\u89c1\u8fc7\u8fd9\u4e9b\u7279\u6b8a\u7b26\u53f7\uff0c\u4f46\u662f\u6bd5\u7adf\u6709\u5fae\u8c03\u9636\u6bb5\u7684\u53c2\u6570\u8c03\u6574\uff0c\u6a21\u578b\u4f1a\u5b66\u7740\u6162\u6162\u7406\u89e3\u8fd9\u4e9b\u7b26\u53f7\u7684\u610f\u601d\u3002\u73b0\u5728\uff0c\u5728 GPT-2 \u4e2d\uff0c\u8981\u505a\u7684\u662f zero-shot\uff0c\u4e5f\u5c31\u662f\u6ca1\u6709\u4efb\u4f55\u8c03\u6574\u7684\u8fc7\u7a0b\u4e86\u3002\u8fd9\u65f6\u6211\u4eec\u5728\u6784\u9020\u8f93\u5165\u65f6\u5c31\u4e0d\u80fd\u7528\u90a3\u4e9b\u5728\u9884\u8bad\u7ec3\u65f6\u6ca1\u6709\u51fa\u73b0\u8fc7\u7684\u7279\u6b8a\u7b26\u53f7\u4e86\u3002\u6240\u5e78\u81ea\u7136\u8bed\u8a00\u5904\u7406\u7684\u7075\u6d3b\u6027\u5f88\u5f3a\uff0c\u6211\u4eec\u53ea\u8981\u628a\u60f3\u8981\u6a21\u578b\u505a\u7684\u4efb\u52a1 \u201c\u544a\u8bc9\u201d \u6a21\u578b\u5373\u53ef\uff0c\u5982\u679c\u6709\u8db3\u591f\u91cf\u9884\u8bad\u7ec3\u6587\u672c\u652f\u6491\uff0c\u6a21\u578b\u60f3\u5fc5\u662f\u80fd\u7406\u89e3\u6211\u4eec\u7684\u8981\u6c42\u7684\u3002<\/strong><\/p>\n\n\n\n<p>\u4e3e\u4e2a\u673a\u5668\u7ffb\u8bd1\u7684\u4f8b\u5b50\uff0c\u8981\u7528 GPT-2 \u505a zero-shot \u7684\u673a\u5668\u7ffb\u8bd1\uff0c\u53ea\u8981\u5c06\u8f93\u5165\u7ed9\u6a21\u578b\u7684\u6587\u672c\u6784\u9020\u6210&nbsp;<code>translate english to chinese, [englist text], [chinese text]<\/code>&nbsp;\u5c31\u597d\u4e86\u3002\u6bd4\u5982\uff1a<code>translate english to chinese, [machine learning], [\u673a\u5668\u5b66\u4e60]<\/code>&nbsp;\u3002\u8fd9\u79cd\u505a\u6cd5\u5c31\u662f\u65e5\u540e\u9f0e\u9f0e\u5927\u540d\u7684 prompt\u3002<\/p>\n\n\n\n<p>\u5728\u8bad\u7ec3\u6570\u636e\u7684\u6536\u96c6\u90e8\u5206\uff0c\u4f5c\u8005\u63d0\u5230\u4ed6\u4eec\u6ca1\u6709\u4f7f\u7528 Common Crawl \u7684\u516c\u5f00\u7f51\u9875\u722c\u53d6\u6570\u636e\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u6570\u636e\u566a\u58f0\u592a\u591a\uff0c\u592a\u591a\u65e0\u610f\u4e49\u7684\u5185\u5bb9\u3002\u4ed6\u4eec\u662f\u53bb Reddit \u722c\u53d6\u4e86\u5927\u91cf\u6709\u610f\u4e49\u7684\u6587\u672c\u3002\u4f5c\u8005\u8fd8\u6307\u51fa\uff0c\u5728 Reddit \u7684\u9ad8\u8d28\u91cf\u6587\u672c\u4e2d\uff0c\u5f88\u53ef\u80fd\u5df2\u7ecf\u6709\u7c7b\u4f3c zero-shot \u6784\u9020\u65b9\u5f0f\u7684\u6837\u672c\u4f9b\u6a21\u578b\u5b66\u4e60\u3002\u4e00\u4e2a\u673a\u5668\u7ffb\u8bd1\u7684\u4f8b\u5b50\u5982\u4e0b\u6240\u793a\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>In a now-deleted post from Aug. 16, Soheil Eid, Tory candidate in the riding of Joliette, wrote in French: \u201d<strong>Mentez mentez, il en restera toujours quelque chose,<\/strong>\u201d which translates as, \u201d<strong>Lie lie and something will always remain.<\/strong>\u201d<\/p><\/blockquote>\n\n\n\n<h3>\u5b9e\u9a8c<\/h3>\n\n\n\n<p>\u5728\u4e0e\u540c\u6837\u4e3a zero-shot \u6a21\u578b\u7684\u5bf9\u6bd4\u4e0a\uff0c\u80af\u5b9a\u662f\u540a\u6253\u4e86\u4e4b\u524d\u7684 SOTA\uff0c\u8fd9\u91cc\u5c31\u4e0d\u5c55\u793a\u4e86\u3002\u53ef\u4ee5\u5173\u6ce8\u4e00\u4e0b\u4e0b\u9762 GPT-2 \u6a21\u578b zero-shot \u6027\u80fd\u5173\u4e8e\u6a21\u578b\u89c4\u6a21\u7684\u66f2\u7ebf\u3002\u5728\u6709\u4e9b\u4efb\u52a1\u4e0a\u5df2\u7ecf\u63a5\u8fd1\u3001\u8d85\u8fc7\u4e4b\u524d\u67d0\u4e9b\u6709\u76d1\u7763\u7684\u65b9\u6cd5\uff1b\u5728\u6bd4\u8f83\u56f0\u96be\u7684\u4efb\u52a1\u4e0a\uff0c\u6bd4\u5982\u5f00\u653e\u57df\u95ee\u7b54\uff0c\u5b8c\u5168\u8fd8\u770b\u4e0d\u5230\u522b\u4eba\u7684\u5f71\u5b50\u3002\u7136\u800c\uff0c\u770b\u770b\u66f2\u7ebf\u672b\u7aef\u6027\u80fd\u968f\u6a21\u578b\u89c4\u6a21\u63d0\u5347\u7684\u8d8b\u52bf\uff0c\u5b8c\u5168\u6ca1\u6709\u6536\u655b\u7684\u610f\u601d\uff0c\u8fd9\u6700\u540e\u4e00\u6bb5\u7fd8\u8d77\u7684\u66f2\u7ebf\uff0c\u662d\u793a\u7740 GPT-3 \u7ee7\u7eed\u5927\u529b\u51fa\u5947\u8ff9\uff0c\u4ece\u91cf\u53d8\u5230\u8d28\u53d8\u7684\u5e0c\u671b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic1.zhimg.com\/v2-092f1dc7ddf4ba6ac27254017cf2c970_r.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<h3>GPT-3<\/h3>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>Paper\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2005.14165\" target=\"_blank\" rel=\"noreferrer noopener\">Language Models are Few-Shot Learners<\/a><\/p><\/blockquote>\n\n\n\n<h3>\u524d\u8a00<\/h3>\n\n\n\n<p>\u6839\u636e\u6c90\u795e\u5173\u4e8e\u8bba\u6587\u4ef7\u503c\u7ed9\u51fa\u7684\u516c\u5f0f\uff1a\u8bba\u6587\u4ef7\u503c = \u6709\u6548\u6027 * \u65b0\u610f\u5ea6 * \u95ee\u9898\u7684\u5927\u5c0f\u3002GPT-2 \u867d\u7136\u901a\u8fc7 zero-shot \u7684\u8bbe\u5b9a\uff0c\u5c06\u81ea\u5df1\u7684\u65b0\u610f\u5ea6\u51f8\u663e\u4e86\u51fa\u6765\uff0c\u4f46\u662f\u6709\u6548\u6027\uff08\u7edd\u5bf9\u6027\u80fd\uff09\u8fd8\u662f\u4e0d\u592a\u4ee4\u4eba\u6ee1\u610f\u3002GPT-3 \uff0c\u4f17\u6240\u5468\u77e5\u7684 \u201c\u5927\u529b\u51fa\u5947\u8ff9\u201d \u5f0f\u7684\u6587\u7ae0\uff0c\u901a\u8fc7\u6d77\u91cf\u6570\u636e\u8bad\u7ec3\u4e86\u4e00\u4e2a 175 Billion \u53c2\u6570\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c\u6027\u80fd\u76f4\u63a5\u62c9\u6ee1\u3002\u751a\u81f3\u6709\u70b9\u4ece\u91cf\u53d8\u5230\u8d28\u53d8\u7684\u610f\u601d\uff0cGPT-3 \u901a\u8fc7\u81ea\u56de\u5f52\u5f0f\u8bed\u8a00\u6a21\u578b\u7684\u751f\u6210\u80fd\u529b\uff0c\u53ef\u4ee5\u751f\u6210\u4e00\u4e9b\u50cf\u6a21\u50cf\u6837\u7684\u6587\u7ae0\uff0c\u6709\u65f6\u4eba\u7c7b\u90fd\u5f88\u96be\u8bfb\u51fa\u8fd9\u4e9b\u6587\u7ae0\u662f\u51fa\u81ea\u4e8e AI \u6a21\u578b\u751f\u6210\uff0c\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 GPT-3 \u80fd\u591f\u6210\u4e3a NLP \u9886\u57df\u6700\u706b\u51fa\u5708\u7684\u6a21\u578b\uff0c\u6587\u672c\u751f\u6210\u80fd\u529b\u4f7f\u5f97\u5b83\u73a9\u6cd5\u4f17\u591a\u3002\u5728\u4efb\u52a1\u8bbe\u5b9a\u4e0a\uff0cGPT-3 \u6ca1\u6709\u56fa\u5b88\u4e8e GPT-2 \u7684 zero-shot \u65b9\u5f0f\u3002\u56e0\u4e3a\u5373\u4f7f\u5bf9\u4e8e\u4eba\u7c7b\u6765\u8bf4\uff0c\u8981\u5b8c\u6210\u4e00\u4e2a\u65b0\u4efb\u52a1\uff0c\u5982\u679c\u4e00\u4e2a\u793a\u4f8b\u4e5f\u4e0d\u7ed9\u7684\u8bdd\uff0c\u4e5f\u6709\u70b9\u5f3a\u4eba\u6240\u96be\u4e86\u3002\u5982\u6807\u9898\u6240\u793a\uff0cGPT-3 \u91c7\u7528\u4e86 few-shot \u7684\u4efb\u52a1\u8bbe\u5b9a\uff0c\u5373\u7ed9\u51fa\u4e0b\u6e38\u4efb\u52a1\u7684\u4e00\u4e24\u4e2a\u4f8b\u5b50\uff0c\u7136\u540e\u8981\u6c42\u6a21\u578b\u5bf9\u8be5\u4efb\u52a1\u7684\u65b0\u95ee\u9898\u7ed9\u51fa\u9884\u6d4b\u3002\u5f53\u7136\uff0c\u5982\u6b64\u5927\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u5373\u4f7f\u662f\u4e00\u4e24\u4e2a\u6837\u672c\uff0c\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u5fae\u8c03\u6a21\u578b\u6743\u91cd\u4e5f\u5f88\u8d39\u52b2\u3002\u56e0\u6b64\uff0cGPT-3 \u4e2d\u6240\u8c13\u7684 \u201cfew-shot\u201d\uff0c\u4e0e\u4e00\u822c\u7684\u6839\u636e\u652f\u6301\u96c6\uff08\u4e0b\u6e38\u4efb\u52a1\u793a\u4f8b\uff09\u8fdb\u884c\u68af\u5ea6\u4e0b\u964d\u66f4\u65b0\u53c2\u6570\u7684 few-shot \u65b9\u6cd5\u4e0d\u540c<strong>\uff0c\u5b83\u662f\u5229\u7528\u81ea\u7136\u8bed\u8a00\u7684\u7075\u6d3b\u6027\uff0c\u5c06\u652f\u6301\u96c6\u793a\u4f8b\u653e\u5230 prompt \u91cc\uff0c\u8ba9\u6a21\u578b\u81ea\u5df1\u7406\u89e3\u793a\u4f8b\uff0c\u5b8c\u6210\u4e0b\u6e38\u4efb\u52a1 few-shot \u9884\u6d4b\u3002<\/strong><\/p>\n\n\n\n<p>\u4e0b\u56fe\u5c55\u793a\u4e86 GPT-3 \u5728\u4e0d\u540c\u7684 NLP \u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u968f\u6a21\u578b\u89c4\u6a21\u7684\u53d8\u5316\uff0c\u6a59\u3001\u7eff\u3001\u84dd\u5206\u522b\u4ee3\u8868 few-\/one-\/zero shot \u65b9\u5f0f\uff0c\u6de1\u5316\u7684\u66f2\u7ebf\u662f\u5728\u4e0d\u540c\u4efb\u52a1\u4e0a\u5404\u81ea\u7684\u51c6\u786e\u7387\u3002\u5b9e\u7ebf\u662f\u5e73\u5747\u51c6\u786e\u7387\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\uff0c\u6027\u80fd\u8fd8\u662f\u6709\u4e00\u5b9a\u63d0\u5347\u7684\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic1.zhimg.com\/v2-46caa4983003977c391973f8f1d13424_r.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<h3>\u65b9\u6cd5<\/h3>\n\n\n\n<p>GPT-3 \u7684\u9884\u8bad\u7ec3\u65b9\u5f0f\u548c\u4e4b\u524d\u8fd8\u662f\u4e00\u6837\u7684\uff0c\u6a21\u578b\u7ed3\u6784\u4e5f\u6539\u52a8\u4e0d\u5927\u3002\u8fd8\u662f\u5728 Transformer \u89e3\u7801\u5668\u4e0a\u505a\u6807\u51c6\u8bed\u8a00\u6a21\u578b\u7684\u9884\u8bad\u7ec3\uff0c\u4f46\u662f\u6a21\u578b\u89c4\u6a21\u548c\u6570\u636e\u89c4\u6a21\u5927\u4e86\u51e0\u4e2a\u6570\u91cf\u7ea7\u3002\u8fd9\u91cc\u6211\u4eec\u8fd8\u662f\u4e3b\u8981\u6765\u770b\u4e00\u4e0b GPT-3 \u4e2d\u6240\u8c13\u7684 few-\/one-\/zero- shot \u65b9\u5f0f\u5206\u522b\u662f\u4ec0\u4e48\u610f\u601d\u3002<\/p>\n\n\n\n<p>\u4e0b\u56fe\u5c55\u793a\u4e86 GPT-3 \u4e2d\u7684 few-\/one-\/zero- shot \u65b9\u5f0f\u4e0e\u5e38\u89c4\u7684\u5fae\u8c03\u65b9\u5f0f\u3002<\/p>\n\n\n\n<ul><li>\u5fae\u8c03\u65b9\u5f0f\u7684\u5c0f\u6837\u672c\u5b66\u4e60\uff0c\u9700\u8981\u6839\u636e\u7ed9\u51fa\u7684\u4e0b\u6e38\u4efb\u52a1\u6837\u672c\u548c\u6807\u6ce8\uff0c\u6784\u9020\u635f\u5931\u51fd\u6570\uff0c\u65b9\u5411\u4f20\u64ad\u68af\u5ea6\uff0c\u66f4\u65b0\u6a21\u578b\u6743\u91cd\uff0c\u7136\u540e\u8fdb\u884c\u9884\u6d4b\u3002GPT-3 \u4e2d\u5b8c\u5168\u6ca1\u6709\u91c7\u53d6\u8fd9\u79cd\u65b9\u5f0f\u3002<\/li><li>Zero-shot\uff0c\u7ed9\u5b9a\u4efb\u52a1\u63cf\u8ff0\uff0c\u5982&nbsp;<code>Translate English to French<\/code>\uff0c\u7136\u540e\u76f4\u63a5\u7ed9\u51fa\u95ee\u9898\uff0c\u8981\u6c42\u6a21\u578b\u7ed9\u51fa\u7b54\u6848\u3002\u8fd9\u79cd\u65b9\u5f0f\u4e0e GPT-2 \u4e00\u81f4\u3002<\/li><li><strong>One-shot\uff0c\u7ed9\u5b9a\u4efb\u52a1\u63cf\u8ff0\uff0c\u7136\u540e\u7ed9\u4e00\u4e2a\u4f8b\u5b50\uff0c\u5305\u62ec\u95ee\u9898\u548c\u7b54\u6848\uff0c\u5982&nbsp;<code>sea otter =&gt; loutre de mer<\/code>\uff0c\u4e4b\u540e\u518d\u7ed9\u51fa\u95ee\u9898\uff0c\u5c06\u4e0a\u8ff0\u6574\u4e00\u6bb5\u6587\u672c\u4f5c\u4e3a\u8f93\u5165\uff0c\u8981\u6c42\u6a21\u578b\u7ed9\u51fa\u7b54\u6848\u3002\u8fd9\u79cd\u65b9\u5f0f\u671f\u671b\u6a21\u578b\u5229\u7528\u9884\u8bad\u7ec3\u9636\u6bb5\u6d77\u91cf\u7684\u6587\u672c\u6570\u636e\u79ef\u7d2f\u548c Tranformer \u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u7406\u89e3\u95ee\u9898\u548c\u793a\u4f8b\uff0c\u7136\u540e\u4eff\u7167\u793a\u4f8b\u7ed9\u51fa\u9884\u6d4b\u3002<br>\u7b14\u8005\u8ba4\u4e3a\u8fd9\u79cd\u65b9\u5f0f\u53ef\u884c\u7684\u6839\u672c\u539f\u56e0\u662f\u81ea\u7136\u8bed\u8a00\u7684\u7075\u6d3b\u6027\u548c\u751f\u6210\u5f0f\u6a21\u578b\u7684\u521b\u9020\u6027\uff0c\u4f7f\u5f97\u6211\u4eec\u80fd\u591f\u76f4\u63a5\u8ddf\u6a21\u578b\u8fdb\u884c\u4ea4\u4e92\uff0c\u628a\u8981\u505a\u4ec0\u4e48\u4efb\u52a1\u3001\u4efb\u52a1\u793a\u4f8b\u76f4\u63a5 \u201c\u6253\u5b57\u544a\u8bc9\u5b83\u201d\u3002\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\uff0c\u597d\u50cf\u5f88\u96be\u505a\u5230\u7c7b\u4f3c\u7684\u4e8b\u60c5\u3002<\/strong><\/li><li>Few-shot\uff0c\u4e0e One-shot \u7c7b\u4f3c\uff0c\u53ea\u662f\u7ed9\u7684\u793a\u4f8b\u66f4\u591a\u3002<\/li><\/ul>\n\n\n\n<p><strong>GPT-3 \u4e2d\u7684\u4efb\u52a1\u8bbe\u5b9a\u5f88\u60ca\u8273\uff0c\u4f46\u662f\u7ec6\u60f3\u4e4b\u4e0b\uff0c\u4e5f\u662f\u65e0\u5948\u4e4b\u4e3e\u5e76\u4e14\u4e5f\u6709\u7f3a\u70b9\u3002\u4e00\u65b9\u9762\uff0c\u6a21\u578b\u89c4\u6a21\u5b9e\u5728\u592a\u5927\uff0c\u5fae\u8c03\u6765\u66f4\u65b0\u6743\u91cd\u53c2\u6570\u4e0d\u53ef\u884c\uff0c\u53ea\u597d\u91c7\u7528 few-shot \u7684\u65b9\u5f0f\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u6a21\u578b\u6743\u91cd\u4e0d\u80fd\u66f4\u65b0\uff0c\u6bcf\u6b21\u7406\u89e3\u4e0b\u6e38\u4efb\u52a1\u4e4b\u540e\u4e0d\u80fd\u4fdd\u5b58\u4e0b\u6765\uff0c\u4e5f\u5c31\u662f\u8bf4\u6bcf\u6b21\u505a\u540c\u4e00\u4e2a\u4e0b\u6e38\u4efb\u52a1\u90fd\u8981\u7ed9\u540c\u6837\u7684\u4f8b\u5b50\u3002\u8fd8\u6709\uff0c\u4e0b\u6e38\u4efb\u52a1\u7684\u793a\u4f8b\u4e5f\u4e0d\u80fd\u592a\u591a\uff0c\u56e0\u4e3a\u6a21\u578b\u53ef\u80fd\u65e0\u6cd5\u5904\u7406\u8fc7\u957f\u7684\u8f93\u5165\u5e8f\u5217\u3002\u5982\u679c\u5728\u6211\u4eec\u7684\u5b9e\u9645\u4e0b\u6e38\u4efb\u52a1\u4e2d\u786e\u5b9e\u6709\u4e0d\u5c11\u53ef\u4f9b\u5b66\u4e60\u7684\u6837\u672c\uff0cGPT-3 \u6050\u6015\u4e0d\u662f\u4e00\u4e2a\u597d\u7684\u9009\u62e9\u3002\u6240\u4ee5\uff0c\u867d\u7136 GPT-3 \u80fd\u505a\u5230\u7684\u4e8b\u60c5\u4f3c\u4e4e\u542c\u8d77\u6765\u66f4\u63a5\u8fd1 \u201c\u4eba\u5de5\u667a\u80fd\u201d\uff0c\u4f46\u662f\u76f8\u5173\u7684\u8ddf\u8fdb\u5de5\u4f5c\u5e76\u4e0d\u591a\u3002<\/strong><\/p>\n\n\n\n<p>\u8fd9\u5e94\u8be5\u662f\u4f5c\u8005\u4eec\u5145\u5206\u6316\u6398\u6a21\u578b\u80fd\u529b\uff0c\u89c4\u907f\u6a21\u578b\u7f3a\u70b9\uff0c\u626c\u957f\u907f\u77ed\u8bbe\u8ba1\u51fa\u7684\u4efb\u52a1\u8bbe\u5b9a\uff0c\u8fd9\u79cd\u601d\u8def\u503c\u5f97\u5b66\u4e60\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic3.zhimg.com\/v2-d13537e7374cfcd3b90bbf2bf68f4c8a_r.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u5173\u4e8e\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u7684\u8c03\u53c2\u3001\u6570\u636e\u51c6\u5907\u4e0e\u6e05\u6d17\u3001\u5de5\u7a0b\u5b9e\u8df5\u7b49\uff0cGPT-3 \u8bba\u6587\u4e2d\u4e5f\u6709\u8ba8\u8bba\uff0c\u8fd9\u91cc\u5c31\u4e0d\u63d0\u4e86\uff0c\u6709\u5174\u8da3\u53ef\u4ee5\u53c2\u8003\u539f\u6587\u3002<\/p>\n\n\n\n<p>\u518d\u540e\u9762\u5173\u4e8e\u5b9e\u9a8c\u3001GPT-3 \u7684\u4e0d\u8db3\u4ee5\u53ca\u53ef\u80fd\u7684\u793e\u4f1a\u5f71\u54cd\u4f5c\u8005\u5199\u4e86\u5f88\u591a\uff0c\u672c\u6587\u4e3b\u8981\u5173\u6ce8\u7b97\u6cd5\u90e8\u5206\uff0c\u540e\u9762\u5c31\u4e0d\u4e00\u4e00\u4ecb\u7ecd\u4e86\uff0c\u540c\u6837\u8bf7\u53c2\u8003\u539f\u6587\u3002<\/p>\n\n\n\n<h3>Ref<\/h3>\n\n\n\n<ul><li><a href=\"https:\/\/www.bilibili.com\/video\/BV1AF411b7xQ\" target=\"_blank\" rel=\"noreferrer noopener\">GPT\uff0cGPT-2\uff0cGPT-3 \u8bba\u6587\u7cbe\u8bfb\u3010\u8bba\u6587\u7cbe\u8bfb\u3011<\/a><\/li><li><a href=\"https:\/\/s3-us-west-2.amazonaws.com\/openai-assets\/research-covers\/language-unsupervised\/language_understanding_paper.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Improving Language Understanding by Generative Pre-Training<\/a><\/li><li><a href=\"https:\/\/d4mucfpksywv.cloudfront.net\/better-language-models\/language_models_are_unsupervised_multitask_learners.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Language Models are Unsupervised Multitask Learners<\/a><\/li><li><a href=\"https:\/\/arxiv.org\/abs\/2005.14165\" target=\"_blank\" rel=\"noreferrer noopener\">Language Models are Few-Shot Learners<\/a><\/li><\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u7684\u4e3b\u8981\u53c2\u8003\u662f\u674e\u6c90\u8001\u5e08\u5173\u4e8e GPT \u7cfb\u5217\u7684\u89e3\u8bfb\uff1a\u8bba\u6587\u7cbe\u8bfb GPT1:&nbsp;Improving Lang &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/12\/29\/gp123\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">GPT\u7cfb\u5217\u8bba\u6587\uff1a\u751f\u6210\u5f0f\u9884\u8bad\u7ec3\u4e0e\u96f6\u6837\u672c\u5b66\u4e60<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21,4,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/11297"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=11297"}],"version-history":[{"count":44,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/11297\/revisions"}],"predecessor-version":[{"id":11426,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/11297\/revisions\/11426"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=11297"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=11297"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=11297"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}