{"id":34660,"date":"2024-12-10T16:57:02","date_gmt":"2024-12-10T08:57:02","guid":{"rendered":"https:\/\/17aitech.com\/?p=34660"},"modified":"2024-12-10T16:57:02","modified_gmt":"2024-12-10T08:57:02","slug":"%e7%9b%9b%e5%90%8d%e4%b8%80%e6%97%b6%e7%9a%84bert%e5%93%aa%e5%8e%bb%e4%ba%86%ef%bc%9f%e8%bf%99%e4%b8%aa%e9%97%ae%e9%a2%98%e7%9a%84%e7%ad%94%e6%a1%88%e6%98%ad%e7%a4%ba%e4%ba%86llm%e8%8c%83%e5%bc%8f","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34660","title":{"rendered":"\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-22-14\" target=\"_blank\">\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"47\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>\u7f16\u7801\u5668\u6a21\u578b\u54ea\u53bb\u4e86\uff1f\u5982\u679c BERT \u6548\u679c\u597d\uff0c\u90a3\u4e3a\u4ec0\u4e48\u4e0d\u6269\u5c55\u5b83\uff1f\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u6216\u4ec5\u7f16\u7801\u5668\u6a21\u578b\u600e\u4e48\u6837\u4e86\uff1f<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png\"><\/a><\/p>\n<p>\u5728\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLM\uff09\u9886\u57df\uff0c\u73b0\u5728\u662f\u4ec5\u89e3\u7801\u5668\u6a21\u578b\uff08\u5982 GPT \u7cfb\u5217\u6a21\u578b\uff09\u72ec\u9886\u98ce\u9a9a\u7684\u65f6\u4ee3\u3002\u90a3\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u6216\u4ec5\u7f16\u7801\u5668\u6a21\u578b\u53d1\u5c55\u5982\u4f55\u5462\uff1f\u4e3a\u4ec0\u4e48\u66fe\u7ecf\u76db\u540d\u4e00\u65f6\u7684 BERT \u5374\u6e10\u6e10\u5c11\u6709\u4eba\u5173\u6ce8\u4e86\uff1f<\/p>\n<p>\u8fd1\u65e5\uff0cAI \u521b\u4e1a\u516c\u53f8 Reka \u7684\u9996\u5e2d\u79d1\u5b66\u5bb6\u548c\u8054\u5408\u521b\u59cb\u4eba Yi Tay \u53d1\u5e03\u4e86\u4e00\u7bc7\u535a\u5ba2\u6587\u7ae0\uff0c\u5206\u4eab\u4e86\u4ed6\u7684\u770b\u6cd5\u3002Yi Tay \u5728\u53c2\u4e0e\u521b\u7acb Reka \u4e4b\u524d\u66fe\u5728 Google Research \u548c\u8c37\u6b4c\u5927\u8111\u5de5\u4f5c\u8fc7\u4e09\u5e74\u591a\u65f6\u95f4\uff0c\u53c2\u4e0e\u8fc7 PaLM\u3001UL2\u3001Flan-2\u3001Bard \u7b49\u8457\u540d LLM \u4ee5\u53ca PaLI-X \u548c ViT-22B \u7b49\u591a\u6a21\u6001\u6a21\u578b\u7684\u7814\u53d1\u5de5\u4f5c\u3002\u4ee5\u4e0b\u4e3a\u4ed6\u7684\u535a\u5ba2\u6587\u7ae0\u5185\u5bb9\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-fb46722b953fef55591ebfd39ed590fe.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-fb46722b953fef55591ebfd39ed590fe.png\"><\/a><\/p>\n<p><strong>\u57fa\u7840\u7b80\u4ecb<\/strong><\/p>\n<p>\u603b\u4f53\u4e0a\u770b\uff0c\u8fc7\u53bb\u8fd9\u4e9b\u5e74\u7684 LLM \u6a21\u578b\u67b6\u6784\u4e3b\u8981\u5206\u4e3a\u4e09\u5927\u8303\u5f0f\uff1a\u4ec5\u7f16\u7801\u5668\u6a21\u578b\uff08\u5982 BERT\uff09\u3001\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u6a21\u578b\uff08\u5982 T5\uff09\u3001\u4ec5\u89e3\u7801\u5668\u6a21\u578b\uff08\u5982 GPT \u7cfb\u5217\u6a21\u578b\uff09\u3002\u4eba\u4eec\u5e38\u5e38\u641e\u4e0d\u6e05\u695a\u8fd9\u4e9b\uff0c\u5e76\u4e14\u5bf9\u8fd9\u4e9b\u5206\u7c7b\u65b9\u6cd5\u548c\u67b6\u6784\u6709\u6240\u8bef\u89e3\u3002<\/p>\n<p>\u9996\u5148\u8981\u7406\u89e3\u7684\u4e00\u70b9\u662f\uff1a\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u6a21\u578b\u5b9e\u9645\u4e0a\u4e5f\u662f<mark data-type=\"concepts\" data-id=\"e70f77cf-89e6-4a0b-b689-c5757a1ecf6c\">\u81ea\u56de\u5f52\u6a21\u578b<\/mark>\u3002\u5728\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u6a21\u578b\u4e2d\uff0c\u89e3\u7801\u5668\u672c\u8d28\u4e0a\u4f9d\u7136\u662f\u56e0\u679c\u89e3\u7801\u5668\u3002\u5176\u65e0\u9700\u9884\u586b\u5145\u89e3\u7801\u5668\u6a21\u578b\uff0c\u800c\u662f\u4f1a\u5c06\u67d0\u4e9b\u6587\u672c\u5378\u8f7d\u5230\u7f16\u7801\u5668\uff0c\u7136\u540e\u518d\u901a\u8fc7\u4ea4\u53c9\u6ce8\u610f\u529b\u53d1\u9001\u7ed9\u89e3\u7801\u5668\u3002\u662f\u7684\uff0cT5 \u6a21\u578b\u4e5f\u662f<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff01<\/p>\n<p>\u8fd9\u7c7b\u6a21\u578b\u7684\u4e00\u79cd\u53d8\u4f53\u662f\u524d\u7f00<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08Prefix Language Model\uff09\uff0c\u7b80\u79f0 PrefixLM\uff0c\u5176\u5de5\u4f5c\u65b9\u5f0f\u51e0\u4e4e\u4e00\u6837\uff0c\u53ea\u662f\u6ca1\u6709\u4ea4\u53c9\u6ce8\u610f\u529b\uff08\u4ee5\u53ca\u5176\u5b83\u4e00\u4e9b\u5c0f\u7ec6\u8282\uff0c\u6bd4\u5982\u7f16\u7801\u5668 \/ \u89e3\u7801\u5668\u4e4b\u95f4\u5171\u4eab<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\u4ee5\u53ca\u6ca1\u6709\u7f16\u7801\u5668\u74f6\u9888\uff09\u3002PrefixLM \u6709\u65f6\u4e5f\u88ab\u79f0\u4e3a\u975e\u56e0\u679c\u89e3\u7801\u5668\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u3001\u4ec5\u89e3\u7801\u5668\u6a21\u578b\u548c PrefixLM \u6574\u4f53\u4e0a\u5dee\u522b\u4e0d\u5927\uff01<\/p>\n<p>\u5728 Hyung Won \u8fd1\u671f\u7684\u7cbe\u5f69\u8bb2\u5ea7\u4e2d\uff0c\u4ed6\u5a34\u719f\u5730\u89e3\u91ca\u4e86\u8fd9\u4e9b\u6a21\u578b\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u8be6\u60c5\u53ef\u53c2\u9605<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u7684\u62a5\u9053\uff1a\u300a<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650922111&amp;idx=2&amp;sn=69ae67a08b5ffd782d7bd25d94e6ed7a&amp;chksm=84e41801b393911720f779edaa704cc703a4871b96dd75eff9b599470090e8eb5d50034cebe0&amp;scene=21#wechat_redirect\" target=\"_blank\">AI \u7814\u7a76\u7684\u4e3b\u8981\u63a8\u52a8\u529b\u4f1a\u662f\u4ec0\u4e48\uff1fChatGPT \u56e2\u961f\u7814\u7a76\u79d1\u5b66\u5bb6\uff1a\u7b97\u529b\u6210\u672c\u4e0b\u964d<\/a>\u300b<\/p>\n<p>\u540c\u65f6\uff0cBERT \u8fd9\u6837\u7684\u4ec5\u7f16\u7801\u5668\u6a21\u578b\u7684\u53bb\u566a\u65b9\u5f0f\u4e0d\u4e00\u6837\uff08\u5373 in-place\uff09\uff1b\u5e76\u4e14\u4ece\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u8bb2\uff0c\u4ec5\u7f16\u7801\u5668\u6a21\u578b\u8981\u5728\u9884\u8bad\u7ec3\u4e4b\u540e\u771f\u6b63\u53d1\u6325\u4f5c\u7528\uff0c\u9700\u8981\u4f9d\u9760\u5206\u7c7b\u300c\u4efb\u52a1\u300d\u5934\u3002\u540e\u6765\uff0cT5 \u7b49\u6a21\u578b\u91c7\u7528\u4e86\u4e00\u79cd\u300c\u4fee\u6539\u7248\u300d\u7684\u53bb\u566a\u76ee\u6807\uff0c\u5176\u4f7f\u7528\u4e86\u4e00\u79cd<mark data-type=\"tech_methods\" data-id=\"d5bb7d14-efba-43e8-80a1-ddd7dfd88677\">\u5e8f\u5217\u5230\u5e8f\u5217<\/mark>\u7684\u683c\u5f0f\u3002<\/p>\n<p>\u4e3a\u6b64\uff0c\u9700\u8981\u6307\u51fa\uff1aT5 \u4e2d\u7684\u53bb\u566a\u5e76\u975e\u4e00\u79cd\u65b0\u7684<mark data-type=\"concepts\" data-id=\"a3ee878c-7013-408b-8e6b-7b1038057f8f\">\u76ee\u6807\u51fd\u6570<\/mark>\uff08\u5728<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u610f\u4e49\u4e0a\uff09\uff0c\u800c\u662f\u4e00\u79cd\u8de8\u8f93\u5165\u7684\u6570\u636e\u53d8\u6362\uff0c\u5373\u4f60\u4e5f\u53ef\u4ee5\u4f7f\u7528\u4e00\u4e2a\u56e0\u679c\u89e3\u7801\u5668\u8bad\u7ec3\u8de8\u5ea6\u635f\u574f\u76ee\u6807\uff08span corruption objective \uff09\u3002<\/p>\n<p>\u4eba\u4eec\u603b\u662f\u5047\u8bbe\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u6a21\u578b\u5fc5\u5b9a\u662f\u53bb\u566a\u6a21\u578b\uff0c\u90e8\u5206\u539f\u56e0\u662f T5 \u5b9e\u5728\u8fc7\u4e8e\u5177\u6709\u4ee3\u8868\u6027\u3002\u4f46\u4e8b\u5b9e\u5e76\u4e0d\u603b\u662f\u5982\u6b64\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528\u5e38\u89c4\u7684\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\uff08\u6bd4\u5982\u56e0\u679c\u8bed\u8a00\u5efa\u6a21\uff09\u8bad\u7ec3\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u3002\u53cd\u8fc7\u6765\uff0c\u4e5f\u53ef\u4ee5\u4f7f\u7528\u8de8\u5ea6\u635f\u574f\u4efb\u52a1\u8bad\u7ec3\u56e0\u679c\u89e3\u7801\u5668\u3002\u6b63\u5982\u6211\u524d\u9762\u8bf4\u7684\u90a3\u6837\uff0c\u8fd9\u57fa\u672c\u4e0a\u5c31\u662f\u4e00\u79cd\u6570\u636e\u53d8\u6362\u3002<\/p>\n<p>\u8fd8\u6709\u4e00\u70b9\u503c\u5f97\u6ce8\u610f\uff1a\u4e00\u822c\u6765\u8bf4\uff0c\u6709 2N \u4e2a<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u7684\u8ba1\u7b97\u6210\u672c\u4e0e\u6709 N \u4e2a<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684\u4ec5\u89e3\u7801\u5668\u6a21\u578b\u4e00\u6837\uff0c\u8fd9\u6837\u4e00\u6765\uff0c\u5b83\u4eec\u7684 FLOP \u548c<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u91cf\u4e4b\u6bd4\u5c31\u4e0d\u4e00\u6837\u4e86\u3002\u8fd9\u5c31\u50cf\u662f\u5728\u8f93\u5165\u548c\u76ee\u6807\u4e4b\u95f4\u5206\u914d\u300c\u6a21\u578b\u7a00\u758f\u6027\u300d\u3002<\/p>\n<p>\u8fd9\u4e0d\u662f\u4ec0\u4e48\u65b0\u4e1c\u897f\uff0c\u4e5f\u4e0d\u662f\u6211\u81ea\u5df1\u60f3\u51fa\u6765\u7684\u30022019 \u5e74\u7684 T5 \u8bba\u6587\u4e2d\u5c31\u6709\uff0c\u5e76\u4e14 UL2 \u8bba\u6587\u4e5f\u518d\u6b21\u5f3a\u8c03\u4e86\u8fd9\u4e00\u70b9\u3002<\/p>\n<p>\u76ee\u524d\u6765\u8bf4\uff0c\u5f88\u9ad8\u5174\u80fd\u628a\u8fd9\u4e00\u70b9\u8bf4\u6e05\u695a\u3002\u73b0\u5728\u6765\u8bf4\u76ee\u6807\u3002<\/p>\n<p><strong>\u5173\u4e8e\u53bb\u566a\u76ee\u6807\uff08\u5b83\u6ca1\u8d77\u4f5c\u7528\u5417\uff1f\u65e0\u6cd5\u6269\u5c55\u5417\uff1f\u8fd8\u662f\u592a\u5bb9\u6613\u4e86\uff1f\uff09<\/strong><\/p>\n<p>\u8fd9\u91cc\u7684\u53bb\u566a\u76ee\u6807\u662f\u6307\u300c\u8de8\u5ea6\u635f\u574f\u300d\u4efb\u52a1\u7684\u4efb\u610f\u53d8\u4f53\u3002\u8fd9\u6709\u65f6\u5019\u88ab\u79f0\u4e3a\u300c\u586b\u5145\u300d\u6216\u300c\u586b\u7a7a\u300d\u3002\u8868\u8fbe\u5b83\u7684\u65b9\u5f0f\u6709\u5f88\u591a\uff0c\u6bd4\u5982\u8de8\u5ea6\u957f\u5ea6\u3001\u968f\u673a\u6027\u3001sentinel token \u7b49\u3002\u60f3\u5fc5\u4f60\u5df2\u660e\u767d\u5176\u4e2d\u5173\u952e\u3002<\/p>\n<p>\u5c3d\u7ba1 BERT \u5f0f\u6a21\u578b\u7684\u53bb\u566a\u76ee\u6807\u57fa\u672c\u662f\u5c31\u5730\u7684\uff08in-place\uff0c\u6bd4\u5982\u5206\u7c7b\u5934\u4f4d\u4e8e\u63a9\u7801 token \u4e0a\uff09\uff0c\u4f46\u300cT5 \u98ce\u683c\u300d\u8981\u66f4\u73b0\u4ee3\u4e00\u70b9\uff0c\u5373\u901a\u8fc7\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u6216\u4ec5\u89e3\u7801\u5668\u6a21\u578b\u6765\u5904\u7406\u6570\u636e\u53d8\u6362\u3002\u5728\u8fd9\u6837\u7684\u6570\u636e\u53d8\u6362\u4e2d\uff0c\u88ab\u63a9\u853d\u7684 token \u53ea\u662f\u4f1a\u88ab\u300c\u79fb\u56de\u53bb\u300d\u4ee5\u4fbf\u6a21\u578b\u7ed9\u51fa\u9884\u6d4b\u3002<\/p>\n<p>\u9884\u8bad\u7ec3\u7684\u4e3b\u8981\u76ee\u6807\u662f\u4ee5\u5c3d\u53ef\u80fd\u6700\u9ad8\u6548\u548c\u6709\u6548\u7684\u65b9\u5f0f\u6784\u5efa\u4e0e\u4e0b\u6e38\u4efb\u52a1\u5bf9\u9f50\u7684\u5185\u90e8\u8868\u5f81\u3002\u8fd9\u79cd\u5185\u90e8\u8868\u5f81\u8d8a\u597d\uff0c\u5c31\u66f4\u5bb9\u6613\u5c06\u8fd9\u4e9b\u5b66\u4e60\u5230\u7684\u8868\u5f81\u7528\u4e8e\u540e\u7eed\u4efb\u52a1\u3002\u6211\u4eec\u90fd\u77e5\u9053\uff0c\u7b80\u5355\u7684\u4e0b\u4e00\u8bcd\u9884\u6d4b\u300c\u56e0\u679c\u8bed\u8a00\u5efa\u6a21\u300d\u76ee\u6807\u8868\u73b0\u51fa\u8272\uff0c\u5e76\u4e14\u5df2\u6210\u4e3a LLM \u9769\u547d\u7684\u6838\u5fc3\u3002\u73b0\u5728\u7684\u95ee\u9898\u662f\u53bb\u566a\u76ee\u6807\u662f\u5426\u540c\u6837\u51fa\u8272\u3002<\/p>\n<p>\u6839\u636e\u516c\u5f00\u4fe1\u606f\uff0c\u6211\u4eec\u77e5\u9053 T5-11B \u7684\u6548\u679c\u76f8\u5f53\u597d\uff0c\u5373\u4f7f\u5728\u5bf9\u9f50\u548c\u7ecf\u8fc7\u76d1\u7763\u5f0f\u5fae\u8c03\u4e4b\u540e\uff08Flan-T5 XXL \u7684 MMLU \u5206\u6570\u662f 55+\uff0c\u5728\u5f53\u65f6\u6765\u8bf4\uff0c\u8fd9\u4e2a\u89c4\u6a21\u7684\u6a21\u578b\u5df2\u7ecf\u76f8\u5f53\u597d\u4e86\uff09\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u53ef\u4ee5\u5f97\u51fa\u8fd9\u6837\u7684\u7ed3\u8bba\uff1a\u53bb\u566a\u76ee\u6807\u7684\u8fc1\u79fb\u8fc7\u7a0b\uff08\u9884\u8bad\u7ec3\u2192\u5bf9\u9f50\uff09\u5728\u8fd9\u4e2a\u89c4\u6a21\u4e0a\u76f8\u5bf9\u6765\u8bf4\u6548\u679c\u4e0d\u9519\u3002<\/p>\n<p>\u6211\u7684\u770b\u6cd5\u662f\uff0c\u53bb\u566a\u76ee\u6807\u7684\u6548\u679c\u5f88\u597d\uff0c\u4f46\u8fd8\u4e0d\u8db3\u4ee5\u5355\u72ec\u4f5c\u4e3a\u76ee\u6807\u3002\u4e00\u4e2a\u5de8\u5927\u7684\u7f3a\u70b9\u6e90\u81ea\u6240\u8c13\u7684\u66f4\u5c11\u7684\u300c\u635f\u5931\u66b4\u9732\uff08loss exposure\uff09\u300d\u3002\u5728\u53bb\u566a\u76ee\u6807\u4e2d\uff0c\u4ec5\u6709\u5c11\u91cf token \u4f1a\u88ab\u63a9\u853d\u548c\u5f97\u5230\u5b66\u4e60\uff08\u5373\u88ab\u8003\u8651\u5230\u635f\u5931\u4e2d\uff09\u3002\u53cd\u8fc7\u6765\uff0c\u5728\u5e38\u89c4\u7684\u8bed\u8a00\u5efa\u6a21\u4e2d\uff0c\u8fd9\u63a5\u8fd1\u4e8e 100%\u3002\u8fd9\u4f7f\u5f97\u6bcf\u4e2a FLOP \u7684\u6837\u672c\u6548\u7387\u975e\u5e38\u4f4e\uff0c\u8fd9\u4f7f\u5f97\u5728 flop \u57fa\u7840\u4e0a\u7684\u6bd4\u8f83\u4e2d\uff0c\u53bb\u566a\u76ee\u6807\u7684\u52a3\u52bf\u5f88\u5927\u3002<\/p>\n<p>\u53bb\u566a\u76ee\u6807\u7684\u53e6\u4e00\u4e2a\u7f3a\u70b9\u662f\u5176\u6bd4\u5e38\u89c4\u8bed\u8a00\u5efa\u6a21\u66f4\u4e0d\u81ea\u7136\uff0c\u56e0\u4e3a\u5b83\u4f1a\u4ee5\u4e00\u79cd\u5947\u602a\u7684\u65b9\u5f0f\u91cd\u65b0\u8bbe\u5b9a\u8f93\u5165 \/ \u8f93\u51fa\u7684\u683c\u5f0f\uff0c\u8fd9\u4f7f\u5f97\u5b83\u4eec\u4e0d\u592a\u9002\u5408\u5c11\u6837\u672c\u5b66\u4e60\u3002\uff08\u4f46\u5728\u5c11\u6837\u672c\u4efb\u52a1\u4e0a\uff0c\u4ecd\u53ef\u80fd\u901a\u8fc7\u8c03\u6574\u8fd9\u4e9b\u6a21\u578b\u8ba9\u5176\u8868\u73b0\u5f97\u76f8\u5f53\u4f18\u826f\u3002\uff09\u56e0\u6b64\uff0c\u6211\u8ba4\u4e3a\u53bb\u566a\u76ee\u6807\u5e94\u8be5\u53ea\u80fd\u7528\u4f5c\u5e38\u89c4\u8bed\u8a00\u5efa\u6a21\u7684\u8865\u5145\u76ee\u6807\u3002<\/p>\n<p><strong>\u7edf\u4e00\u7684\u65e9\u671f\u4ee5\u53ca BERT \u7c7b\u6a21\u578b\u6d88\u5931\u7684\u539f\u56e0<\/strong><\/p>\n<p>\u7c7b\u4f3c BERT \u7684\u6a21\u578b\u9010\u6e10\u6d88\u5931\uff0c\u73b0\u5728\u5df2\u7ecf\u6ca1\u591a\u5c11\u4eba\u518d\u8c08\u5b83\u4eec\u4e86\u3002\u8fd9\u4e5f\u80fd\u89e3\u91ca\u4e3a\u4ec0\u4e48\u6211\u4eec\u73b0\u5728\u770b\u4e0d\u5230\u8d85\u5927\u89c4\u6a21\u7684 BERT \u6a21\u578b\u4e86\u3002\u539f\u56e0\u662f\u4ec0\u4e48\uff1f\u8fd9\u5f88\u5927\u7a0b\u5ea6\u4e0a\u662f\u56e0\u4e3a\u4efb\u52a1 \/ \u5efa\u6a21\u8303\u5f0f\u7684\u7edf\u4e00\u548c\u8f6c\u53d8\u3002BERT \u5f0f\u6a21\u578b\u5f88\u7e41\u7410\uff0c\u4f46 BERT \u6a21\u578b\u88ab\u5f03\u7528\u7684\u771f\u6b63\u539f\u56e0\u662f\uff1a\u4eba\u4eec\u5e0c\u671b\u4e00\u6b21\u6027\u5b8c\u6210\u6240\u6709\u4efb\u52a1\uff0c\u56e0\u6b64\u91c7\u7528\u4e86\u4e00\u79cd\u66f4\u597d\u7684\u53bb\u566a\u65b9\u6cd5 \u2014\u2014 \u4f7f\u7528<mark data-type=\"concepts\" data-id=\"e70f77cf-89e6-4a0b-b689-c5757a1ecf6c\">\u81ea\u56de\u5f52\u6a21\u578b<\/mark>\u3002<\/p>\n<p>\u5728 2018-2021 \u5e74\u671f\u95f4\uff0c\u51fa\u73b0\u4e86\u4e00\u79cd\u9690\u542b\u7684\u8303\u5f0f\u8f6c\u53d8\uff1a\u4ece\u5355\u4efb\u52a1\u5fae\u8c03\u8f6c\u53d8\u6210\u5927\u89c4\u6a21\u591a\u4efb\u52a1\u6a21\u578b\u3002\u8fd9\u6162\u6162\u5730\u5c06\u6211\u4eec\u5bfc\u5411\u4e86\u7edf\u4e00\u7684 SFT \u6a21\u578b\uff0c\u8fd9\u5c31\u662f\u6211\u4eec\u5982\u4eca\u770b\u5230\u7684\u901a\u7528\u6a21\u578b\u3002\u4f7f\u7528 BERT \u5374\u5f88\u96be\u505a\u5230\u8fd9\u4e00\u70b9\u3002\u6211\u8ba4\u4e3a\u8fd9\u4e0e\u300c\u53bb\u566a\u300d\u5173\u7cfb\u4e0d\u5927\u3002\u5bf9\u4e8e\u8fd8\u60f3\u4f7f\u7528\u8fd9\u6837\u7684\u6a21\u578b\uff08\u5373 T5\uff09\u7684\u4eba\uff0c\u4ed6\u4eec\u627e\u5230\u4e86\u4e00\u79cd\u91cd\u65b0\u8868\u8fbe\u53bb\u566a\u9884\u8bad\u7ec3\u4efb\u52a1\u7684\u65b9\u6cd5\uff0c\u8fd9\u4f7f\u5f97\u5982\u4eca BERT \u5f0f\u6a21\u578b\u5df2\u7ecf\u57fa\u672c\u4e0a\u88ab\u5f03\u7528\u4e86\uff0c\u56e0\u4e3a\u6211\u4eec\u5df2\u6709\u66f4\u597d\u7684\u66ff\u4ee3\u65b9\u6848\u3002<\/p>\n<p>\u66f4\u786e\u5207\u5730\u8bf4\uff0c\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u548c\u4ec5\u89e3\u7801\u5668\u6a21\u578b\u65e0\u9700\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u5206\u7c7b\u5934\u5c31\u80fd\u7528\u4e8e\u591a\u79cd\u4efb\u52a1\u3002\u5bf9\u4e8e\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\uff0c\u7814\u7a76\u8005\u548c\u5de5\u7a0b\u5e08\u5f00\u59cb\u53d1\u73b0\u653e\u5f03\u7f16\u7801\u5668\u7684\u6548\u679c\u5c31\u548c BERT \u7f16\u7801\u5668\u5dee\u4e0d\u591a\u3002\u6b64\u5916\uff0c\u8fd9\u8fd8\u80fd\u4fdd\u7559\u53cc\u5411\u6ce8\u610f\u529b\u7684\u4f18\u52bf \u2014\u2014 \u8be5\u4f18\u52bf\u8ba9 BERT \u5728\u5c0f\u89c4\u6a21\uff08\u5f80\u5f80\u662f\u751f\u4ea7\u89c4\u6a21\uff09\u4e0a\u53ef\u4e0e GPT \u7ade\u4e89\u3002<\/p>\n<p><strong>\u53bb\u566a\u76ee\u6807\u7684\u4ef7\u503c<\/strong><\/p>\n<p>\u53bb\u566a\u9884\u8bad\u7ec3\u76ee\u6807\u4e5f\u80fd\u4ee5\u4e00\u79cd\u7c7b\u4f3c\u5e38\u89c4\u8bed\u8a00\u5efa\u6a21\u7684\u65b9\u5f0f\u5b66\u4e60\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u3002\u4f46\u662f\uff0c\u4e0d\u540c\u4e8e\u5e38\u89c4\u56e0\u679c\u8bed\u8a00\u5efa\u6a21\uff0c\u8fd9\u9700\u8981\u5bf9\u5e8f\u5217\u4f7f\u7528\u4e00\u79cd\u6570\u636e\u53d8\u6362\uff0c\u4f7f\u5f97\u6a21\u578b\u53ef\u4ee5\u5b66\u4e60\u300c\u586b\u7a7a\u300d\uff0c\u800c\u4e0d\u662f\u7b80\u5355\u5730\u9884\u6d4b\u4ece\u5de6\u5230\u53f3\u7684\u81ea\u7136\u6587\u672c\u3002<\/p>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u53bb\u566a\u76ee\u6807\u6709\u65f6\u4e5f\u79f0\u4e3a\u300c\u586b\u5145\u4efb\u52a1\u300d\uff0c\u6709\u65f6\u4f1a\u4e0e\u5e38\u89c4\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\u4e00\u8d77\u5728\u9884\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6df7\u5408\u4f7f\u7528\u3002<\/p>\n<p>\u867d\u7136\u786e\u5207\u7684\u914d\u7f6e\u548c\u5b9e\u73b0\u7ec6\u8282\u53ef\u80fd\u6709\u6240\u4e0d\u540c\uff0c\u4f46\u5f53\u4eca\u7684\u73b0\u4ee3 LLM \u53ef\u80fd\u5728\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u7ec4\u5408\u4f7f\u7528\u8bed\u8a00\u5efa\u6a21\u548c\u586b\u5145\u3002\u6709\u8da3\u7684\u662f\uff0c\u8fd9\u79cd\u300c<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark> + \u586b\u5145\u300d\u7684\u6df7\u5408\u5b9e\u9645\u4e0a\u4e5f\u5927\u6982\u5728\u540c\u4e00\u65f6\u671f\u56db\u5904\u4f20\u64ad\uff08\u5982 UL2\u3001FIM\u3001GLM\u3001CM3\uff09\uff0c\u8bb8\u591a\u56e2\u961f\u90fd\u5e26\u6765\u4e86\u81ea\u5df1\u72ec\u5177\u7279\u8272\u7684\u6df7\u5408\u65b9\u6848\u3002\u987a\u5e26\u4e00\u63d0\uff0c\u76ee\u524d\u5df2\u77e5\u7684\u4ee5\u8fd9\u79cd\u65b9\u5f0f\u8bad\u7ec3\u7684\u6700\u5927\u6a21\u578b\u5f88\u53ef\u80fd\u662f PaLM-2\u3002<\/p>\n<p>\u8fd8\u9700\u8981\u8bf4\u660e\u4e00\u70b9\uff0c\u9884\u8bad\u7ec3\u4efb\u52a1\u6df7\u5408\u4e5f\u53ef\u4ee5\u6309\u987a\u5e8f<mark data-type=\"tech_methods\" data-id=\"efb128f3-0b7c-45a8-a6ae-065fa93c6556\">\u5806\u53e0<\/mark>\uff0c\u4e0d\u4e00\u5b9a\u5fc5\u987b\u540c\u65f6\u6df7\u5408\uff0c\u6bd4\u5982 Flan-T5 \u8d77\u521d\u662f\u5728 1T \u8de8\u5ea6\u635f\u574f token \u4e0a\u8bad\u7ec3\u7684\uff0c\u7136\u540e\u6362\u5230\u524d\u9988\u8bed\u8a00\u5efa\u6a21\u76ee\u6807\u7684 100B token\uff0c\u4e4b\u540e\u518d\u8fdb\u884c flan \u6307\u4ee4\u5fae\u8c03\u3002\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u8bb2\uff0c\u8fd9\u9002\u5408\u6df7\u5408\u53bb\u566a \/ LM \u76ee\u6807\u6a21\u578b\u3002\u9700\u8981\u660e\u786e\u7684\u662f\uff0c\u524d\u7f00\u8bed\u8a00\u5efa\u6a21\u76ee\u6807\uff08\u4e0d\u8981\u4e0e\u67b6\u6784\u6df7\u6dc6\uff09\u5355\u7eaf\u53ea\u662f\u56e0\u679c\u8bed\u8a00\u5efa\u6a21\uff0c\u5176\u6709\u4e00\u4e2a\u968f\u673a\u786e\u5b9a\u548c\u53d1\u9001\u5230\u8f93\u5165\u7aef\u7684\u5206\u5272\u70b9\uff08\u6ca1\u6709\u635f\u5931\u548c\u975e\u56e0\u679c\u63a9\u7801\uff09\u3002<\/p>\n<p>\u987a\u4fbf\u4e00\u63d0\uff0c\u586b\u5145\u53ef\u80fd\u8d77\u6e90\u4e8e\u4ee3\u7801 LLM \u9886\u57df\uff0c\u5176\u4e2d\u300c\u586b\u7a7a\u300d\u66f4\u50cf\u662f\u5199\u4ee3\u7801\u6240\u9700\u7684\u4e00\u4e2a\u529f\u80fd\u3002\u540c\u65f6\uff0cUL2 \u7684\u52a8\u673a\u66f4\u591a\u662f\u5c06\u53bb\u566a\u76ee\u6807\u548c\u53cc\u5411 LLM \u64c5\u957f\u7684\u4efb\u52a1\u7c7b\u522b\u4e0e\u56fa\u6709\u7684\u751f\u6210\u4efb\u52a1\uff08\u4f8b\u5982\u603b\u7ed3\u6216\u5f00\u653e\u5f0f\u751f\u6210\uff09\u7edf\u4e00\u8d77\u6765\u3002\u8fd9\u79cd\u81ea\u56de\u5f52\u5f0f\u7684\u89e3\u7801\u300c\u5411\u540e\u79fb\u300d\u7684\u4f18\u70b9\u662f\uff1a\u5176\u4e0d\u4ec5\u80fd\u8ba9\u6a21\u578b\u5b66\u4e60\u66f4\u957f\u7a0b\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u8fd8\u80fd\u8ba9\u5176\u9690\u5f0f\u5730\u53d7\u76ca\u4e8e\u975e\u663e\u5f0f\u7684\u53cc\u5411\u6ce8\u610f\u529b\uff08\u56e0\u4e3a\u4e3a\u4e86\u586b\u7a7a\uff0c\u4f60\u5df2\u7ecf\u770b\u8fc7\u4e86\u672a\u6765\uff09\u3002<\/p>\n<p>\u6709\u4e00\u4e2a\u4f20\u8bf4\u4e2d\u7684\u7ecf\u9a8c\uff1a\u53bb\u566a\u76ee\u6807\u5b66\u4e60\u7684\u8868\u5f81\u5728\u7279\u5b9a\u4efb\u52a1\u7c7b\u522b\u4e0a\u8868\u73b0\u66f4\u597d\uff0c\u6709\u65f6\u5019\u8fd8\u4f1a\u6709\u66f4\u9ad8\u7684\u6837\u672c\u6548\u7387\u3002\u5728 U-PaLM \u8bba\u6587\u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u5c11\u91cf\u7684\u8de8\u5ea6\u635f\u574f up-training \u5982\u4f55\u6539\u53d8\u5728\u4e00\u7ec4 BIG-Bench \u4efb\u52a1\u4e0a\u7684\u884c\u4e3a\u548c\u6d8c\u73b0\u73b0\u8c61\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u5bf9\u4f7f\u7528\u8fd9\u4e00\u76ee\u6807\u8bad\u7ec3\u7684\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u901a\u5e38\u53ef\u4ee5\u5f97\u5230\u66f4\u597d\u7684\u76d1\u7763\u5f0f\u5fae\u8c03\u6a21\u578b\uff0c\u5c24\u5176\u662f\u5f53\u89c4\u6a21\u8f83\u5c0f\u65f6\u3002<\/p>\n<p>\u5728\u5355\u4efb\u52a1\u5fae\u8c03\u65b9\u9762\uff0c\u53ef\u4ee5\u770b\u5230 PaLM-1 62B \u6a21\u578b\u88ab\u5c0f\u5f97\u591a\u7684 T5 \u6a21\u578b\u51fb\u8d25\u3002\u5728\u76f8\u5bf9\u8f83\u5c0f\u7684\u89c4\u6a21\u4e0a\uff0c\u300c\u53cc\u5411\u6ce8\u610f\u529b + \u53bb\u566a\u76ee\u6807\u300d\u662f\u4e00\u8bb0\u6f02\u4eae\u7684\u7ec4\u5408\u62f3\uff01\u6211\u76f8\u4fe1\u5f88\u591a\u5b9e\u8df5\u8005\u4e5f\u6ce8\u610f\u5230\u4e86\u8fd9\u79cd\u60c5\u51b5\uff0c\u5c24\u5176\u662f\u5728\u751f\u4ea7\u5e94\u7528\u4e2d\u3002<\/p>\n<p><strong>\u53cc\u5411\u6ce8\u610f\u529b\u5982\u4f55\u5462\uff1f<\/strong><\/p>\n<p>\u5bf9<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u6765\u8bf4\uff0c\u53cc\u5411\u6ce8\u610f\u529b\u662f\u4e00\u79cd\u6709\u8da3\u7684\u300c\u5f52\u7eb3\u504f\u7f6e\u300d\u2014\u2014 \u4eba\u4eec\u5e38\u5e38\u5c06\u5176\u4e0e\u76ee\u6807\u548c\u6a21\u578b\u9aa8\u5e72\u6df7\u6dc6\u3002\u5728\u4e0d\u540c\u7684\u8ba1\u7b97\u9886\u57df\uff0c\u5f52\u7eb3\u504f\u7f6e\u7684\u7528\u9014\u4e5f\u5404\u4e0d\u76f8\u540c\uff0c\u5e76\u4e14\u4e5f\u53ef\u80fd\u5bf9\u6269\u5c55\u66f2\u7ebf\u9020\u6210\u4e0d\u540c\u7684\u5f71\u54cd\u3002\u8bdd\u867d\u5982\u6b64\uff0c\u76f8\u6bd4\u4e8e\u8f83\u5c0f\u89c4\u6a21\uff0c\u53cc\u5411\u6ce8\u610f\u529b\u5728\u89c4\u6a21\u8f83\u5927\u65f6\u53ef\u80fd\u5c31\u6ca1\u90a3\u4e48\u91cd\u8981\u4e86\uff0c\u6216\u8005\u53ef\u80fd\u5bf9\u4e0d\u540c\u7684\u4efb\u52a1\u6216\u6a21\u6001\u6709\u4e0d\u540c\u7684\u5f71\u54cd\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0cPaliGemma \u4f7f\u7528\u4e86 PrefixLM \u67b6\u6784\u3002<\/p>\n<p>Hyung Won \u4e5f\u5728\u4ed6\u7684\u6f14\u8bb2\u4e2d\u6307\u51fa\uff1aPrefixLM \u6a21\u578b\uff08\u4f7f\u7528\u53cc\u5411\u6ce8\u610f\u529b\u7684\u4ec5\u89e3\u7801\u5668\u6a21\u578b\uff09\u4e5f\u5b58\u5728\u7f13\u5b58\u95ee\u9898\uff0c\u8fd9\u662f\u8fd9\u7c7b\u67b6\u6784\u7684\u4e00\u4e2a\u56fa\u6709\u7f3a\u9677\u3002\u4f46\u662f\uff0c\u6211\u8ba4\u4e3a\u6709\u5f88\u591a\u65b9\u6cd5\u53ef\u4ee5\u89e3\u51b3\u8fd9\u4e2a\u7f3a\u9677\uff0c\u4f46\u8fd9\u8d85\u51fa\u4e86\u672c\u6587\u7684\u8303\u56f4\u3002<\/p>\n<p><strong>\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u67b6\u6784\u7684\u4f18\u7f3a\u70b9<\/strong><\/p>\n<p>\u76f8\u6bd4\u4e8e\u4ec5\u89e3\u7801\u5668\u6a21\u578b\uff0c\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u67b6\u6784\u6709\u4f18\u52bf\u4e5f\u6709\u52a3\u52bf\u3002\u7b2c\u4e00\u79cd\u60c5\u51b5\u662f\u7f16\u7801\u5668\u7aef\u4e0d\u53d7\u56e0\u679c\u63a9\u7801\u7684\u9650\u5236\u3002\u5728\u67d0\u79cd\u7a0b\u5ea6\u4e0a\uff0c\u4f60\u53ef\u4ee5\u5728\u6ce8\u610f\u529b\u5c42\u4e0a\u653e\u5f00\u624b\u811a\uff0c\u6fc0\u8fdb\u5730\u6267\u884c<mark data-type=\"concepts\" data-id=\"0a4cedf0-0ee0-4406-946e-2877950da91d\">\u6c60\u5316<\/mark>\u6216\u4efb\u4f55\u5f62\u5f0f\u7684\u7ebf\u6027\u6ce8\u610f\u529b\uff0c\u800c\u4e0d\u5fc5\u62c5\u5fc3\u81ea\u56de\u5f52\u7684\u8bbe\u8ba1\u9650\u5236\u3002\u8fd9\u662f\u4e00\u79cd\u5c06\u4e0d\u592a\u91cd\u8981\u7684\u300c\u4e0a\u4e0b\u6587\u300d\u5378\u8f7d\u5230\u7f16\u7801\u5668\u7684\u597d\u65b9\u6cd5\u3002\u4f60\u4e5f\u80fd\u628a\u7f16\u7801\u5668\u505a\u5c0f\u70b9\uff0c\u8fd9\u4e5f\u662f\u4e2a\u4f18\u52bf\u3002<\/p>\n<p>\u5fc5\u9700\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u67b6\u6784\u7684\u4e00\u4e2a\u4f8b\u5b50\u662f Charformer\uff0c\u5176\u4e2d\u5927\u80c6\u4f7f\u7528\u4e86\u7f16\u7801\u5668\u5e76\u7f13\u89e3\u4e86\u5b57\u8282\u5c42\u9762\u7684\u6a21\u578b\u7684\u901f\u5ea6\u52a3\u52bf\u3002\u5728\u7f16\u7801\u5668\u65b9\u9762\u8fdb\u884c\u521b\u65b0\u53ef\u4ee5\u5feb\u901f\u83b7\u76ca\uff0c\u540c\u65f6\u65e0\u9700\u62c5\u5fc3\u56e0\u679c\u63a9\u7801\u7684\u91cd\u5927\u7f3a\u9677\u3002<\/p>\n<p>\u540c\u65f6\uff0c\u76f8\u6bd4\u4e8e PrefixLM\uff0c\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u7684\u4e00\u4e2a\u7f3a\u70b9\u662f\u8f93\u5165\u548c\u76ee\u6807\u5fc5\u987b\u5206\u914d\u56fa\u5b9a\u7684\u9884\u7b97\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5982\u679c\u8f93\u5165\u9884\u7b97\u662f 1024 token\uff0c\u90a3\u4e48\u7f16\u7801\u5668\u7aef\u5c31\u5fc5\u987b\u586b\u5145\u5230\u8fd9\u4e2a\u503c\uff0c\u800c\u8fd9\u53ef\u80fd\u4f1a\u6d6a\u8d39\u5927\u91cf\u8ba1\u7b97\u3002\u76f8\u53cd\uff0c\u5728 PrefixLM \u4e2d\uff0c\u8f93\u5165\u548c\u76ee\u6807\u53ef\u4ee5\u76f4\u63a5\u8fde\u63a5\u8d77\u6765\uff0c\u4ece\u800c\u53ef\u4ee5\u7f13\u89e3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n<p><strong>\u4e0e\u5f53\u4eca\u6a21\u578b\u7684\u76f8\u5173\u6027\u548c\u5173\u952e\u8981\u70b9<\/strong><\/p>\n<p>\u5f53\u4eca\u65f6\u4ee3\uff0c\u8981\u6210\u4e3a\u4e00\u4f4d\u5408\u683c\u7684 LLM \u7814\u7a76\u8005\u548c\u5b9e\u8df5\u8005\uff0c\u4e00\u9879\u5173\u952e\u80fd\u529b\u662f\u80fd\u540c\u65f6\u4ece\u67b6\u6784\u65b9\u9762\u548c\u9884\u8bad\u7ec3\u65b9\u9762\u63a8\u65ad\u5f52\u7eb3\u504f\u7f6e\u3002\u7406\u89e3\u5176\u4e2d\u5fae\u5999\u7684\u5dee\u5f02\u53ef\u5e2e\u52a9\u4eba\u4eec\u8fdb\u884c\u5916\u63a8\u548c\u6301\u7eed\u521b\u65b0\u3002<\/p>\n<p>\u4ee5\u4e0b\u662f\u6211\u7684\u5173\u952e\u8981\u70b9\uff1a<\/p>\n<ol>\n<li>\n<p>\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u548c\u4ec5\u89e3\u7801\u5668\u6a21\u578b\u90fd\u662f<mark data-type=\"concepts\" data-id=\"e70f77cf-89e6-4a0b-b689-c5757a1ecf6c\">\u81ea\u56de\u5f52\u6a21\u578b<\/mark>\uff0c\u5b83\u4eec\u5728\u5b9e\u73b0\u5c42\u9762\u4e0a\u6709\u5dee\u5f02\uff0c\u4e5f\u6709\u5404\u81ea\u7684\u4f18\u7f3a\u70b9\u3002\u5b83\u4eec\u662f\u7565\u6709\u4e0d\u540c\u7684\u5f52\u7eb3\u504f\u7f6e\u3002\u81f3\u4e8e\u9009\u7528\u54ea\u4e00\u79cd\uff0c\u8fd9\u53d6\u51b3\u4e8e\u4e0b\u6e38\u7528\u4f8b\u548c\u5e94\u7528\u9650\u5236\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5bf9\u4e8e\u5927\u591a\u6570 LLM \u7528\u4f8b\u548c\u5229\u57fa\u7528\u4f8b\u800c\u8a00\uff0c\u53ef\u4ee5\u8ba4\u4e3a BERT \u5f0f\u7684\u7f16\u7801\u5668\u6a21\u578b\u5df2\u7ecf\u8fc7\u65f6\u3002<\/p>\n<\/li>\n<li>\n<p>\u53bb\u566a\u76ee\u6807\u4e3b\u8981\u80fd\u4f5c\u4e3a\u56e0\u679c<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u8865\u5145\u3002\u5b83\u4eec\u5df2\u7ecf\u88ab\u6210\u529f\u7528\u4f5c\u8bad\u7ec3\u9636\u6bb5\u7684\u300c\u652f\u6301\u76ee\u6807\u300d\u3002\u4f7f\u7528\u53bb\u566a\u76ee\u6807\u8bad\u7ec3\u56e0\u679c<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u901a\u5e38\u80fd\u5e26\u6765\u67d0\u79cd\u7a0b\u5ea6\u7684\u5e2e\u52a9\u3002\u5c3d\u7ba1\u8fd9\u5728\u4ee3\u7801\u6a21\u578b\u9886\u57df\u975e\u5e38\u5e38\u89c1\uff08\u5373\u4ee3\u7801\u586b\u5145\uff09\uff0c\u4f46\u5bf9\u4e8e\u5982\u4eca\u7684\u901a\u7528\u6a21\u578b\uff0c\u4f7f\u7528\u56e0\u679c<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u52a0\u67d0\u4e2a\u53bb\u566a\u76ee\u6807\u6765\u8fdb\u884c\u9884\u8bad\u7ec3\u7684\u505a\u6cd5\u4e5f\u633a\u5e38\u89c1\u3002<\/p>\n<\/li>\n<li>\n<p>\u53cc\u5411\u6ce8\u610f\u529b\u80fd\u7ed9\u8f83\u5c0f\u89c4\u6a21\u6a21\u578b\u5e26\u6765\u5f88\u5927\u52a9\u76ca\uff0c\u4f46\u5bf9\u8f83\u5927\u6a21\u578b\u6765\u8bf4\u53ef\u6709\u53ef\u65e0\u3002\u8fd9\u5927\u90fd\u662f\u4f20\u8a00\u3002\u6211\u8ba4\u4e3a\u53cc\u5411\u6ce8\u610f\u529b\u5177\u6709\u4e00\u79cd\u5f52\u7eb3\u504f\u7f6e\uff0c\u5c31\u7c7b\u4f3c\u4e8e\u5bf9 Transformer \u6a21\u578b\u505a\u7684\u8bb8\u591a\u5176\u5b83\u7c7b\u578b\u7684\u4fee\u6539\u3002<\/p>\n<\/li>\n<\/ol>\n<p>\u6700\u540e\uff0c\u603b\u7ed3\u4e00\u4e0b\u3002\u76ee\u524d\u5df2\u7ecf\u6ca1\u6709\u5927\u89c4\u6a21\u7248\u672c\u7684 BERT \u6a21\u578b\u5728\u8fd0\u8425\u4e86\uff1aBERT \u6a21\u578b\u5df2\u88ab\u5f03\u7528\uff0c\u53d6\u800c\u4ee3\u4e4b\u7684\u662f\u66f4\u7075\u6d3b\u7684\u53bb\u566a\uff08\u81ea\u56de\u5f52\uff09T5 \u6a21\u578b\u3002\u8fd9\u4e3b\u8981\u662f\u7531\u4e8e\u8303\u5f0f\u7edf\u4e00\uff0c\u5373\u4eba\u4eec\u66f4\u559c\u6b22\u4f7f\u7528\u4e00\u4e2a\u901a\u7528\u6a21\u578b\u53bb\u6267\u884c\u5404\u79cd\u4efb\u52a1\uff08\u800c\u4e0d\u662f\u4f7f\u7528\u7279\u5b9a\u4e8e\u67d0\u4efb\u52a1\u7684\u6a21\u578b\uff09\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u81ea\u56de\u5f52\u53bb\u566a\u6709\u65f6\u53ef\u4f5c\u4e3a\u56e0\u679c<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u526f\u76ee\u6807\u3002<\/p>\n<p><em><sup>\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/www.yitay.net\/blog\/model-architecture-blogpost-encoders-prefixlm-denoising<\/sup><\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-22-14\" target=\"_blank\">\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[73,68,55],"class_list":["post-34660","post","type-post","status-publish","format-standard","hentry","category-news","tag-73","tag-68","tag-55"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=34660\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34660\",\"url\":\"https:\/\/17aitech.com\/?p=34660\",\"name\":\"\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34660#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34660#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png\",\"datePublished\":\"2024-12-10T08:57:02+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34660#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34660\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34660#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34660#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=34660","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34660","url":"https:\/\/17aitech.com\/?p=34660","name":"\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34660#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34660#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png","datePublished":"2024-12-10T08:57:02+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34660#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34660"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34660#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d792aae6677ee7609a32b06685a09e3f.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34660#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u76db\u540d\u4e00\u65f6\u7684BERT\u54ea\u53bb\u4e86\uff1f\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\u662d\u793a\u4e86LLM\u8303\u5f0f\u7684\u8f6c\u53d8"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34660","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34660"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34660\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34660"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34660"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34660"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}