{"id":5094,"date":"2024-11-15T12:00:31","date_gmt":"2024-11-15T04:00:31","guid":{"rendered":"https:\/\/nullthought.net\/?p=5094"},"modified":"2025-02-23T11:08:03","modified_gmt":"2025-02-23T03:08:03","slug":"%e9%80%9a%e8%bf%87%e4%bb%a4%e7%89%8c%e9%80%89%e6%8b%a9%ef%bc%88token-selection%ef%bc%89%ef%bc%8c%e5%ae%9e%e7%8e%b0transformer%e7%9a%84%e5%86%85%e5%ad%98%e9%ab%98%e6%95%88%e5%be%ae%e8%b0%83","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5094","title":{"rendered":"\u901a\u8fc7\u4ee4\u724c\u9009\u62e9\uff08Token Selection\uff09\uff0c\u5b9e\u73b0Transformer\u7684\u5185\u5b58\u9ad8\u6548\u5fae\u8c03"},"content":{"rendered":"\n<p>\u5fae\u8c03\u662f\u5229\u7528\u9884\u8bad\u7ec3\u7684\u5927\u578b\u6a21\u578b\u5728\u7279\u5b9a\u4e0b\u6e38\u4efb\u52a1\u4e0a\u83b7\u5f97\u6700\u4f73\u6027\u80fd\u7684\u4e00\u79cd\u6709\u6548\u65b9\u6cd5\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u548c\u5176\u4ed6\u57fa\u4e8eTransformer\u7684\u6a21\u578b\u3002\u7136\u800c\uff0c\u5fae\u8c03\u8fd9\u4e9b\u5927\u578b\u6a21\u578b\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u8d44\u6e90\u548c\u5185\u5b58\u8d44\u6e90\uff0c\u5c24\u5176\u662f\u5f53\u6d89\u53ca\u5230\u6570\u5341\u4ebf\u53c2\u6570\u7684\u6a21\u578b\u65f6\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8eGPT-3\uff08175B\uff09\u6216Llama\uff0865B\uff09\u8fd9\u6837\u7684\u5927\u6a21\u578b\uff0c\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u9700\u8981\u5206\u522b\u9ad8\u8fbe1200GB\u548c780GB\u7684GPU\u5185\u5b58\u3002\u8fd9\u79cd\u8d44\u6e90\u9700\u6c42\u663e\u7136\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u662f\u6781\u5176\u6602\u8d35\u548c\u96be\u4ee5\u5b9e\u73b0\u7684\uff0c\u5c24\u5176\u5bf9\u4e8e\u90a3\u4e9b\u8d44\u6e90\u6709\u9650\u7684\u7ec4\u7ec7\u6216\u4e2a\u4eba\u5f00\u53d1\u8005\u800c\u8a00\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u5728\u4fdd\u8bc1\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6\u51cf\u5c11\u5185\u5b58\u6d88\u8017\u6210\u4e3a\u4e86\u7814\u7a76\u4eba\u5458\u5173\u6ce8\u7684\u91cd\u70b9\u3002<\/p>\n\n\n\n<p>\u73b0\u6709\u7684\u4e00\u4e9b\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff08Parameter-Efficient Fine-Tuning\uff0cPEFT\uff09\u65b9\u6cd5\uff0c\u6bd4\u5982<a href=\"https:\/\/nullthought.net\/?p=4066\" target=\"_blank\" rel=\"noreferrer noopener\">LoRA<\/a>\u3001QLoRA\u7b49\uff0c\u5df2\u7ecf\u5c1d\u8bd5\u901a\u8fc7\u51cf\u5c11\u53c2\u6570\u91cf\u6216\u8fdb\u884c\u53c2\u6570\u91cf\u5316\u6765\u964d\u4f4e\u5185\u5b58\u9700\u6c42\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u4ecd\u7136\u9700\u8981\u5728\u524d\u5411\u4f20\u64ad\u65f6\u7f13\u5b58\u6240\u6709\u7684\u4e2d\u95f4\u6fc0\u6d3b\uff0c\u4ee5\u4fbf\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u7528\u4e8e\u68af\u5ea6\u8ba1\u7b97\u3002\u8fd9\u4e9b\u4e2d\u95f4\u6fc0\u6d3b\u5360\u636e\u4e86\u5927\u91cf\u7684\u5185\u5b58\uff0c\u5c24\u5176\u662f\u5728\u5927\u6a21\u578b\u4e0a\u3002\u56e0\u6b64\uff0c\u672c\u6587<strong><a href=\"https:\/\/aclanthology.org\/2024.emnlp-main.1202.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Memory-Efficient Fine-Tuning of Transformers via Token Selection<\/a><\/strong>\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u5185\u5b58\u9ad8\u6548\u5fae\u8c03\u65b9\u6cd5\uff0c\u540d\u4e3a<strong>TOKENTUNE<\/strong>\uff0c\u65e8\u5728\u901a\u8fc7\u9009\u62e9\u4e00\u90e8\u5206\u8f93\u5165\u4ee4\u724c\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u51cf\u5c11\u5bf9\u4e2d\u95f4\u6fc0\u6d3b\u7684\u5b58\u50a8\u9700\u6c42\uff0c\u4ece\u800c\u964d\u4f4e\u6574\u4f53\u7684\u5185\u5b58\u6d88\u8017\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aAntoine Simoulin, , Namyong Park, Xiaoyi Liu, Grey Yang\uff0c\u5747\u6765\u81eaMeta\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"882\" height=\"504\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/11\/image-14.png\" alt=\"\" class=\"wp-image-5095\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/11\/image-14.png 882w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/11\/image-14-300x171.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/11\/image-14-768x439.png 768w\" sizes=\"auto, (max-width: 882px) 100vw, 882px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/aclanthology.org\/2024.emnlp-main.1202.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Memory-Efficient Fine-Tuning of Transformers via Token Selection<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u5982\u4e0b\u4e3a\u8bba\u6587\u6982\u8981\u5185\u5bb9\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e00\u3001\u65b9\u6cd5\u4ecb\u7ecd\uff1aTOKENTUNE<\/h3>\n\n\n\n<p><strong>TOKENTUNE \u7684\u57fa\u672c\u7406\u5ff5\u662f\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u4ec5\u5bf9\u9009\u5b9a\u7684\u8f93\u5165\u4ee4\u724c\u8fdb\u884c\u68af\u5ea6\u8ba1\u7b97\uff0c\u8fd9\u610f\u5473\u7740\u5728\u524d\u5411\u4f20\u64ad\u65f6\u53ea\u9700\u7f13\u5b58\u8fd9\u4e9b\u9009\u5b9a\u4ee4\u724c\u7684\u4e2d\u95f4\u6fc0\u6d3b\uff0c\u4ece\u800c\u5927\u5e45\u964d\u4f4e\u5185\u5b58\u5f00\u9500<\/strong>\u3002\u4e3a\u4e86\u66f4\u597d\u5730\u7406\u89e3TOKENTUNE\u7684\u5de5\u4f5c\u673a\u5236\uff0c\u4e0b\u9762\u5c06\u4ece\u51e0\u4e2a\u65b9\u9762\u8be6\u7ec6\u63cf\u8ff0\u5176\u8fc7\u7a0b\u4e0e\u65b9\u6cd5\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u4ee4\u724c\u9009\u62e9\u7b56\u7565<\/h4>\n\n\n\n<p><strong>\u4ee4\u724c\u9009\u62e9\u7b56\u7565<\/strong> \u662fTOKENTUNE\u7684\u6838\u5fc3\u90e8\u5206\u3002\u5728\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u65f6\uff0c\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u4ee4\u724c\u88ab\u5212\u5206\u4e3a\u4e24\u90e8\u5206\uff1a\u4e00\u90e8\u5206\u88ab\u9009\u4e3a\u201c\u53c2\u4e0e\u5fae\u8c03\u7684\u4ee4\u724c\u201d\uff0c\u53e6\u4e00\u90e8\u5206\u5219\u88ab\u201c\u51bb\u7ed3\u201d\uff0c\u5373\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u4e0d\u53c2\u4e0e\u68af\u5ea6\u8ba1\u7b97\u3002\u5177\u4f53\u6765\u8bf4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bba\u6587\u91c7\u7528\u4e86\u4e00\u79cd<strong>\u968f\u673a\u4ee4\u724c\u9009\u62e9\u7b56\u7565<\/strong>\uff0c\u5373\u5728\u6bcf\u4e2a\u8f93\u5165\u5e8f\u5217\u4e2d\u968f\u673a\u9009\u62e9\u82e5\u5e72\u4e2a\u4ee4\u724c\u53c2\u4e0e\u53cd\u5411\u4f20\u64ad\u3002\u8fd9\u4e9b\u9009\u5b9a\u7684\u4ee4\u724c\u5c06\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u751f\u6210\u68af\u5ea6\uff0c\u800c\u672a\u9009\u4e2d\u7684\u90e8\u5206\u5219\u5728\u8ba1\u7b97\u65f6\u88ab\u5ffd\u7565\u3002<\/li>\n\n\n\n<li>\u968f\u673a\u9009\u62e9\u4ee4\u724c\u7684\u8fc7\u7a0b\u4f7f\u7528\u7684\u662f\u5747\u5300\u5206\u5e03\uff0c\u4ee5\u786e\u4fdd\u6ca1\u6709\u4efb\u4f55\u4ee4\u724c\u88ab\u504f\u8892\u3002\u7136\u800c\uff0c\u4e3a\u4e86\u4f7f\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u7406\u89e3\u6574\u4e2a\u8f93\u5165\u8bed\u4e49\uff0c\u5728\u6bcf\u6b21\u9009\u62e9\u65f6\uff0c\u603b\u662f\u5305\u542b\u8868\u793a\u5168\u5c40\u4fe1\u606f\u7684\u7279\u6b8a\u4ee4\u724c\uff08\u5982CLS\u4ee4\u724c\uff09\u3002<\/li>\n\n\n\n<li>\u5728\u5177\u4f53\u5b9e\u73b0\u4e0a\uff0c\u8bba\u6587\u4e2d\u63d0\u5230\u4f7f\u7528PyTorch\u4e2d\u7684<code>torch.no_grad()<\/code>\u673a\u5236\u6765\u907f\u514d\u4e3a\u672a\u9009\u4e2d\u4ee4\u724c\u7684\u8ba1\u7b97\u751f\u6210\u68af\u5ea6\uff0c\u4ece\u800c\u964d\u4f4e\u5185\u5b58\u4f7f\u7528\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2. \u5bf9\u6a21\u578b\u4e0d\u540c\u5c42\u7684\u5904\u7406<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u4f7f\u4ee4\u724c\u9009\u62e9\u7b56\u7565\u80fd\u591f\u5728Transformer\u67b6\u6784\u4e2d\u5f97\u5230\u6709\u6548\u5e94\u7528\uff0cTOKENTUNE\u5bf9\u4e0d\u540c\u7c7b\u578b\u7684\u5c42\uff08\u5982\u5bc6\u96c6\u5c42\u548c\u6ce8\u610f\u529b\u5c42\uff09\u8fdb\u884c\u4e86\u7279\u522b\u5904\u7406\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5bc6\u96c6\u5c42\uff08Dense Layer\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\uff0c\u666e\u901a\u7684\u5fae\u8c03\u65b9\u6cd5\u9700\u8981\u8ba1\u7b97\u6240\u6709\u8f93\u5165\u7684\u68af\u5ea6\uff0c\u8fd9\u5305\u62ec\u8ba1\u7b97\u4e0e\u8f93\u5165\u3001\u6743\u91cd\u3001\u504f\u7f6e\u7b49\u76f8\u5173\u7684\u6240\u6709\u4e2d\u95f4\u6fc0\u6d3b\u3002\u901a\u8fc7\u53ea\u9009\u62e9\u90e8\u5206\u4ee4\u724c\uff0cTOKENTUNE\u51cf\u5c11\u4e86\u53cd\u5411\u4f20\u64ad\u6240\u9700\u7f13\u5b58\u7684\u6fc0\u6d3b\u6570\u91cf\u3002<\/li>\n\n\n\n<li>\u5728\u5b9e\u73b0\u4e2d\uff0c\u4f5c\u8005\u5c06\u6fc0\u6d3b\u62c6\u5206\u4e3a\u4e24\u90e8\u5206\uff1a\u9009\u5b9a\u4ee4\u724c\u7684\u6fc0\u6d3b\u548c\u672a\u9009\u5b9a\u4ee4\u724c\u7684\u6fc0\u6d3b\u3002\u5728\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\uff0c\u8ba1\u7b97\u6240\u6709\u6fc0\u6d3b\uff0c\u4f46\u5728\u53cd\u5411\u4f20\u64ad\u65f6\uff0c\u4ec5\u5bf9\u9009\u5b9a\u90e8\u5206\u7684\u6fc0\u6d3b\u751f\u6210\u68af\u5ea6\uff0c\u4ece\u800c\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u6ce8\u610f\u529b\u5c42\uff08Attention Layer\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6ce8\u610f\u529b\u673a\u5236\u662fTransformer\u6a21\u578b\u7684\u91cd\u8981\u7ec4\u6210\u90e8\u5206\uff0c\u6d89\u53ca\u5bf9\u6574\u4e2a\u8f93\u5165\u5e8f\u5217\u8fdb\u884c\u64cd\u4f5c\u3002\u4e3a\u4e86\u9002\u5e94\u90e8\u5206\u4ee4\u724c\u7684\u9009\u62e9\u7b56\u7565\uff0cTOKENTUNE\u5bf9\u67e5\u8be2\u3001\u952e\u548c\u503c\uff08Query, Key, Value\uff09\u7684\u8ba1\u7b97\u4e5f\u8fdb\u884c\u4e86\u62c6\u5206\u3002<\/li>\n\n\n\n<li>\u5728\u524d\u5411\u4f20\u64ad\u65f6\uff0c\u6240\u6709\u7684\u67e5\u8be2\u3001\u952e\u548c\u503c\u90fd\u88ab\u8ba1\u7b97\uff0c\u4f46\u5728\u53cd\u5411\u4f20\u64ad\u65f6\uff0c\u53ea\u8ba1\u7b97\u9009\u5b9a\u4ee4\u724c\u7684\u68af\u5ea6\u3002\u8fd9\u79cd\u5904\u7406\u65b9\u5f0f\u4f7f\u5f97\u6a21\u578b\u5728\u4fdd\u7559\u5168\u5c40\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u540c\u65f6\uff0c\u663e\u8457\u51cf\u5c11\u4e86\u53cd\u5411\u4f20\u64ad\u4e2d\u7684\u5185\u5b58\u9700\u6c42\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e8c\u3001\u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c\u5206\u6790<\/h3>\n\n\n\n<p>\u8bba\u6587\u901a\u8fc7\u5927\u91cf\u5b9e\u9a8c\u9a8c\u8bc1\u4e86TOKENTUNE\u5728\u4e0d\u540c\u89c4\u6a21\u7684\u6a21\u578b\u4e0a\u7684\u6709\u6548\u6027\uff0c\u6db5\u76d6\u4e86\u4ece\u4e2d\u7b49\u89c4\u6a21\uff08\u5982BERT\uff09\u5230\u5927\u89c4\u6a21\u6a21\u578b\uff08\u5982Llama\uff09\u7684\u591a\u4e2a\u4e0b\u6e38\u4efb\u52a1\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u5fae\u8c03\u4e2d\u7b49\u89c4\u6a21\u6a21\u578b\u7684\u5b9e\u9a8c<\/h4>\n\n\n\n<p>\u5728\u5bf9\u4e2d\u7b49\u89c4\u6a21\u7684\u7f16\u7801\u5668\u6a21\u578b\uff08\u5982BERT\uff09\u8fdb\u884c\u5b9e\u9a8c\u65f6\uff0c\u7814\u7a76\u8005\u9009\u62e9\u4e86GLUE\u57fa\u51c6\u6d4b\u8bd5\u4f5c\u4e3a\u8bc4\u4f30\u57fa\u51c6\u3002GLUE\u57fa\u51c6\u6d4b\u8bd5\u662f\u4e00\u7ec4\u5305\u542b\u591a\u79cd\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u4efb\u52a1\u7684\u6570\u636e\u96c6\uff0c\u7528\u4e8e\u8bc4\u4f30\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002\u5b9e\u9a8c\u8bbe\u8ba1\u4e2d\u4f7f\u7528\u4e86BERT-large\u6a21\u578b\uff0c\u5e76\u5bf9\u6bd4\u4e86\u4ee5\u4e0b\u51e0\u79cd\u5fae\u8c03\u65b9\u6cd5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5168\u91cf\u5fae\u8c03\uff08Full Fine-Tuning\uff09<\/strong>\uff1a\u5bf9\u6a21\u578b\u7684\u6240\u6709\u53c2\u6570\u8fdb\u884c\u66f4\u65b0\uff0c\u5185\u5b58\u9700\u6c42\u9ad8\uff0c\u4f46\u901a\u5e38\u80fd\u8fbe\u5230\u6700\u4f73\u6027\u80fd\u3002<\/li>\n\n\n\n<li><strong>\u4e0d\u540cPEFT\u65b9\u6cd5<\/strong>\uff1a\u5305\u62ecLoRA\u3001Ladder Side Tuning\u7b49\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u5c1d\u8bd5\u901a\u8fc7\u51cf\u5c11\u53c2\u6570\u91cf\u6765\u964d\u4f4e\u5185\u5b58\u9700\u6c42\u3002<\/li>\n\n\n\n<li><strong>TOKENTUNE<\/strong>\uff1a\u4ec5\u9009\u62e916\u4e2a\u8f93\u5165\u4f4d\u7f6e\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u9a8c\u8bc1\u5176\u5728\u51cf\u5c11\u5185\u5b58\u6d88\u8017\u7684\u540c\u65f6\u662f\u5426\u80fd\u4fdd\u6301\u4e0e\u5168\u91cf\u5fae\u8c03\u76f8\u8fd1\u7684\u6027\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>TOKENTUNE\u5728GLUE\u4e0a\u7684\u5e73\u5747\u5f97\u5206\u63a5\u8fd1\u5168\u91cf\u5fae\u8c03\uff0c\u5e76\u4e14\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\uff08\u5982CoLA\u4efb\u52a1\uff09\u6027\u80fd\u4e0e\u5176\u4ed6PEFT\u65b9\u6cd5\u6301\u5e73\u751a\u81f3\u66f4\u4f18\u3002<\/li>\n\n\n\n<li>\u5728GPU\u5185\u5b58\u7684\u4f7f\u7528\u4e0a\uff0c\u4e0e\u5168\u91cf\u5fae\u8c03\u76f8\u6bd4\uff0cTOKENTUNE\u4ec5\u4f7f\u7528\u4e86\u7ea6\u4e00\u534a\u7684\u5185\u5b58\uff0c\u8fd9\u610f\u5473\u7740\u5176\u53ef\u4ee5\u5728\u66f4\u4f4e\u7684\u786c\u4ef6\u8981\u6c42\u4e0b\u5b9e\u73b0\u5bf9\u6a21\u578b\u7684\u9ad8\u6548\u5fae\u8c03\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2. \u5fae\u8c03\u5927\u89c4\u6a21\u6a21\u578b\u7684\u5b9e\u9a8c<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u9a8c\u8bc1TOKENTUNE\u5728\u5927\u89c4\u6a21\u6a21\u578b\u4e0a\u7684\u8868\u73b0\uff0c\u7814\u7a76\u8005\u8fd8\u5bf9Llama2-7B\u6a21\u578b\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u7279\u522b\u662f\u901a\u8fc7<strong>\u6307\u4ee4\u5fae\u8c03\uff08Instruction Tuning\uff09\u548c\u5c11\u6837\u672c\u8bc4\u4f30\uff08Few-Shot Evaluation\uff09<\/strong>\u7684\u65b9\u5f0f\uff0c\u6765\u8bc4\u4f30\u6a21\u578b\u5728\u5904\u7406\u590d\u6742\u4efb\u52a1\u4e0a\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6307\u4ee4\u5fae\u8c03<\/strong>\uff1a\u4f7f\u7528\u5f00\u6e90\u7684Open-Platypus\u6570\u636e\u96c6\u5bf9Llama2-7B\u8fdb\u884c\u5fae\u8c03\uff0c\u5e76\u7ed3\u5408LoRA\u548cQLoRA\u7b49\u65b9\u6cd5\u8fdb\u4e00\u6b65\u964d\u4f4e\u5185\u5b58\u9700\u6c42\u3002<\/li>\n\n\n\n<li><strong>\u5c11\u6837\u672c\u8bc4\u4f30<\/strong>\uff1a\u5728\u591a\u4e2a\u4efb\u52a1\uff08\u5982MMLU\u3001ARC\u3001HellaSwag\u3001TruthfulQA\u7b49\uff09\u4e0a\u8fdb\u884c\u5c11\u6837\u672c\u6d4b\u8bd5\uff0c\u4ee5\u8bc4\u4f30\u6a21\u578b\u5728\u53ea\u7ed9\u51fa\u5c11\u91cf\u793a\u4f8b\u60c5\u51b5\u4e0b\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>TOKENTUNE\u4e0eLoRA\u3001QLoRA\u7ed3\u5408\u4f7f\u7528\u65f6\uff0c\u53ef\u4ee5\u663e\u8457\u964d\u4f4e\u5927\u89c4\u6a21\u6a21\u578b\u7684\u5185\u5b58\u9700\u6c42\u3002\u5728Llama2-7B\u6a21\u578b\u7684\u5fae\u8c03\u4e2d\uff0c\u901a\u8fc7\u7ed3\u5408QLoRA\u548cTOKENTUNE\uff0c\u5185\u5b58\u9700\u6c42\u964d\u4f4e\u81f3\u5168\u91cf\u5fae\u8c03\u7684\u4e09\u5206\u4e4b\u4e00\u3002<\/li>\n\n\n\n<li>\u5728\u5c11\u6837\u672c\u4efb\u52a1\u7684\u8868\u73b0\u4e0a\uff0cTOKENTUNE\u4e0e\u5176\u4ed6\u65b9\u6cd5\u76f8\u5f53\uff0c\u751a\u81f3\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u7565\u6709\u63d0\u9ad8\uff0c\u8fd9\u8868\u660e\u5728\u51cf\u5c11\u5185\u5b58\u9700\u6c42\u7684\u540c\u65f6\uff0c\u6027\u80fd\u635f\u5931\u53ef\u4ee5\u5ffd\u7565\u4e0d\u8ba1\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">3. \u4ee4\u724c\u9009\u62e9\u6bd4\u4f8b\u7684\u5f71\u54cd<\/h4>\n\n\n\n<p>\u5b9e\u9a8c\u8fd8\u7814\u7a76\u4e86\u4e0d\u540c\u4ee4\u724c\u9009\u62e9\u6bd4\u4f8b\u5bf9\u6a21\u578b\u6027\u80fd\u548c\u5185\u5b58\u9700\u6c42\u7684\u5f71\u54cd\u3002\u5728GLUE\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u901a\u8fc7\u9009\u62e9\u4e0d\u540c\u6570\u91cf\u7684\u8f93\u5165\u4ee4\u724c\u8fdb\u884c\u5fae\u8c03\uff0c\u53d1\u73b0\u5f53\u9009\u5b9a\u4ee4\u724c\u7684\u6bd4\u4f8b\u589e\u52a0\u65f6\uff0c\u6a21\u578b\u7684\u6027\u80fd\u9010\u6e10\u63d0\u9ad8\uff0c\u4f46\u5728\u8fbe\u5230\u67d0\u4e00\u9608\u503c\u540e\uff0c\u6027\u80fd\u7684\u63d0\u5347\u8d8b\u4e8e\u5e73\u7a33\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f8b\u5982\uff0c\u5f53\u9009\u62e916\u4e2a\u4ee4\u724c\u65f6\uff0c\u6a21\u578b\u7684\u6027\u80fd\u5df2\u7ecf\u63a5\u8fd1\u4e8e\u9009\u62e9\u5168\u90e8\u4ee4\u724c\u65f6\u7684\u6027\u80fd\uff0c\u4f46\u5185\u5b58\u6d88\u8017\u5374\u663e\u8457\u51cf\u5c11\u3002\u56e0\u6b64\uff0c\u5bf9\u4e8e\u5927\u591a\u6570\u4efb\u52a1\uff0c\u9009\u62e920%-30%\u7684\u4ee4\u724c\u4f5c\u4e3a\u9ed8\u8ba4\u6bd4\u4f8b\u53ef\u4ee5\u5728\u6027\u80fd\u548c\u5185\u5b58\u4e4b\u95f4\u53d6\u5f97\u826f\u597d\u7684\u5e73\u8861\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e09\u3001\u65b9\u6cd5\u4f18\u52bf\u4e0e\u5c40\u9650\u6027<\/h3>\n\n\n\n<p><strong>\u4f18\u52bf<\/strong>\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5185\u5b58\u6548\u7387\u663e\u8457\u63d0\u9ad8<\/strong>\uff1a\u901a\u8fc7\u4ec5\u9009\u62e9\u90e8\u5206\u8f93\u5165\u4ee4\u724c\u8fdb\u884c\u68af\u5ea6\u8ba1\u7b97\uff0cTOKENTUNE\u663e\u8457\u964d\u4f4e\u4e86\u5b58\u50a8\u4e2d\u95f4\u6fc0\u6d3b\u7684\u5185\u5b58\u9700\u6c42\uff0c\u5c24\u5176\u662f\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u3002<\/li>\n\n\n\n<li><strong>\u4e0e\u73b0\u6709\u65b9\u6cd5\u517c\u5bb9<\/strong>\uff1aTOKENTUNE\u53ef\u4ee5\u4e0e\u5176\u4ed6PEFT\u65b9\u6cd5\u7ed3\u5408\uff0c\u5982LoRA\u548cQLoRA\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u51cf\u5c11\u5185\u5b58\u4f7f\u7528\u3002\u8fd9\u79cd\u517c\u5bb9\u6027\u4f7f\u5f97\u5176\u5e94\u7528\u66f4\u52a0\u7075\u6d3b\uff0c\u53ef\u4ee5\u6839\u636e\u4efb\u52a1\u548c\u786c\u4ef6\u6761\u4ef6\u9009\u62e9\u5408\u9002\u7684\u7ec4\u5408\u7b56\u7565\u3002<\/li>\n\n\n\n<li><strong>\u6027\u80fd\u4fdd\u6301\u4e0d\u53d8\u6216\u7565\u6709\u63d0\u5347<\/strong>\uff1a\u5c3d\u7ba1\u51cf\u5c11\u4e86\u5185\u5b58\u9700\u6c42\uff0c\u4f46\u5728\u5b9e\u9a8c\u4e2d\uff0cTOKENTUNE\u5728\u591a\u4e2a\u4efb\u52a1\u4e0a\u90fd\u8868\u73b0\u51fa\u4e0e\u5168\u91cf\u5fae\u8c03\u76f8\u8fd1\u7684\u6027\u80fd\u3002\u8fd9\u8868\u660e\u8be5\u65b9\u6cd5\u5728\u4e0d\u727a\u7272\u6a21\u578b\u80fd\u529b\u7684\u60c5\u51b5\u4e0b\uff0c\u5b9e\u73b0\u4e86\u66f4\u9ad8\u7684\u8ba1\u7b97\u6548\u7387\u3002<\/li>\n<\/ol>\n\n\n\n<p><strong>\u5c40\u9650\u6027<\/strong>\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u9009\u62e9\u4ee4\u724c\u7684\u7b56\u7565\u7b80\u5355<\/strong>\uff1a\u5f53\u524d\u7248\u672c\u7684TOKENTUNE\u4f7f\u7528\u7684\u662f\u968f\u673a\u9009\u62e9\u4ee4\u724c\u7684\u7b56\u7565\uff0c\u8fd9\u79cd\u65b9\u6cd5\u867d\u7136\u7b80\u5355\u6709\u6548\uff0c\u4f46\u53ef\u80fd\u4e0d\u662f\u6700\u4f18\u7684\u3002\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u5c1d\u8bd5\u5f15\u5165\u66f4\u667a\u80fd\u7684\u9009\u62e9\u673a\u5236\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6027\u80fd\u3002<\/li>\n\n\n\n<li><strong>\u5bf9\u7279\u5b9a\u4efb\u52a1\u7684\u4f9d\u8d56\u6027<\/strong>\uff1a\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\uff0c\u51cf\u5c11\u4ee4\u724c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u9700\u8981\u7cbe\u786e\u6355\u6349\u4e0a\u4e0b\u6587\u5173\u7cfb\u7684\u4efb\u52a1\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u9488\u5bf9\u4e0d\u540c\u4efb\u52a1\u4f18\u5316\u4ee4\u724c\u9009\u62e9\u7b56\u7565\u4ecd\u7136\u662f\u4e00\u4e2a\u5f00\u653e\u7684\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u5176\u4ed6\u9886\u57df\u7684\u5e94\u7528\u5c1a\u5f85\u9a8c\u8bc1<\/strong>\uff1a\u867d\u7136TOKENTUNE\u5728\u8bed\u8a00\u6a21\u578b\u7684\u5fae\u8c03\u4e2d\u8868\u73b0\u826f\u597d\uff0c\u4f46\u5176\u5728\u5176\u4ed6\u9886\u57df\uff08\u5982\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u7684Transformer\u67b6\u6784\uff09\u4e2d\u7684\u6548\u679c\u5c1a\u672a\u5f97\u5230\u9a8c\u8bc1\u3002\u672a\u6765\u7684\u5de5\u4f5c\u53ef\u4ee5\u63a2\u7d22\u5176\u5728\u66f4\u5e7f\u6cdb\u9886\u57df\u4e2d\u7684\u9002\u7528\u6027\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">\u56db\u3001\u7ed3\u8bba\u4e0e\u672a\u6765\u5de5\u4f5c<\/h3>\n\n\n\n<p><strong>\u7ed3\u8bba<\/strong>\uff1a\u672c\u6587\u63d0\u51fa\u4e86TOKENTUNE\uff0c\u4e00\u79cd\u901a\u8fc7\u9009\u62e9\u8f93\u5165\u4ee4\u724c\u5b9e\u73b0\u5185\u5b58\u9ad8\u6548\u5fae\u8c03\u7684\u521b\u65b0\u65b9\u6cd5\u3002\u901a\u8fc7\u5728\u524d\u5411\u4f20\u64ad\u4e2d\u53ea\u7f13\u5b58\u90e8\u5206\u4e2d\u95f4\u6fc0\u6d3b\u72b6\u6001\uff0cTOKENTUNE\u663e\u8457\u964d\u4f4e\u4e86\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u5bf9GPU\u5185\u5b58\u7684\u9700\u6c42\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u8be5\u65b9\u6cd5\u80fd\u591f\u4e0e\u5176\u4ed6PEFT\u65b9\u6cd5\u7ed3\u5408\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u5185\u5b58\u5f00\u9500\uff0c\u540c\u65f6\u4fdd\u6301\u4e0e\u5168\u91cf\u5fae\u8c03\u76f8\u8fd1\u7684\u6a21\u578b\u6027\u80fd\u3002<\/p>\n\n\n\n<p><strong>\u672a\u6765\u5de5\u4f5c<\/strong>\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4f18\u5316\u4ee4\u724c\u9009\u62e9\u7b56\u7565<\/strong>\uff1a\u672a\u6765\u53ef\u4ee5\u63a2\u7d22\u66f4\u667a\u80fd\u7684\u4ee4\u724c\u9009\u62e9\u673a\u5236\uff0c\u6bd4\u5982\u57fa\u4e8e\u4ee4\u724c\u7684\u91cd\u8981\u6027\u8bc4\u5206\u6216\u4e0a\u4e0b\u6587\u4f9d\u8d56\u5173\u7cfb\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u5fae\u8c03\u7684\u6548\u7387\u548c\u6548\u679c\u3002<\/li>\n\n\n\n<li><strong>\u6269\u5c55\u5230\u5176\u4ed6\u9886\u57df<\/strong>\uff1a\u76ee\u524d\u7684\u7814\u7a76\u4e3b\u8981\u96c6\u4e2d\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\uff0c\u672a\u6765\u53ef\u4ee5\u5c06\u8be5\u65b9\u6cd5\u6269\u5c55\u5230\u5176\u4ed6\u5e94\u7528\u9886\u57df\uff0c\u5982\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u7684\u5927\u89c4\u6a21\u6a21\u578b\uff0c\u9a8c\u8bc1\u5176\u5728\u4e0d\u540c\u9886\u57df\u4e2d\u7684\u9002\u7528\u6027\u548c\u6709\u6548\u6027\u3002<\/li>\n\n\n\n<li><strong>\u7406\u8bba\u5206\u6790\u4e0e\u89e3\u91ca\u6027<\/strong>\uff1a\u8fdb\u4e00\u6b65\u7684\u5de5\u4f5c\u53ef\u4ee5\u4ece\u7406\u8bba\u4e0a\u5206\u6790\u4e3a\u4ec0\u4e48\u4ec5\u9009\u62e9\u90e8\u5206\u4ee4\u724c\u5c31\u53ef\u4ee5\u4fdd\u6301\u4e0e\u5168\u91cf\u5fae\u8c03\u76f8\u8fd1\u7684\u6027\u80fd\uff0c\u4ee5\u53ca\u5982\u4f55\u89e3\u91ca\u8fd9\u79cd\u73b0\u8c61\u80cc\u540e\u7684\u673a\u5236\uff0c\u8fd9\u5bf9\u4e8e\u7406\u89e3\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u5185\u5728\u5de5\u4f5c\u539f\u7406\u5177\u6709\u91cd\u8981\u610f\u4e49\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u5fae\u8c03\u662f\u5229\u7528\u9884\u8bad\u7ec3\u7684\u5927\u578b\u6a21\u578b\u5728\u7279\u5b9a\u4e0b\u6e38\u4efb\u52a1\u4e0a\u83b7\u5f97\u6700\u4f73\u6027\u80fd\u7684\u4e00\u79cd\u6709\u6548\u65b9\u6cd5\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u548c\u5176\u4ed6 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8],"tags":[39,77,95,86],"class_list":["post-5094","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-meta","tag-transformer","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u5fae\u8c03\u662f\u5229\u7528\u9884\u8bad\u7ec3\u7684\u5927\u578b\u6a21\u578b\u5728\u7279\u5b9a\u4e0b\u6e38\u4efb\u52a1\u4e0a\u83b7\u5f97\u6700\u4f73\u6027\u80fd\u7684\u4e00\u79cd\u6709\u6548\u65b9\u6cd5\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u548c\u5176\u4ed6&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5094","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5094"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5094\/revisions"}],"predecessor-version":[{"id":5097,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5094\/revisions\/5097"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5094"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5094"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5094"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}