{"id":4314,"date":"2024-07-17T14:52:18","date_gmt":"2024-07-17T06:52:18","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=4314"},"modified":"2024-07-17T15:48:37","modified_gmt":"2024-07-17T07:48:37","slug":"smollm-%e9%80%9f%e5%ba%a6%e6%9e%81%e5%bf%ab%ef%bc%8c%e5%8a%9f%e8%83%bd%e5%bc%ba%e5%a4%a7","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=4314","title":{"rendered":"SmolLM &#8211; \u901f\u5ea6\u6781\u5feb\uff0c\u529f\u80fd\u5f3a\u5927"},"content":{"rendered":"\n<p>\u8fd9\u7bc7\u535a\u6587\u4ecb\u7ecd\u4e86&nbsp;<a href=\"https:\/\/huggingface.co\/collections\/HuggingFaceTB\/smollm-models-6695016cad7167254ce15966\">SmolLM\uff0c<\/a>\u8fd9\u662f\u4e00\u4e2a\u5177\u6709 135M\u3001360M \u548c 1.7B \u53c2\u6570\u7684\u6700\u5148\u8fdb\u7684\u5c0f\u578b\u6a21\u578b\u7cfb\u5217\uff0c\u5728\u65b0\u7684\u9ad8\u8d28\u91cf\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u8bad\u7ec3\u3002\u5b83\u6db5\u76d6\u4e86\u6570\u636e\u7ba1\u7406\u3001\u6a21\u578b\u8bc4\u4f30\u548c\u4f7f\u7528\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#introduction\"><\/a>\u4ecb\u7ecd<\/h2>\n\n\n\n<p>\u4eba\u4eec\u5bf9\u53ef\u4ee5\u5728\u672c\u5730\u8bbe\u5907\u4e0a\u8fd0\u884c\u7684\u5c0f\u578b\u8bed\u8a00\u6a21\u578b\u8d8a\u6765\u8d8a\u611f\u5174\u8da3\u3002\u8fd9\u79cd\u8d8b\u52bf\u6d89\u53ca\u84b8\u998f\u6216\u91cf\u5316\u7b49\u6280\u672f\u6765\u538b\u7f29\u5927\u578b\u6a21\u578b\uff0c\u4ee5\u53ca\u5728\u5927\u578b\u6570\u636e\u96c6\u4e0a\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u5c0f\u578b\u6a21\u578b\u3002\u8fd9\u4e9b\u65b9\u6cd5\u652f\u6301\u65b0\u7684\u5e94\u7528\u7a0b\u5e8f\uff0c\u540c\u65f6\u5927\u5927\u964d\u4f4e\u63a8\u7406\u6210\u672c\u5e76\u6539\u5584\u7528\u6237\u9690\u79c1\u3002<\/p>\n\n\n\n<p>Microsoft \u7684 Phi \u7cfb\u5217\u3001\u963f\u91cc\u5df4\u5df4\u7684 Qwen2\uff08\u5c0f\u4e8e 2B\uff09\u548c Meta \u7684 MobileLLM \u8868\u660e\uff0c\u7ecf\u8fc7\u6df1\u601d\u719f\u8651\u7684\u8bbe\u8ba1\u548c\u8bad\u7ec3\uff0c\u5c0f\u578b\u6a21\u578b\u53ef\u4ee5\u53d6\u5f97\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u7ed3\u679c\u3002\u4f46\u662f\uff0c\u6709\u5173\u8fd9\u4e9b\u6a21\u578b\u7684\u6570\u636e\u7ba1\u7406\u548c\u8bad\u7ec3\u7684\u5927\u591a\u6570\u7ec6\u8282\u5c1a\u672a\u516c\u5f00\u3002<\/p>\n\n\n\n<p>\u5728\u8fd9\u7bc7\u535a\u6587\u4e2d\uff0c\u6211\u4eec\u5f88\u9ad8\u5174\u5730\u4ecb\u7ecd&nbsp;<a href=\"https:\/\/huggingface.co\/collections\/HuggingFaceTB\/smollm-models-6695016cad7167254ce15966\">SmolLM\uff0c<\/a>\u8fd9\u662f\u4e00\u7cfb\u5217\u6700\u5148\u8fdb\u7684\u5c0f\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u6709\u4e09\u79cd\u5927\u5c0f\uff1a135M\u3001360M \u548c 1.7B \u53c2\u6570\u3002\u8fd9\u4e9b\u6a21\u578b\u5efa\u7acb\u5728\u7cbe\u5fc3\u7b56\u5212\u7684\u9ad8\u8d28\u91cf\u8bad\u7ec3\u8bed\u6599\u5e93\u4e4b\u4e0a\uff0c\u6211\u4eec\u5c06\u5176\u4f5c\u4e3a&nbsp;<a href=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/smollm-corpus\">SmolLM-Corpus<\/a>&nbsp;\u53d1\u5e03\u3002Smollm \u8bed\u6599\u5e93\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Cosmopedia v2<\/strong>\uff1a\u7531 Mixtral \u751f\u6210\u7684\u5408\u6210\u6559\u79d1\u4e66\u548c\u6545\u4e8b\u96c6\uff0828B \u4ee3\u5e01\uff09<\/li>\n\n\n\n<li><strong>Python-Edu<\/strong>\uff1a\u6765\u81ea The Stack \u7684\u6559\u80b2 Python \u793a\u4f8b\uff084B \u4ee4\u724c\uff09<\/li>\n\n\n\n<li><strong>FineWeb-Edu\uff08\u91cd\u590d\u6570\u636e\u5220\u9664\uff09\uff1a<\/strong>\u6765\u81ea FineWeb \u7684\u6559\u80b2\u7f51\u7edc\u793a\u4f8b\uff08220B \u4ee4\u724c\uff09<\/li>\n<\/ul>\n\n\n\n<p>\u6211\u4eec\u7684\u8bc4\u4f30\u8868\u660e\uff0cSmolLM\u6a21\u578b\u5728\u5404\u79cd\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5728\u5176\u5927\u5c0f\u7c7b\u522b\u4e2d\u4f18\u4e8e\u5176\u4ed6\u6a21\u578b\uff0c\u6d4b\u8bd5\u4e86\u5e38\u8bc6\u63a8\u7406\u548c\u4e16\u754c\u77e5\u8bc6\u3002\u5728\u8fd9\u7bc7\u535a\u6587\u4e2d\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u8bad\u7ec3\u8bed\u6599\u5e93\u4e2d\u6bcf\u4e2a\u5b50\u96c6\u7684\u7b56\u5c55\uff0c\u7136\u540e\u8ba8\u8bba SmolLM \u6a21\u578b\u7684\u8bad\u7ec3\u548c\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled.png\" alt=\"\"><br><em>\u5728\u4e0d\u540c\u7684\u63a8\u7406\u548c\u5e38\u8bc6\u57fa\u51c6\u4e0a\u8bc4\u4f30 SmolLM \u6a21\u578b\u3002<\/em><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#data-curation\"><\/a>\u6570\u636e\u7ba1\u7406<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#from-cosmopedia-v1-to-v2\"><\/a>\u4ece Cosmopedia v1 \u5230 v2<\/h3>\n\n\n\n<p><a href=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/smollm-corpus\">Cosmopedia v2<\/a>&nbsp;\u662f Cosmopedia \u7684\u589e\u5f3a\u7248\u672c\uff0cCosmopedia \u662f\u6700\u5927\u7684\u9884\u8bad\u7ec3\u5408\u6210\u6570\u636e\u96c6\uff0c\u7531 Mixtral-8x7B-Instruct-v0.1 \u751f\u6210\u7684\u8d85\u8fc7 3000 \u4e07\u672c\u6559\u79d1\u4e66\u3001\u535a\u5ba2\u6587\u7ae0\u548c\u6545\u4e8b\u7ec4\u6210\u3002\u5927\u591a\u6570\u6837\u672c\u662f\u901a\u8fc7\u63d0\u793a\u6a21\u578b\u4f7f\u7528\u79f0\u4e3a\u201c\u79cd\u5b50\u6837\u672c\u201d\u7684\u7f51\u9875\u751f\u6210\u7279\u5b9a\u4e3b\u9898\u7684\u5185\u5bb9\u800c\u751f\u6210\u7684\uff0c\u5982\u56fe 1 \u6240\u793a\u3002\u6211\u4eec\u4f7f\u7528 Web \u793a\u4f8b\u6765\u589e\u52a0\u591a\u6837\u6027\u5e76\u6269\u5927\u63d0\u793a\u8303\u56f4\u3002\u60a8\u53ef\u4ee5\u5728\u8fd9\u7bc7<a href=\"https:\/\/huggingface.co\/blog\/cosmopedia\">\u535a\u6587<\/a>\u4e2d\u627e\u5230\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%201.png\" alt=\"\"><br><em>\u56fe 1.Cosmopedia \u63d0\u793a\u793a\u4f8b\u3002<\/em><\/p>\n\n\n\n<p>\u4e3a\u4e86\u6539\u8fdb v2 \u4e2d\u7684\u6570\u636e\u96c6\uff0c\u6211\u4eec\u5c1d\u8bd5\u4e86\u4e24\u79cd\u7b56\u7565\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528\u529f\u80fd\u66f4\u5f3a\u5927\u7684\u6a21\u578b\u548c\u76f8\u540c\u7684\u63d0\u793a<\/li>\n\n\n\n<li>\u4f18\u5316\u63d0\u793a\u672c\u8eab<\/li>\n<\/ul>\n\n\n\n<p>\u5bf9\u4e8e\u7b2c\u4e00\u79cd\u7b56\u7565\uff0c\u6211\u4eec\u5c1d\u8bd5\u4e86 llama3-70B-Instruct\u3001Mixtral-8x22B-Instruct-v0.1 \u548c Qwen1.5-72B-Chat\uff0c\u4f46\u5728\u8fd9\u4e9b\u66ff\u4ee3\u65b9\u6848\u751f\u6210\u7684\u6559\u79d1\u4e66\u4e0a\u8bad\u7ec3\u6a21\u578b\u65f6\u53d1\u73b0\u6ca1\u6709\u663e\u7740\u6539\u8fdb\u3002\u56e0\u6b64\uff0c\u5728\u672c\u8282\u7684\u5176\u4f59\u90e8\u5206\uff0c\u6211\u4eec\u5c06\u91cd\u70b9\u4ecb\u7ecd\u7b2c\u4e8c\u79cd\u7b56\u7565\uff1a\u6211\u4eec\u5982\u4f55\u6539\u8fdb\u63d0\u793a\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#the-search-for-better-topics-and-seed-samples\"><\/a>\u5bfb\u627e\u66f4\u597d\u7684\u4e3b\u9898\u548c\u79cd\u5b50\u6837\u672c<\/h4>\n\n\n\n<p>\u6bcf\u4e2a\u63d0\u793a\u90fd\u6709\u4e09\u4e2a\u4e3b\u8981\u7ec4\u4ef6\uff1a\u4e3b\u9898\u3001\u79cd\u5b50\u793a\u4f8b\u548c\u751f\u6210\u6837\u5f0f\uff0c\u7528\u4e8e\u6307\u5b9a\u76ee\u6807\u53d7\u4f17\u548c\u6211\u4eec\u5e0c\u671b\u6a21\u578b\u751f\u6210\u7684\u5185\u5bb9\u7c7b\u578b\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u786e\u4fdd\u4e16\u4ee3\u4e00\u81f4\uff0c\u6211\u4eec\u9700\u8981\u4e0e\u7ed9\u5b9a\u4e3b\u9898\u5bc6\u5207\u76f8\u5173\u7684\u79cd\u5b50\u6837\u672c\u3002\u5728 Cosmopedia v1 \u4e2d\uff0c\u6211\u4eec\u5bf9 FineWeb \u793a\u4f8b\u8fd0\u884c\u4e86\u805a\u7c7b\u5206\u6790\uff0c\u4ee5\u8bc6\u522b\u4e3b\u9898\u548c\u76f8\u5e94\u7684 Web \u793a\u4f8b\uff0c\u5982\u56fe 2 \u6240\u793a\u3002\u6b64\u65b9\u6cd5\u6709\u4e24\u4e2a\u4e3b\u8981\u5c40\u9650\u6027\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u4e3b\u9898\u5217\u8868\u53cd\u6620\u4e86 web\/FineWeb \u96c6\u7fa4\uff0c\u867d\u7136\u5f88\u5168\u9762\uff0c\u4f46\u53ef\u80fd\u4f1a\u9650\u5236\u6211\u4eec\u5bf9\u4e3b\u9898\u7684\u63a7\u5236\u3002<\/li>\n\n\n\n<li>\u6bcf\u4e2a\u96c6\u7fa4\u4e2d\u7684 Web \u6837\u672c\u4e0d\u4f1a\u8fdb\u4e00\u6b65\u7b5b\u9009\uff0c\u53ef\u80fd\u5305\u62ec\u4e00\u4e9b\u4f4e\u8d28\u91cf\u7684\u6837\u672c\u3002<\/li>\n<\/ol>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%202.png\" alt=\"\"><br><em>\u56fe2.FineWeb \u96c6\u7fa4\u3002<\/em><\/p>\n\n\n\n<p>\u5728 v2 \u4e2d\uff0c\u6211\u4eec\u6ca1\u6709\u91c7\u7528\u8fd9\u79cd\u65e0\u76d1\u7763\u805a\u7c7b\u65b9\u6cd5\uff0c\u800c\u662f\u4f7f\u7528 BISAC \u56fe\u4e66\u5206\u7c7b\u4ece\u9884\u5b9a\u4e49\u7684 34,000 \u4e2a\u4e3b\u9898\u5217\u8868\u5f00\u59cb\uff0c<a href=\"https:\/\/www.bisg.org\/complete-bisac-subject-headings-list\">BISAC \u56fe\u4e66\u5206\u7c7b<\/a>\u662f\u4e00\u79cd\u7528\u4e8e\u6309\u4e3b\u9898\u5bf9\u4e66\u7c4d\u8fdb\u884c\u5206\u7c7b\u7684\u6807\u51c6\uff0c\u65e2\u5168\u9762\u53c8\u6ce8\u91cd\u6559\u80b2\u3002\u6211\u4eec\u4ece\u5c5e\u4e8e 51 \u4e2a\u7c7b\u522b\u7684 5,000 \u4e2a\u4e3b\u9898\u5f00\u59cb\uff0c\u5e76\u8981\u6c42 Mixtral \u4e3a\u67d0\u4e9b\u4e3b\u9898\u751f\u6210\u5b50\u4e3b\u9898\u3002\u4ee5\u4e0b\u662f\u6bcf\u4e2a\u7c7b\u522b\u4e2d\u5b50\u4e3b\u9898\u7684\u6700\u7ec8\u5206\u5e03\uff1a<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%203.png\" alt=\"\"><br><em>\u56fe3.\u6bcf\u4e2a\u7528\u4e8e\u63d0\u793a\u7684\u9876\u7ea7\u7c7b\u522b\u7684\u4e3b\u9898\u5206\u5e03\u3002<\/em><\/p>\n\n\n\n<p>\u5b9a\u4e49\u4e3b\u9898\u540e\uff0c\u6211\u4eec\u4ecd\u7136\u9700\u8981\u627e\u5230\u4e0e\u5b83\u4eec\u76f8\u5173\u7684\u7f51\u9875\u3002\u5c31\u50cf\u4f7f\u7528\u641c\u7d22\u5f15\u64ce\u67e5\u627e\u7279\u5b9a\u4e3b\u9898\u7684\u5185\u5bb9\u4e00\u6837\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86\u4e00\u4e2a\u641c\u7d22\u5de5\u5177\u6765\u68c0\u7d22\u6bcf\u4e2a\u4e3b\u9898\u6700\u76f8\u5173\u7684\u9875\u9762\u3002\u6211\u4eec\u4f7f\u7528 BISAC \u7c7b\u522b\u53ca\u5176\u5b50\u4e3b\u9898\u4f5c\u4e3a FineWeb CC-MAIN-2024-10 \u548c CC-MAIN-2023-50 \u8f6c\u50a8\u7684\u67e5\u8be2\u6765\u8fd0\u884c\u6b64\u5de5\u5177\uff0c\u8fd9\u4e9b\u8f6c\u50a8\u603b\u5171\u5305\u542b\u8d85\u8fc7 5.2 \u4ebf\u4e2a\u6837\u672c\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u67e5\u8be2\uff0c\u6211\u4eec\u68c0\u7d22\u4e86 1,000 \u4e2a\u9875\u9762\uff0c\u786e\u4fdd\u6211\u4eec\u53ea\u68c0\u7d22\u5230\u6700\u76f8\u5173\u7684\u5185\u5bb9\u3002<a href=\"https:\/\/github.com\/huggingface\/cosmopedia\/tree\/main\/fulltext_search\">\u6b64\u5904<\/a>\u63d0\u4f9b\u4e86\u7528\u4e8e\u90e8\u7f72\u548c\u8fd0\u884c\u641c\u7d22\u5de5\u5177\u7684\u4ee3\u7801\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u7f16\u8bd1\u4e86 34,000 \u4e2a\u4e3b\u9898\u7684 3400 \u4e07\u4e2a\u7f51\u9875\u3002\u4e0b\u4e00\u6b65\u662f\u786e\u5b9a\u54ea\u79cd\u4e16\u4ee3\u98ce\u683c\u6548\u679c\u6700\u597d\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%204.png\" alt=\"\"><br><em>\u56fe4.\u4e3b\u9898\u53ca\u5176\u5728\u201c\u533b\u7597\u201d\u7c7b\u522b\u4e2d\u68c0\u7d22\u5230\u7684\u6837\u672c\u3002<\/em><\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#generation-style\"><\/a>\u751f\u6210\u98ce\u683c<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u786e\u5b9a\u6700\u6709\u6548\u7684\u751f\u6210\u65b9\u5f0f\uff0c\u6211\u4eec\u901a\u8fc7\u5728 Cosmopedia v1 \u4e0d\u540c\u5b50\u96c6\u7684 8B \u4ee3\u5e01\u4e0a\u8bad\u7ec3 1.8B \u6a21\u578b\u6765\u8fdb\u884c\u6d88\u878d\u7814\u7a76\u3002\u5bf9\u4e8e\u65b0\u751f\u6210\u7684\u6570\u636e\uff0c\u6211\u4eec\u53ea\u751f\u6210\u4e86 2B \u4ee4\u724c\u5e76\u8bad\u7ec3\u4e86 4 \u4e2a epoch \u4ee5\u8282\u7701\u65f6\u95f4\uff08\u4f7f\u7528 Mixtral \u751f\u6210 2B \u4ee4\u724c\u5927\u7ea6\u9700\u8981 1000 \u4e2a GPU \u5c0f\u65f6\uff09\u3002\u6211\u4eec\u4f7f\u7528\u4e86\u4e0e<a href=\"https:\/\/huggingface.co\/spaces\/HuggingFaceFW\/blogpost-fineweb-v1\">FineWeb\u6d88\u878d\u6a21\u578b<\/a>\u76f8\u540c\u7684\u8bad\u7ec3\u548c\u8bc4\u4f30\u8bbe\u7f6e\u3002\u6211\u4eec\u7528\u4e24\u4e2a\u4e0d\u540c\u7684\u79cd\u5b50\u5bf9\u6bcf\u4e2a\u5b9e\u9a8c\u8fdb\u884c\u4e86\u4e24\u6b21\uff0c\u5e76\u5e73\u5747\u4e86\u4e24\u6b21\u8fd0\u884c\u4e4b\u95f4\u7684\u5206\u6570\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u6bd4\u8f83\u4e86 Cosmopedia v1 \u7684\u4ee5\u4e0b\u5b50\u96c6\u7684\u6027\u80fd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7f51\u7edc\u6559\u79d1\u4e66\u5b50\u96c6<\/li>\n\n\n\n<li>\u6545\u4e8b\u5b50\u96c6<\/li>\n\n\n\n<li>Stanford &amp; OpenStax \u5b50\u96c6<\/li>\n<\/ul>\n\n\n\n<p>\u6211\u4eec\u53d1\u73b0\uff0c\u4e0e\u57fa\u4e8e\u7f51\u7edc\u7684\u6559\u79d1\u4e66\u76f8\u6bd4\uff0c\u57fa\u4e8e\u65af\u5766\u798f\u5927\u5b66\u548c OpenStax \u7b49\u7cbe\u9009\u6765\u6e90\u7684\u4e3b\u9898\u548c\u79cd\u5b50\u6837\u672c\u7684\u6559\u79d1\u4e66\u63d0\u4f9b\u4e86\u6700\u4f73\u7684\u6574\u4f53\u6027\u80fd\uff0c\u4ece\u800c\u83b7\u5f97\u4e86 MMLU \u548c ARC \u57fa\u51c6\u3002\u6545\u4e8b\u4f3c\u4e4e\u6709\u52a9\u4e8e\u5efa\u7acb\u5e38\u8bc6\u57fa\u51c6\u3002\u5728 v2 \u4e2d\u5b9e\u73b0\u65b0\u4e3b\u9898\u548c\u79cd\u5b50\u6837\u672c\u68c0\u7d22\u65b9\u6cd5\u540e\uff0c\u6211\u4eec\u80fd\u591f\u4f7f\u7528 Web \u79cd\u5b50\u5339\u914d\u7cbe\u9009\u6e90\u7684\u6027\u80fd\uff0c\u4ece\u800c\u786e\u8ba4\u65b0\u63d0\u793a\u7684\u8d28\u91cf\u3002<\/p>\n\n\n\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u63a2\u8ba8\u4e86\u54ea\u79cd\u53d7\u4f17\u98ce\u683c\u6548\u679c\u6700\u597d\u3002\u6211\u4eec\u4f7f\u7528\u76f8\u540c\u7684\u7f51\u7edc\u6559\u79d1\u4e66\u63d0\u793a\u751f\u6210\u6559\u79d1\u4e66\uff0c\u4f46\u9488\u5bf9\u4e24\u4e2a\u4e0d\u540c\u7684\u53d7\u4f17\uff1a\u4e2d\u5b66\u751f\u548c\u5927\u5b66\u751f\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u5728\u4e3b\u8981\u9488\u5bf9\u4e2d\u5b66\u751f\u7684\u6559\u79d1\u4e66\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u5728\u9664MMLU\u4ee5\u5916\u7684\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u90fd\u83b7\u5f97\u4e86\u6700\u9ad8\u5206\u3002\u8fd9\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u4e8b\u5b9e\u6765\u89e3\u91ca\uff1a\u5176\u4e2d\u5927\u591a\u6570\u6d4b\u8bd5\u57fa\u672c\u5e38\u8bc6\u548c\u521d\u7ea7\u5230\u4e2d\u7ea7\u79d1\u5b66\u77e5\u8bc6\uff0c\u800c MMLU \u5305\u542b\u4e00\u4e9b\u9700\u8981\u9ad8\u7ea7\u77e5\u8bc6\u548c\u4e13\u4e1a\u77e5\u8bc6\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%205.png\" alt=\"\"><br><em>\u9488\u5bf9\u4e0d\u540c\u53d7\u4f17\u7684\u6559\u79d1\u4e66\u8bc4\u4f30\u3002<\/em><\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%206.png\" alt=\"\"><br><em>\u9488\u5bf9\u4e0d\u540c\u53d7\u4f17\u7684\u6559\u79d1\u4e66\u8bc4\u4f30\u3002<\/em><\/p>\n\n\n\n<p>\u5bf9\u4e8e v2\uff0c\u6211\u4eec\u51b3\u5b9a\u4e3a\u4e2d\u5b66\u751f\u751f\u6210 40% \u7684\u5185\u5bb9\uff0c\u4e3a\u5927\u5b66\u751f\u751f\u6210 30% \u7684\u5185\u5bb9\uff0c\u4ee5\u53ca 30% \u4f5c\u4e3a\u5176\u4ed6\u53d7\u4f17\u548c\u98ce\u683c\u7684\u6df7\u5408\uff0c\u5305\u62ec\u6211\u4eec\u4ece Cosmopedia v1 \u501f\u6765\u7684\u5b50\u96c6\uff0c\u4f8b\u5982\u6545\u4e8b\u548c\u57fa\u4e8e\u65af\u5766\u798f\u8bfe\u7a0b\u7684\u6559\u79d1\u4e66\u3002\u6b64\u5916\uff0c\u6211\u4eec\u57fa\u4e8e\u6765\u81ea AutoMathText \u6570\u636e\u96c6\u7684 Python \u79cd\u5b50\u6837\u672c\u751f\u6210\u4e86 1B \u4ee3\u7801\u6559\u79d1\u4e66\u3002<\/p>\n\n\n\n<p>\u6700\u7ec8\uff0c\u6211\u4eec\u5236\u4f5c\u4e86 3900 \u4e07\u4efd\u5408\u6210\u6587\u6863\uff0c\u5305\u62ec 28B \u4e2a\u6559\u79d1\u4e66\u3001\u6545\u4e8b\u3001\u6587\u7ae0\u548c\u4ee3\u7801\uff0c\u53d7\u4f17\u8303\u56f4\u5e7f\u6cdb\uff0c\u4e3b\u9898\u8d85\u8fc7 34,000 \u4e2a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#fineweb-edu\"><\/a>FineWeb-\u6559\u80b2<\/h3>\n\n\n\n<p>FineWeb-Edu \u662f\u6211\u4eec\u51e0\u4e2a\u6708\u524d\u53d1\u5e03\u7684\u4e00\u4e2a\u6570\u636e\u96c6\uff0c\u5176\u4e2d\u5305\u542b FineWeb&nbsp;<a href=\"https:\/\/huggingface.co\/spaces\/HuggingFaceFW\/blogpost-fineweb-v1\">\u7684\u6280\u672f\u62a5\u544a\u3002<\/a>\u5b83\u7531\u4ece\ud83c\udf77 FineWeb\u6570\u636e\u96c6\u4e2d\u8fc7\u6ee4\u51fa\u6765\u7684<strong>1.3T\u6559\u80b2\u7f51\u9875\u4ee4\u724c<\/strong>\u7ec4\u6210\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u4f7f\u7528 Llama3-70B-Inform \u751f\u6210\u7684\u6ce8\u91ca\u5f00\u53d1\u4e86\u4e00\u4e2a<a href=\"https:\/\/huggingface.co\/HuggingFaceFW\/fineweb-edu-classifier\"><strong>\u6559\u80b2\u8d28\u91cf\u5206\u7c7b\u5668<\/strong><\/a>\u3002\u7136\u540e\uff0c\u6211\u4eec\u4f7f\u7528\u6b64\u5206\u7c7b\u5668\u4ec5\u4fdd\u7559 FineWeb \u4e2d\u6700\u5177\u6559\u80b2\u610f\u4e49\u7684\u7f51\u9875\u3002FineWeb-Edu \u5728\u6d41\u884c\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u4f18\u4e8e FineWeb\uff0c\u5e76\u663e\u793a\u4e86\u5728\u5408\u6210\u6570\u636e\u4e0a\u8bad\u7ec3\u7684\u5206\u7c7b\u5668\u7684\u5f3a\u5927\u529f\u80fd\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%207.png\" alt=\"\"><br><em>FineWeb-Edu \u4e0e\u5176\u4ed6\u5f00\u653e Web \u6570\u636e\u96c6\u7684\u6bd4\u8f83\u3002<\/em><\/p>\n\n\n\n<p>\u5728 Smollm-Corpus \u4e2d\uff0c\u6211\u4eec\u5305\u542b\u6765\u81ea FineWeb \u7684 220B \u91cd\u590d\u6570\u636e\u5220\u9664\u4ee4\u724c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#stack-edu-python\"><\/a>\u5806\u6808-edu-python\u7684<\/h3>\n\n\n\n<p>\u6211\u4eec\u5c06 FineWeb-Edu \u7684\u76f8\u540c\u7406\u5ff5\u5e94\u7528\u5230\u4ee3\u7801\u4e2d\u3002\u6211\u4eec\u4f7f\u7528 Llama3 \u6ce8\u91ca\u4e86 The Stack \u6570\u636e\u96c6\u4e2d\u7684 500,000 \u4e2a python \u6837\u672c\uff0c\u5e76\u4f7f\u7528\u4e0e FineWeb-Edu \u5206\u7c7b\u5668\u76f8\u540c\u7684\u914d\u65b9\u6765\u8bad\u7ec3<a href=\"https:\/\/huggingface.co\/HuggingFaceTB\/python-edu-scorer\">\u6559\u80b2\u5206\u7c7b\u5668<\/a>\u3002\u7136\u540e\uff0c\u6211\u4eec\u5c06\u8fd9\u4e2a\u5206\u7c7b\u5668\u5e94\u7528\u4e8e StarCoder \u6a21\u578b\u8bad\u7ec3\u8bed\u6599\u5e93\u7684 Python \u5b50\u96c6\u3002\u4ece\u53ef\u7528\u7684 40B Python \u4ee3\u5e01\u4e2d\uff0c\u6211\u4eec\u53ea\u4fdd\u7559\u4e86\u5f97\u5206\u4e3a 4 \u6216\u66f4\u9ad8\u7684\u6837\u672c\uff0c\u4ece\u800c\u5f62\u6210\u4e86\u4e00\u4e2a\u7cbe\u7ec6\u7684 4B \u4ee3\u5e01\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n<p>\u4e0b\u56fe\u5c06 Python-Edu \u4e0e\u672a\u7ecf\u7b5b\u9009\u7684 Python \u4ee3\u7801\u4ee5\u53ca\u4f7f\u7528\u4e0d\u592a\u4e25\u683c\u7684\u9608\u503c 3 \u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u5728 Python-Edu \u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u6bd4\u5728\u672a\u8fc7\u6ee4\u7684 Python \u4ee3\u7801\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u6536\u655b\u901f\u5ea6\u5feb 3 \u500d\u4ee5\u4e0a\uff0c\u4ec5\u4f7f\u7528 12B \u6807\u8bb0\u540e\u5c31\u5b9e\u73b0\u4e86 16% \u7684pass@1\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%208.png\" alt=\"\"><br><em>Python-Edu \u4e0e\u672a\u8fc7\u6ee4\u7684 Python \u4ee3\u7801\u7684\u6bd4\u8f83\u3002<\/em><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#training\"><\/a>\u8bad\u7ec3<\/h2>\n\n\n\n<p>SmolLM \u6a21\u578b\u6709\u4e09\u79cd\u5c3a\u5bf8\u53ef\u4f9b\u9009\u62e9\uff0c\u5e76\u6839\u636e\u4ee5\u4e0b\u6570\u636e\u7ec4\u5408\u8fdb\u884c\u8bad\u7ec3\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>135M \u548c 360M \u6a21\u578b\uff0c\u6bcf\u4e2a\u6a21\u578b\u90fd\u4f7f\u7528&nbsp;<a href=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/smollm-corpus\">Smollm-Corpus<\/a>&nbsp;\u7684 600B \u4ee3\u5e01\u8fdb\u884c\u8bad\u7ec3<\/li>\n\n\n\n<li>1.7B \u6a21\u578b\uff0c\u5728 Smollm-Corpus \u7684 1T \u4ee3\u5e01\u4e0a\u8bad\u7ec3<\/li>\n<\/ul>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%209.png\" alt=\"\"><br><em>SmolLM \u6a21\u578b\u7684\u8bad\u7ec3\u6df7\u5408\u7269\u3002<\/em><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#hyperparameters-choice\"><\/a>\u8d85\u53c2\u6570\u9009\u62e9<\/h3>\n\n\n\n<p>\u6211\u4eec\u4f7f\u7528\u68af\u5f62\u5b66\u4e60\u901f\u7387\u8c03\u5ea6\u5668\uff0c\u51b7\u5374\u9636\u6bb5\u7b49\u4e8e\u603b\u8bad\u7ec3\u65f6\u95f4\u7684 20%\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e2a\u65f6\u95f4\u8868\u7684\u539f\u59cb\u5b9e\u9a8c\u662f\u5728\u8f83\u5c0f\u7684\u89c4\u6a21\u4e0b\u8fdb\u884c\u7684\uff0c\u6211\u4eec\u5df2\u7ecf\u5c06\u5176\u8c03\u6574\u4e3a\u66f4\u5927\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e 135M \u548c 360M \u53c2\u6570\u6a21\u578b\u7684\u67b6\u6784\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u7c7b\u4f3c\u4e8e&nbsp;<a href=\"https:\/\/arxiv.org\/abs\/2402.14905\">MobileLLM<\/a>&nbsp;\u7684\u8bbe\u8ba1\uff0c\u7ed3\u5408\u4e86\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b \uff08GQA\uff09 \u5e76\u4f18\u5148\u8003\u8651\u6df1\u5ea6\u800c\u4e0d\u662f\u5bbd\u5ea6\u30021.7B \u53c2\u6570\u6a21\u578b\u4f7f\u7528\u66f4\u4f20\u7edf\u7684\u67b6\u6784\u3002\u5bf9\u4e8e\u8fd9\u4e09\u4e2a\u6a21\u578b\uff0c\u6211\u4eec\u4f7f\u7528\u5d4c\u5165\u7ed1\u5b9a\u548c 2048 \u4e2a\u4ee4\u724c\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002\u901a\u8fc7\u4e00\u4e9b\u957f\u65f6\u95f4\u7684\u4e0a\u4e0b\u6587\u5fae\u8c03\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u6269\u5c55\u6b64\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002<\/p>\n\n\n\n<p>\u6bcf\u79cd\u578b\u53f7\u5927\u5c0f\u7684\u8be6\u7ec6\u67b6\u6784\u89c4\u8303\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%2010.png\" alt=\"\"><br><em>SmolLM \u6a21\u578b\u7684\u67b6\u6784\u8be6\u7ec6\u4fe1\u606f\u3002<\/em><\/p>\n\n\n\n<p>\u6211\u4eec\u4f7f\u7528\u4e86\u5728 Smollm \u8bed\u6599\u5e93\u4e0a\u8bad\u7ec3\u7684\u5206\u8bcd\u5668\uff0c\u8bcd\u6c47\u91cf\u4e3a 49152\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#experiments\"><\/a>\u5b9e\u9a8c<\/h3>\n\n\n\n<p>\u4f7f\u7528\u68af\u5f62\u8c03\u5ea6\u5668\u7684\u4e00\u4e2a\u4f18\u70b9\u662f\uff0c\u5b83\u53ef\u4ee5\u51cf\u5c11\u6267\u884c\u7f29\u653e\u5b9a\u5f8b\u5b9e\u9a8c\u6240\u9700\u7684\u65f6\u95f4\uff0c\u5982<a href=\"https:\/\/arxiv.org\/pdf\/2405.18392\">H\u00e4gele\u7b49\u4eba<\/a>\u6240\u793a\u3002\u6211\u4eec\u901a\u8fc7\u5bf9\u6700\u5c0f\u578b\u53f7 SmolLM-125M \u7684\u5c0f\u578b\u7f29\u653e\u5b9a\u5f8b\u7814\u7a76\u6765\u8bf4\u660e\u8fd9\u4e00\u70b9\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u968f\u7740\u8bad\u7ec3\u65f6\u95f4\u7684\u5ef6\u957f\uff0c\u6027\u80fd\u4f1a\u7ee7\u7eed\u63d0\u9ad8\uff0c\u751a\u81f3\u8d85\u8fc7\u9f99\u732b\u7684\u6700\u4f73\u70b9\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u51b3\u5b9a\u5728 1 \u4e07\u4ebf\u4e2a\u4ee3\u5e01\u4e0a\u4f7f\u7528 1.7B \u6a21\u578b\uff0c\u5728 600B \u4ee3\u5e01\u4e0a\u4f7f\u7528 135M \u548c 360M \u6a21\u578b\uff0c\u56e0\u4e3a\u5728\u8fd9\u4e9b\u8f83\u5c0f\u6a21\u578b\u7684\u4e00\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c400B \u4ee3\u5e01\u540e\u7684\u6027\u80fd\u63d0\u5347\u5f00\u59cb\u653e\u7f13\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%2011.png\" alt=\"\"><br><em>\u8bc4\u4f30\u5728\u4e0d\u540c\u6570\u91cf\u7684\u4ee4\u724c\u4e0a\u8bad\u7ec3\u7684 125M SmolLM \u6a21\u578b\u3002<\/em><\/p>\n\n\n\n<p>\u6211\u4eec\u5c1d\u8bd5\u5728\u51b7\u5374\u9636\u6bb5\u6dfb\u52a0 instruct \u6570\u636e\u96c6\u5e76\u5bf9\u7cbe\u9009\u7684 Cosmopedia \u5b50\u96c6\u8fdb\u884c\u4e0a\u91c7\u6837\uff0c\u4f46\u6ca1\u6709\u53d1\u73b0\u4efb\u4f55\u663e\u7740\u6539\u8fdb\u3002\u8fd9\u53ef\u80fd\u662f\u56e0\u4e3a\u4e3b\u8981\u6570\u636e\u7ec4\u5408\u5df2\u7ecf\u5177\u6709\u9ad8\u8d28\u91cf\uff0c\u4ece\u800c\u9650\u5236\u4e86\u8fd9\u4e9b\u66f4\u6539\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8ddf\u8e2a\u6211\u4eec\u7684\u8bad\u7ec3\u8fdb\u5ea6\uff0c\u6211\u4eec\u6bcf 2B \u4e2a\u4ee4\u724c\u8bc4\u4f30\u4e24\u4e2a\u6700\u5c0f\u7684\u6a21\u578b\u3002\u4e0b\u56fe\u663e\u793a\u4e86\u5b83\u4eec\u5728\u51e0\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u8868\u73b0\uff1a<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%2012.png\" alt=\"\"><br><em>SmolLM-135M \u548c SmolLM-360M \u5728\u4e0d\u540c\u57fa\u51c6\u4e0a\u7684\u4e2d\u7ea7\u8bc4\u4f30\u3002<\/em><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#evaluation\"><\/a>\u8bc4\u4f30<\/h2>\n\n\n\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u8bc4\u4f30\u4e86 SmolLM \u6a21\u578b\u5728\u4e0d\u540c\u53c2\u6570\u5927\u5c0f\u4e0b\u7684\u6027\u80fd\uff0c\u5e76\u5c06\u5b83\u4eec\u4e0e\u5404\u81ea\u7c7b\u522b\u4e2d\u7684\u6700\u4f73\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\u3002\u6211\u4eec\u6839\u636e\u4e00\u5957\u4e0d\u540c\u7684\u57fa\u51c6\u8fdb\u884c\u8bc4\u4f30\uff0c\u6d4b\u8bd5\u5e38\u8bc6\u3001\u63a8\u7406\u548c\u4e16\u754c\u77e5\u8bc6\u3002\u6211\u4eec\u5bf9\u6240\u6709\u4f7f\u7528\u6b64<a href=\"https:\/\/github.com\/huggingface\/cosmopedia\/tree\/main\/evaluation\">\u8bbe\u7f6e<\/a>\u548c\u5e93\u7684\u6a21\u578b\u4f7f\u7528\u76f8\u540c\u7684\u8bc4\u4f30\u8bbe\u7f6e\u3002\u5bf9\u4e8e HumanEval\uff0c\u6211\u4eec\u4f7f\u7528 [bigcode-evaluation-harness]\uff08\u6211\u4eec\u4f7f\u7528 temperature 0.2\uff0ctop-p 0.95 \u548c 20 \u4e2a\u6837\u672c\uff09\u548c We use temperature 0.2\uff0ctop-p 0.95 \u548c 20 \u4e2a\u6837\u672c\u3002\u5bf9\u4e8e\u4e0d\u516c\u5f00\u7684 MobileLLM\uff0c\u6211\u4eec\u5c3d\u53ef\u80fd\u4f7f\u7528\u8bba\u6587\u4e2d\u62a5\u544a\u7684\u6570\u5b57\u3002<code>lighteval<\/code><\/p>\n\n\n\n<p>\u6211\u4eec\u53d1\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>SmolLM-135M \u7684\u6027\u80fd\u4f18\u4e8e\u5f53\u524d\u53c2\u6570\u5c0f\u4e8e 200M \u7684\u6700\u4f73\u578b\u53f7 MobileLM-125M\uff0c\u5c3d\u7ba1\u4e0e MobileLM \u7684 1T \u4ee4\u724c\u76f8\u6bd4\uff0c\u5b83\u53ea\u63a5\u53d7\u4e86 600B \u4ee4\u724c\u7684\u8bad\u7ec3\u3002<\/li>\n\n\n\n<li>\u4e0e MobileLM-350M \u548c Qwen2-500M \u76f8\u6bd4\uff0cSmolLM**-**360M \u7684\u6027\u80fd\u4f18\u4e8e\u6240\u6709\u53c2\u6570\u5c0f\u4e8e 500M \u7684\u6a21\u578b\uff0c\u5c3d\u7ba1\u53c2\u6570\u8f83\u5c11\u4e14\u8bad\u7ec3\u7684\u4ee4\u724c\u4e0d\u5230\u4e00\u4e07\u4ebf\u4e2a \uff08600B\uff09\u3002<\/li>\n\n\n\n<li>SmolLM-1.7B\u4f18\u4e8e\u6240\u6709\u5176\u4ed6\u53c2\u6570\u5c0f\u4e8e2B\u7684\u578b\u53f7\uff0c\u5305\u62ecMicrosoft\u7684Phi1.5\uff0cMobileLM-1.5B\u548cQwen2-1.5B\u3002<\/li>\n\n\n\n<li>SmolLM-1.7B \u5728 24 pass@1\u4e2d\u8868\u73b0\u51fa\u5f3a\u5927\u7684 Python \u7f16\u7801\u6027\u80fd\u3002\u6211\u4eec\u6ce8\u610f\u5230 Qwen2-1.5B \u7684\u8bc4\u4f30\u5206\u6570\u4e0e Qwen \u56e2\u961f\u62a5\u544a\u7684 31.1 pass@1\u4e0d\u540c\u3002\u6211\u4eec\u4f7f\u7528 0.2 \u7684\u6e29\u5ea6\uff0ctop-p 0.95 \u548c 20 \u4e2a\u6837\u54c1\u3002<\/li>\n<\/ul>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%2014.png\" alt=\"\"><br><em>SmolLM \u6a21\u578b\u4e0e\u5176\u4ed6 SLM \u7684\u6bd4\u8f83\u3002\u6211\u4eec\u5728\u540c\u4e00\u8bbe\u7f6e\u4e0a\u8bc4\u4f30\u6240\u6709\u6a21\u578b\uff0c\u4f46 MobieLLM \u9664\u5916\uff0c\u5b83\u4e0d\u662f\u516c\u5f00\u7684\u3002<\/em><\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/image.png\" alt=\"\"><br><em>\u5728HumanEval\u4e0a\u8bc4\u4f30SmolLM\u6a21\u578b\u3002<\/em><\/p>\n\n\n\n<p>\u6211\u4eec\u8fd8\u4f7f\u7528\u516c\u5f00\u53ef\u7528\u7684\u5bbd\u677e\u6307\u4ee4\u6570\u636e\u96c6\u5bf9\u6a21\u578b\u8fdb\u884c\u4e86\u6307\u4ee4\u8c03\u6574\u3002\u6211\u4eec\u5728&nbsp;<a href=\"https:\/\/huggingface.co\/datasets\/TIGER-Lab\/WebInstructSub\">WebInstructSub \u6570\u636e\u96c6<\/a>\u7684\u5bbd\u677e\u5b50\u96c6\u4e0a\u8bad\u7ec3\u4e86\u6240\u6709\u4e09\u4e2a\u6a21\u578b\uff0c\u5e76\u7ed3\u5408 StarCoder2-Self-OSS-Instruct\u3002\u5728\u6b64\u4e4b\u540e\uff0c\u6211\u4eec\u5bf9\u4e00\u4e2a\u65f6\u671f\u6267\u884c\u4e86 DPO\uff08\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff09\uff1a\u5bf9 135M \u548c 1.7B \u6a21\u578b\u4f7f\u7528&nbsp;<a href=\"https:\/\/huggingface.co\/datasets\/nvidia\/HelpSteer\">HelpSteer<\/a>\uff0c\u5bf9 360M \u6a21\u578b\u4f7f\u7528&nbsp;<a href=\"https:\/\/huggingface.co\/datasets\/argilla\/dpo-mix-7k\">argilla\/dpo-mix-7k<\/a>\u3002\u6211\u4eec\u9075\u5faa<a href=\"https:\/\/github.com\/huggingface\/alignment-handbook\/blob\/main\/recipes\/zephyr-7b-gemma\/README.md\">\u4e86\u5bf9\u9f50\u624b\u518c<\/a>\u4e2d Zephyr-Gemma \u914d\u65b9\u4e2d\u7684\u8bad\u7ec3\u53c2\u6570\uff0c\u4f46\u5c06 SFT\uff08\u76d1\u7763\u5fae\u8c03\uff09\u5b66\u4e60\u7387\u8c03\u6574\u4e3a 3e-4\u3002<\/p>\n\n\n\n<p>\u4e0b\u8868\u663e\u793a\u4e86 SmolLM-Instruct \u548c\u5176\u4ed6\u6a21\u578b\u5728 IFEval \u57fa\u51c6\u6d4b\u8bd5\uff08\u63d0\u793a\u4e25\u683c\u7cbe\u5ea6\uff09\u4e0a\u7684\u6027\u80fd\u3002Qwen2-1.5B-Instruct \u6a21\u578b\u5f97\u5206\u6700\u9ad8\uff0c\u4e3a 29.94\uff0cSmolLM-Instruct \u6a21\u578b\u5728\u6a21\u578b\u5927\u5c0f\u548c\u6027\u80fd\u4e4b\u95f4\u63d0\u4f9b\u4e86\u826f\u597d\u7684\u5e73\u8861\uff0c\u4ec5\u4f7f\u7528\u516c\u5f00\u53ef\u7528\u7684\u5bbd\u677e\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%2016.png\" alt=\"\"><br><em>\u5728 IFEval \u4e0a\u8bc4\u4f30 SmolLM-Instruct \u6a21\u578b\u3002<\/em><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#how-to-run-locally-\"><\/a>\u5982\u4f55\u5728\u672c\u5730\u8fd0\u884c\uff1f<\/h2>\n\n\n\n<p>\u6211\u4eec\u7684\u578b\u53f7\u8bbe\u8ba1\u4e3a\u5c0f\u578b\uff0c\u53ef\u4ee5\u5728\u5404\u79cd\u786c\u4ef6\u914d\u7f6e\u4e0a\u672c\u5730\u8fd0\u884c\u3002\u4f5c\u4e3a\u53c2\u8003\uff0ciPhone 15 \u6709 6GB \u7684 DRAM\uff0c\u800c iPhone 15 Pro \u6709 16GB\u3002\u8fd9\u4e9b\u5185\u5b58\u8981\u6c42\u4f7f\u6211\u4eec\u7684\u578b\u53f7\u9002\u5408\u90e8\u7f72\u5728\u4ece\u667a\u80fd\u624b\u673a\u5230\u7b14\u8bb0\u672c\u7535\u8111\u7684\u5404\u79cd\u8bbe\u5907\u4e0a\u3002\u6211\u4eec\u5bf9\u4e09\u79cd\u578b\u53f7\u5927\u5c0f\u7684\u5185\u5b58\u5360\u7528\u8fdb\u884c\u4e86\u57fa\u51c6\u6d4b\u8bd5\uff1a<\/p>\n\n\n\n<p><img decoding=\"async\" src=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/images\/resolve\/main\/Untitled%2013.png\" alt=\"\"><br><em>SmolLM \u6a21\u578b\u7684\u5185\u5b58\u5360\u7528\u3002<\/em><\/p>\n\n\n\n<p>\u9664\u4e86 transformers \u68c0\u67e5\u70b9\uff0c\u6211\u4eec\u8fd8\u53d1\u5e03\u4e86 ONNX \u68c0\u67e5\u70b9\uff0c\u5e76\u8ba1\u5212\u6dfb\u52a0\u4e0e \u517c\u5bb9\u7684 GGUF \u7248\u672c\u3002\u60a8\u53ef\u4ee5\u5728&nbsp;<a href=\"https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-135M-Instruct-WebGPU\">https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-135M-Instruct-WebGPU<\/a>&nbsp;\u548c&nbsp;<a href=\"https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-360M-Instruct-WebGPU\">https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-360M-Instruct-WebGPU<\/a>&nbsp;\u627e\u5230 WebGPU \u6f14\u793a SmolLM-135M \u548c Smol-LM360M\u3002<code>llama.cpp<\/code><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#conclusion\"><\/a>\u7ed3\u8bba<\/h2>\n\n\n\n<p>\u5728\u8fd9\u7bc7\u535a\u6587\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86 SmolLM \u6a21\u578b\uff0c\u8fd9\u662f\u4e00\u4e2a\u65b0\u7684\u6700\u5148\u8fdb\u7684\u5c0f\u578b LLM \u7cfb\u5217\u3002\u4ed6\u4eec\u8bc1\u660e\uff0c\u5c0f\u578b\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u5bf9\u9ad8\u8d28\u91cf\u6570\u636e\u96c6\u8fdb\u884c\u9ad8\u6548\u8bad\u7ec3\u6765\u5b9e\u73b0\u9ad8\u6027\u80fd\uff0c\u4ece\u800c\u5728\u5927\u5c0f\u548c\u6027\u80fd\u4e4b\u95f4\u63d0\u4f9b\u5f3a\u6709\u529b\u7684\u5e73\u8861\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><a href=\"https:\/\/huggingface.co\/blog\/smollm#resources\"><\/a>\u8d44\u6e90<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>SmolLM \u6a21\u578b\u96c6\u5408\uff1a&nbsp;<a href=\"https:\/\/huggingface.co\/collections\/HuggingFaceTB\/smollm-models-6695016cad7167254ce15966\">https:\/\/huggingface.co\/collections\/HuggingFaceTB\/smollm-models-6695016cad7167254ce15966<\/a><\/li>\n\n\n\n<li>SmolLM-Corpus \u6570\u636e\u96c6\uff1a<a href=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/smollm-corpus\">https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/smollm-corpus<\/a><\/li>\n\n\n\n<li>WebGPU \u6f14\u793a\uff1a<a href=\"https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-135M-Instruct-WebGPU\">https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-135M-Instruct-WebGPU<\/a>&nbsp;\u548c&nbsp;<a href=\"https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-360M-Instruct-WebGPU\">https:\/\/huggingface.co\/spaces\/HuggingFaceTB\/SmolLM-360M-Instruct-WebGPU<\/a><\/li>\n<\/ul>\n\n\n\n<p>\u6211\u4eec\u535a\u5ba2\u4e2d\u7684\u66f4\u591a\u6587\u7ae0<\/p>\n\n\n\n<p>\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/huggingface.co\/blog\/smollm\">SmolLM &#8211; \u901f\u5ea6\u6781\u5feb\uff0c\u529f\u80fd\u5f3a\u5927 (huggingface.co)<\/a><\/p>\n\n\n\n<p>\u6570\u636e\u96c6\uff1a<a href=\"https:\/\/huggingface.co\/datasets\/HuggingFaceTB\/smollm-corpus\">HuggingFaceTB\/smollm-corpus \u00b7 Datasets at Hugging Face<\/a><\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8fd9\u7bc7\u535a\u6587\u4ecb\u7ecd\u4e86&nbsp;SmolLM\uff0c\u8fd9\u662f\u4e00\u4e2a\u5177\u6709 135M\u3001360M \u548c 1.7B \u53c2\u6570\u7684\u6700\u5148\u8fdb\u7684\u5c0f\u578b\u6a21\u578b [&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":"","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":"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":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[444,445,443,442],"tags":[242,314],"class_list":["post-4314","post","type-post","status-publish","format-standard","hentry","category-ai","category-ainews","category-llm","category-llms","tag-chatgpt","tag-openai-api"],"views":3593,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4314","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4314"}],"version-history":[{"count":3,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4314\/revisions"}],"predecessor-version":[{"id":4317,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4314\/revisions\/4317"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4314"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4314"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4314"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}