{"id":284376,"date":"2023-10-18T09:00:00","date_gmt":"2023-10-18T01:00:00","guid":{"rendered":"https:\/\/lrxjmw.cn\/?p=284376"},"modified":"2023-10-09T14:57:30","modified_gmt":"2023-10-09T06:57:30","slug":"mistral-7b","status":"publish","type":"post","link":"https:\/\/lrxjmw.cn\/mistral-7b.html","title":{"rendered":"Mistral AI\u53d1\u5e03\u4e00\u4e2a\u62e5\u6709 73 \u4ebf\u53c2\u6570\u6a21\u578bMistral 7B"},"content":{"rendered":"<table>\n<tbody>\n<tr>\n<td style=\"padding: 5px 12px; width: 25px; font-size: 22px; text-align: center; background: #E8E8E8; line-height: 30px; color: #999;\">\u5bfc\u8bfb<\/td>\n<td style=\"padding: 10px 10px 12px; line-height: 20px;\"><strong>\u6cd5\u56fd\u4eba\u5de5\u667a\u80fd\u521d\u521b\u516c\u53f8 Mistral AI\u00a0\u5ba3\u5e03\u63a8\u51fa\u5176\u9996\u6b3e\u5927\u8bed\u8a00\u6a21\u578b Mistral 7B \u662f\u4e00\u4e2a\uff0c\u53f7\u79f0\u662f\u8fc4\u4eca\u4e3a\u6b62\u540c\u89c4\u6a21\u4ea7\u54c1\u4e2d\u6700\u5f3a\u5927\u7684\u8bed\u8a00\u6a21\u578b\uff1b\u5728 Apache-2.0 \u8bb8\u53ef\u4e0b\u5f00\u6e90\uff0c\u53ef\u5b8c\u5168\u514d\u8d39\u4f7f\u7528\uff0c\u4e0d\u53d7\u4efb\u4f55\u9650\u5236\u3002<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Mistral AI \u662f\u4e00\u4e2a\u6210\u7acb\u4ec5\u516d\u4e2a\u6708\u7684\u521d\u521b\u516c\u53f8\uff0c\u4e8e 6 \u6708\u4efd\u7b79\u96c6\u4e86 1.18 \u4ebf\u7f8e\u5143\u7684\u5de8\u989d\u79cd\u5b50\u8f6e\u8d44\u91d1\uff0c\u636e\u79f0\u662f\u6b27\u6d32\u5386\u53f2\u4e0a\u6700\u5927\u7684\u79cd\u5b50\u8f6e\u878d\u8d44\u3002Mistral 7B \u662f\u4e00\u4e2a\u62e5\u6709 73 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u3002\u8be5\u516c\u53f8\u58f0\u79f0\u5728\u6db5\u76d6\u4e00\u7cfb\u5217\u4efb\u52a1\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cMistral 7B \u7684\u8868\u73b0\u5747\u663e\u8457\u4f18\u4e8e Llama 2 7B \u548c 13B\uff0c\u5e76\u4e14\u4e0e Llama 34B \u76f8\u5f53\u3002<\/p>\n<p>\u5728\u6db5\u76d6\u6570\u5b66\u3001\u7f8e\u56fd\u5386\u53f2\u3001\u8ba1\u7b97\u673a\u79d1\u5b66\u3001\u6cd5\u5f8b\u7b49 57 \u4e2a\u79d1\u76ee\u7684\u5927\u89c4\u6a21\u591a\u4efb\u52a1\u8bed\u8a00\u7406\u89e3 (MMLU) \u6d4b\u8bd5\u4e2d\uff0cMistral 7B \u6a21\u578b\u7684\u51c6\u786e\u7387\u4e3a 60.1%\uff0cLlama 2 7B \u548c 13B \u7684\u51c6\u786e\u7387\u7565\u9ad8\u4e8e\u5206\u522b\u4e3a 44.4% \u548c 55.6%\u3002<\/p>\n<p>\u5728\u5e38\u8bc6\u63a8\u7406\u548c\u9605\u8bfb\u7406\u89e3\u6d4b\u8bd5\u4e2d\uff0cMistral 7B \u7684\u51c6\u786e\u7387\u4e5f\u8d85\u8fc7\u4e86\u4e24\u4e2a Llama \u6a21\u578b\u3002\u5728\u4e16\u754c\u77e5\u8bc6\u6d4b\u8bd5\u4e2d\uff0cLlama 2 13B \u4e0e Mistral 7B \u4e0d\u76f8\u4e0a\u4e0b\uff0cMistral \u79f0\u8fd9\u53ef\u80fd\u662f\u7531\u4e8e\u6a21\u578b\u7684\u53c2\u6570\u6570\u91cf\u6709\u9650\uff0c\u9650\u5236\u4e86\u5b83\u53ef\u4ee5\u538b\u7f29\u7684\u77e5\u8bc6\u91cf\u3002<\/p>\n<p>\u552f\u4e00 Llama 2 13B \u548c Mistral 7B \u4e0d\u76f8\u4e0a\u4e0b\u7684\u9886\u57df\u5219\u662f\u4e16\u754c\u77e5\u8bc6\u6d4b\u8bd5\uff0cMistral \u58f0\u79f0 \u201c\u8fd9\u53ef\u80fd\u662f\u7531\u4e8e Mistral 7B \u7684\u53c2\u6570\u6570\u91cf\u6709\u9650\uff0c\u4ece\u800c\u9650\u5236\u4e86\u5176\u53ef\u538b\u7f29\u7684\u77e5\u8bc6\u91cf\u3002\u201d<\/p>\n<p>\u5728\u7f16\u7801\u4efb\u52a1\u65b9\u9762\uff0c\u867d\u7136 Mistral \u79f0 Mistral 7B \u7684\u6027\u80fd\u5927\u4e3a\u63d0\u9ad8\uff1b\u4f46\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c\u8868\u660e\uff0c\u5b83\u4ecd\u7136\u6ca1\u6709\u8d85\u8fc7\u7ecf\u8fc7\u5fae\u8c03\u7684 CodeLlama 7B\u3002\u5728 0-shot Humaneval \u548c 3-shot MBPP \u6d4b\u8bd5\u4e2d\uff0cCodeLlama 7B \u7684\u51c6\u786e\u7387\u5206\u522b\u4e3a 31.1% \u548c 52.5%\uff0cMistral 7B \u5219\u5206\u522b\u4e3a 30.5% \u548c 47.5%\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone wp-image-284377 size-full\" src=\"https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-scaled.webp\" alt=\"\" width=\"2560\" height=\"833\" srcset=\"https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-scaled.webp 2560w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-300x98.webp 300w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-1024x333.webp 1024w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-150x49.webp 150w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-768x250.webp 768w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-1536x500.webp 1536w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/11-2048x666.webp 2048w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/> <img loading=\"lazy\" decoding=\"async\" class=\"alignnone wp-image-284378 size-full\" src=\"https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-scaled.webp\" alt=\"\" width=\"2560\" height=\"356\" srcset=\"https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-scaled.webp 2560w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-300x42.webp 300w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-1024x142.webp 1024w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-150x21.webp 150w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-768x107.webp 768w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-1536x213.webp 1536w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/22-2048x285.webp 2048w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/> <img loading=\"lazy\" decoding=\"async\" class=\"alignnone wp-image-284379 size-full\" src=\"https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-scaled.webp\" alt=\"\" width=\"2560\" height=\"1772\" srcset=\"https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-scaled.webp 2560w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-300x208.webp 300w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-1024x709.webp 1024w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-150x104.webp 150w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-768x532.webp 768w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-1536x1063.webp 1536w, https:\/\/lrxjmw.cn\/wp-content\/uploads\/2023\/10\/33-2048x1418.webp 2048w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/p>\n<p>Mistral AI \u8868\u793a\uff0cMistral 7B \u4f7f\u7528\u4e86 Grouped-query attention (GQA) \u5b9e\u73b0\u66f4\u5feb\u7684\u63a8\u7406\uff0c\u5e76\u4f7f\u7528 Sliding Window Attention (SWA) \u4ee5\u8f83\u5c0f\u7684\u6210\u672c\u5904\u7406\u66f4\u957f\u7684\u5e8f\u5217\u3002<\/p>\n<p>\u201cMistral 7B \u91c7\u7528\u4e86 SWA\uff0c\u5176\u4e2d\u6bcf\u4e00\u5c42\u90fd\u5173\u6ce8\u5148\u524d\u7684 4096 \u4e2a\u9690\u85cf\u72b6\u6001\u3002\u4e3b\u8981\u6539\u8fdb\u4ee5\u53ca\u6700\u521d\u7814\u7a76\u7684\u539f\u56e0\u662f O (sliding_window.seq_len) \u7684\u7ebf\u6027\u8ba1\u7b97\u6210\u672c\u3002\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u5bf9 FlashAttention \u548c xFormers \u6240\u505a\u7684\u66f4\u6539\u4f7f\u5f97\u5728\u5e8f\u5217\u957f\u5ea6\u4e3a 16k \u4e14\u7a97\u53e3\u4e3a 4k \u7684\u60c5\u51b5\u4e0b\uff0c\u901f\u5ea6\u63d0\u9ad8\u4e86 2 \u500d\u3002\u201d<\/p>\n<p>\u6b64\u5916\uff0c\u8be5\u516c\u53f8\u8fd8\u8ba1\u5212\u5728\u8fd9\u9879\u5de5\u4f5c\u7684\u57fa\u7840\u4e0a\u53d1\u5e03\u4e00\u4e2a\u66f4\u5927\u7684\u6a21\u578b\uff0c\u80fd\u591f\u66f4\u597d\u5730\u63a8\u7406\u5e76\u652f\u6301\u591a\u79cd\u8bed\u8a00\uff0c\u9884\u8ba1\u5c06\u4e8e 2024 \u5e74\u4eae\u76f8\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Mistral AI \u662f\u4e00\u4e2a\u6210\u7acb\u4ec5\u516d\u4e2a\u6708\u7684\u521d\u521b\u516c\u53f8\uff0c\u4e8e 6 \u6708\u4efd\u7b79\u96c6\u4e86 1.18 \u4ebf\u7f8e\u5143\u7684\u5de8\u989d\u79cd\u5b50\u8f6e\u8d44\u91d1\uff0c\u636e\u79f0 [&hellip;]<\/p>\n","protected":false},"author":323,"featured_media":284380,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[21],"tags":[],"class_list":["post-284376","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"acf":[],"_links":{"self":[{"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/posts\/284376","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/users\/323"}],"replies":[{"embeddable":true,"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/comments?post=284376"}],"version-history":[{"count":2,"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/posts\/284376\/revisions"}],"predecessor-version":[{"id":284866,"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/posts\/284376\/revisions\/284866"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/media\/284380"}],"wp:attachment":[{"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/media?parent=284376"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/categories?post=284376"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lrxjmw.cn\/wp-json\/wp\/v2\/tags?post=284376"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}