{"id":983,"date":"2023-07-23T17:00:00","date_gmt":"2023-07-23T08:00:00","guid":{"rendered":"https:\/\/staka.jp\/wordpress\/?p=983"},"modified":"2023-07-23T17:00:00","modified_gmt":"2023-07-23T08:00:00","slug":"llama-v2%ef%bc%887b-13b-chat%ef%bc%89%e3%81%ae%e6%a9%9f%e6%a2%b0%e7%bf%bb%e8%a8%b3%e6%80%a7%e8%83%bd","status":"publish","type":"post","link":"https:\/\/staka.jp\/wordpress\/?p=983","title":{"rendered":"Llama 2\uff087B, 13B chat\uff09\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd"},"content":{"rendered":"\n<p>Llama 2\u304c\u767a\u8868\u3055\u308c\u305f\uff08\u516c\u5f0f\u30b5\u30a4\u30c8\u306f<a href=\"https:\/\/ai.meta.com\/llama\/\">Llama 2 &#8211; Meta AI<\/a>\u3000huggingface\u3078\u306f<a href=\"https:\/\/huggingface.co\/meta-llama\">meta-llama (Meta Llama 2) (huggingface.co)<\/a>\uff09\u306e\u3067\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u3092\u691c\u8a3c\u3057\u3066\u307f\u305f\u3002\u8ad6\u6587\u306b\u3088\u308b\u3068\u65e5\u672c\u8a9e\u30c7\u30fc\u30bf\u306f0.1%\u7a0b\u5ea6\u3057\u304b\u542b\u307e\u308c\u3066\u3044\u306a\u3044[1]\u3068\u306e\u3053\u3068\u3060\u304c\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u306f\u304b\u306a\u308a\u9ad8\u3044\u3002<\/p>\n\n\n\n<p>\u6027\u80fd\u8a55\u4fa1\u306b\u4f7f\u7528\u3057\u305f\u30c7\u30fc\u30bf\u306f\u524d\u56de\uff08<a href=\"https:\/\/staka.jp\/wordpress\/?p=959\">DAMO PolyLM-13B\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd | \u3077\u308b\u30fc\u3075\u304a\u3076\u3053\u3093\u305b\u3077\u3068 (staka.jp)<\/a>\uff09\u3068\u540c\u3058\u3002\u540c\u69d8\u306e\u6761\u4ef6\u3067\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3057\u3066\u4f7f\u3063\u3066\u3044\u308b\u3002\u4ed6\u306e\u30d5\u30ea\u30fc\u306a\u30e2\u30c7\u30eb\u306b\u6bd4\u3079\u3066\u56de\u7b54\u90e8\u5206\u306e\u62bd\u51fa\u306f\u5bb9\u6613\u306b\u306a\u3063\u3066\u3044\u308b\u304c\u305d\u308c\u3067\u3082\u5931\u6557\u3059\u308b\u3053\u3068\u304c\u3042\u308b[2]\u3002\u305d\u306e\u305f\u3081\u30eb\u30fc\u30eb\u30d9\u30fc\u30b9\u3067\u56de\u7b54\u90e8\u5206\u3092\u62bd\u51fa\u3057\u305f\u7d50\u679c\u3068\u624b\u52d5\u62bd\u51fa\u3057\u305f\u7d50\u679c\u306e2\u901a\u308a\u3067BLEU\u3092\u7b97\u51fa\u3057\u3066\u3044\u308b\u3002<\/p>\n\n\n\n<p>\u74b0\u5883\u306fColab Pro+\u3092\u7528\u3044float16\u3067\u8aad\u307f\u8fbc\u3093\u3060\u3002A100 VRAM 40GB\u3067\u306f\u30e1\u30e2\u30ea\u5236\u7d04\u304c\u53b3\u3057\u3044\u306e\u3082\u524d\u56de\u540c\u69d8\u3002\u7d50\u679c\u306f\u4e0b\u8868\u306e\u901a\u308a\u300213B\u30e2\u30c7\u30eb\u306e\u6027\u80fd\u306f\u9ad8\u304f\uff08\u7279\u306b\u6700\u5927\u306e\u69cb\u621070B\u3067\u3042\u308c\u3070\uff09ChatGPT\uff08GPT-3.5\uff09\u76f8\u5f53\u3068\u3044\u3046\u8a55\u4fa1\u306f\u904e\u5927\u3068\u3044\u3046\u308f\u3051\u3067\u306f\u306a\u3055\u305d\u3046\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><tbody><tr><td><strong>\u30e2\u30c7\u30eb\u30fb\u6761\u4ef6<\/strong><\/td><td><strong>zero shot<\/strong><\/td><td><strong>1 shot<\/strong> [3]<\/td><\/tr><tr><td>GPT-3.5 [4]<\/td><td>26.7<\/td><td><strong>37.0<\/strong><\/td><\/tr><tr><td>Llama-2-7b-chat-hf<\/td><td>11.6<\/td><td>20.1<\/td><\/tr><tr><td>Llama-2-7b-chat-hf<br>\u203b\u624b\u3067\u56de\u7b54\u62bd\u51fa<\/td><td>13.4<\/td><td>23.9<\/td><\/tr><tr><td>Llama-2-13b-chat-hf<\/td><td>19.1<\/td><td>33.1<\/td><\/tr><tr><td>Llama-2-13b-chat-hf<br>\u203b\u624b\u3067\u56de\u7b54\u62bd\u51fa<\/td><td>19.1<\/td><td><strong>35.1<\/strong><\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">Llama 2 7B\/13B, GPT-3.5-TURBO-16K\u306eBLEU<\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d50\u679c<\/h3>\n\n\n\n<p>\u8a71\u984c\u306eLlama 2\u3092\u8a66\u3057\u3066\u307f\u305f\u300213B\u304b\u30641 shot\u8a2d\u5b9a\u3067\u3046\u307e\u304f\u56de\u7b54\u304c\u62bd\u51fa\u3067\u304d\u308c\u3070ChatGPT\uff08GPT-3.5\uff09\u306b\u8feb\u308b\u30ec\u30d9\u30eb\u3002\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u5de5\u592b\u3059\u308c\u3070\u3082\u3046\u5c11\u3057\u30b9\u30b3\u30a2\u3092\u51fa\u305b\u305d\u3046\u3068\u3082\u611f\u3058\u308b\u3002<\/p>\n\n\n\n<p>Llama 2\u306e\u4e8b\u524d\u5b66\u7fd2\u30c8\u30fc\u30af\u30f3\u91cf\u306f2T\u3068\u3055\u308c\u3066\u3044\u3066\u65e5\u672c\u8a9e\u30c7\u30fc\u30bf\u306f\u305d\u306e\u3046\u30610.1%\u3068\u306e\u3053\u3068\u3002\u65e5\u672c\u8a9e\u30c7\u30fc\u30bf\u306e\u5272\u5408\u304c\u5c11\u306a\u304f\u3066\u3082\u4e00\u5b9a\u306e\u6027\u80fd\u304c\u51fa\u305b\u308b\u3053\u3068\u306b\u9a5a\u304d\u3002\u7d76\u5bfe\u91cf\u3068\u3057\u3066\u30822TB\u00d70.1%=2G tokens\u3067\u3042\u308a\u591a\u3044\u3068\u8a00\u3048\u3070\u591a\u3044\u304c\u5927\u91cf\u3068\u3044\u3046\u307b\u3069\u3067\u3082\u306a\u3044\u30027B\u306813B\u3067\u5927\u304d\u306a\u6027\u80fd\u5dee\u304c\u3042\u308b\u4e8b\u3082\u8208\u5473\u6df1\u3044\u3002MPT\u306e\u691c\u8a3c\u3067\u3082\u611f\u3058\u305f\u304c\u6a5f\u68b0\u7ffb\u8a33\u3092\u884c\u3046\u5834\u54087B\u3067\u306f\u53b3\u3057\u3044\u306e\u304b\u3082\u3057\u308c\u306a\u3044\u300270B\u3067\u3069\u3046\u306a\u308b\u304b\u3001\u7279\u306bzero shot\u3067\u306e\u6027\u80fd\u304c\u6539\u5584\u3059\u308b\u304b\u975e\u5e38\u306b\u6c17\u306b\u306a\u308b\u3068\u3053\u308d[5]\u3002<\/p>\n\n\n\n<p>Llama 2\u3092\u7528\u3044\u305f\u6d3e\u751f\u30e2\u30c7\u30eb\u306f\u4eca\u5f8c\u51fa\u3066\u304f\u308b\u306f\u305a\u3067\u65e5\u672c\u8a9e\u80fd\u529b\u306e\u5f37\u5316\u30d0\u30fc\u30b8\u30e7\u30f3\u3082\u958b\u767a\u3055\u308c\u308b\u306f\u305a\u3002\u3059\u3067\u306b<a href=\"https:\/\/stability.ai\/blog\/freewilly-large-instruction-fine-tuned-models\">Meet FreeWilly, Our Large And Mighty Instruction Fine-Tuned Models \u2014 Stability AI<\/a>\u304c\u767a\u8868\u3055\u308c\u3066\u3044\u308b\u3002\uff08\u524d\u56de\u306e\u611f\u60f3\u3068\u540c\u3058\u3060\u304c\uff09\u30d5\u30ea\u30fc\u306eLLM\u306b\u671f\u5f85\u3057\u305f\u3044\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u811a\u6ce8<\/h4>\n\n\n\n<p>[1] <a href=\"https:\/\/ai.meta.com\/research\/publications\/llama-2-open-foundation-and-fine-tuned-chat-models\/\">Llama 2: Open Foundation and Fine-Tuned Chat Models | Meta AI Research<\/a> \u8ad6\u6587\u306e22\u30da\u30fc\u30b8\u76ee<br>[2] \u7279\u306bICL\u5229\u7528\u6642\u306b\u610f\u56f3\u3057\u306a\u3044\u51fa\u529b\u304c\u7d9a\u304f\u3053\u3068\u304c\u591a\u304b\u3063\u305f\u3002\u516c\u5f0f\u306b\u3082\u66f8\u304b\u308c\u3066\u3044\u308b\u901a\u308a<a href=\"https:\/\/github.com\/facebookresearch\/llama\/blob\/main\/llama\/generation.py\">llama\/llama\/generation.py at main \u00b7 facebookresearch\/llama \u00b7 GitHub<\/a>\u3092\u53c2\u8003\u306b\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u7d44\u307f\u7acb\u3066\u308b\u3068\u6539\u5584\u3059\u308b\u3082\u306e\u3068\u601d\u308f\u308c\u308b\u3002\uff08\u6642\u9593\u304c\u7121\u3044\u3053\u3068\u3082\u3042\u3063\u3066\u30d7\u30ed\u30f3\u30d7\u30c8\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u306f\u5341\u5206\u3067\u306f\u306a\u3044\u2026\uff09<br>[3] OpenICL\u3001TopkRetriever\u306b\u3088\u308a\u53d6\u5f97<br>[4] gpt-3.5-turbo-16k-0613<br>[5] 4bit\u91cf\u5b50\u5316\u3057\u3066\u3082A100(40GB)\u3060\u3068\u63a8\u8ad6\u6642\u306bOut of Memory\u306b\u306a\u308b\u2026<br><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Llama 2\u304c\u767a\u8868\u3055\u308c\u305f\uff08\u516c\u5f0f\u30b5\u30a4\u30c8\u306fLlama 2 &#8211; Meta AI\u3000huggingface\u3078\u306fmeta-llama (Meta Llama 2) (huggingface.co)\uff09\u306e\u3067\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u3092\u691c &hellip; <a href=\"https:\/\/staka.jp\/wordpress\/?p=983\" class=\"more-link\"><span class=\"screen-reader-text\">&#8220;Llama 2\uff087B, 13B chat\uff09\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd&#8221; \u306e<\/span>\u7d9a\u304d\u3092\u8aad\u3080<\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2,13],"tags":[37,43],"class_list":["post-983","post","type-post","status-publish","format-standard","hentry","category-ai","category-13","tag-gpt-3-5-turbo","tag-llama-2"],"_links":{"self":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/983","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=983"}],"version-history":[{"count":0,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/983\/revisions"}],"wp:attachment":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=983"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=983"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=983"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}