{"id":928,"date":"2023-06-27T07:15:00","date_gmt":"2023-06-26T22:15:00","guid":{"rendered":"https:\/\/staka.jp\/wordpress\/?p=928"},"modified":"2023-06-27T07:15:00","modified_gmt":"2023-06-26T22:15:00","slug":"mpt-30b-chat-in-context-learning%e3%81%ae%e6%80%a7%e8%83%bd","status":"publish","type":"post","link":"https:\/\/staka.jp\/wordpress\/?p=928","title":{"rendered":"MPT-30B-Chat + In-Context Learning\u306e\u6027\u80fd"},"content":{"rendered":"\n<p>MPT-30B\u306e\u6027\u80fd\u304c\u9ad8\u3044\u3068\u805e\u3044\u3066\u305d\u306eChat\u7248\u3067\u3042\u308b<a href=\"https:\/\/huggingface.co\/mosaicml\/mpt-30b-chat\">mosaicml\/mpt-30b-chat \u00b7 Hugging Face<\/a>\u306e\u6027\u80fd\u3092\u691c\u8a3c\u3057\u3066\u307f\u305f\u3002\u904e\u53bb\u306e\u4e8b\u4f8b\u901a\u308a\u6a5f\u68b0\u7ffb\u8a33\uff08\u82f1\u6587\u548c\u8a33\uff09\u3067GPT-3.5\u3001GPT-4\u3068\u6027\u80fd\u3092\u6bd4\u8f03\u3057\u305f\u3002\u7d50\u679c\u3001\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u306e\u5272\u306b\u9ad8\u3044\u6027\u80fd\u3092\u6301\u3063\u3066\u3044\u308b\u3053\u3068\u304c\u5206\u304b\u3063\u305f[1]\u3002<\/p>\n\n\n\n<p>MPT-30B-Chat\u306e\u30e9\u30a4\u30bb\u30f3\u30b9\u306f<em>CC-By-NC-SA-4.0<\/em>&nbsp;(non-commercial use only)\u3067\u3042\u308b\u3053\u3068\u306b\u6ce8\u610f\u304c\u5fc5\u8981\u3067\u3042\u308b\u3002Apache-2.0\u306eMPT-30B\u3084<em>CC-By-SA-3.0<\/em>\u306eMPT-30B-Instruct\u3068\u306f\u30e9\u30a4\u30bb\u30f3\u30b9\u304c\u7570\u306a\u308b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><tbody><tr><td><strong>\u30e2\u30c7\u30eb \/ shot\u6570 \/ \u5099\u8003<\/strong><\/td><td><strong>BLEU<\/strong><\/td><\/tr><tr><td>MPT-30B-Chat \/ 0 shot \/ \u30e2\u30c7\u30eb\u51fa\u529b\u306e\u307e\u307e<\/td><td>7.7<\/td><\/tr><tr><td>MPT-30B-Chat \/ 0 shot \/ \u30eb\u30fc\u30eb\u3067\u56de\u7b54\u62bd\u51fa<\/td><td>13.0<\/td><\/tr><tr><td>MPT-30B-Chat \/ 0 shot \/ \u624b\u3067\u56de\u7b54\u62bd\u51fa<\/td><td>13.8<\/td><\/tr><tr><td>MPT-30B-Chat \/ 3 shot \/ \u30e2\u30c7\u30eb\u51fa\u529b\u306e\u307e\u307e<\/td><td>16.1<\/td><\/tr><tr><td>MPT-30B-Chat \/ 3 shot \/ \u30eb\u30fc\u30eb\u3067\u56de\u7b54\u62bd\u51fa<\/td><td>29.2<\/td><\/tr><tr><td><strong>MPT-30B-Chat \/ 3 shot \/ \u624b\u3067\u56de\u7b54\u62bd\u51fa<\/strong><\/td><td><strong>30.9<\/strong><\/td><\/tr><tr><td>gpt-3.5-turbo-16k-0613 \/ 3 shot<\/td><td>35.3<\/td><\/tr><tr><td>gpt-4-0613 \/ 3 shot<\/td><td>32.1<\/td><\/tr><tr><td>gpt-4-0314 \/ 3 shot<\/td><td><strong>36.5<\/strong><\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">MPT-30B-Chat, GPT-3.5, GPT-4+OpenICL\u5229\u7528\u6642\u306eBLEU<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u6027\u80fd\u8a55\u4fa1\u306e\u65b9\u6cd5<\/h2>\n\n\n\n<p>\u6027\u80fd\u8a55\u4fa1\u306b\u4f7f\u7528\u3057\u305f\u30c7\u30fc\u30bf\u306f\u4ee5\u524d\uff08<a href=\"https:\/\/staka.jp\/wordpress\/?p=731\">GPT-4\u3092\u7528\u3044\u305f\u7ffb\u8a33\u306e\u691c\u8a3c\uff08vs GPT-3.5 vs FuguMT\uff09 | \u3077\u308b\u30fc\u3075\u304a\u3076\u3053\u3093\u305b\u3077\u3068 (staka.jp)<\/a>\uff09\u3068\u540c\u69d8\u3067\u3042\u308b\u304c\u3001\u51e6\u7406\u6642\u9593\u306e\u95a2\u4fc2\u4e0a[2]\u300150\u4ef6\u3092\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3057\u305f\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u82f1\u8a9e\u6587\u3092\u65e5\u672c\u8a9e\u6587\u306b\u7ffb\u8a33\u3057\u3001\u305d\u306e\u6027\u80fd\u3092\u8a55\u4fa1\u3057\u305f\u3002\u8a55\u4fa1\u6307\u6a19\u306fBLEU\u3067\u3001\u4f7f\u7528\u3057\u305f\u30c4\u30fc\u30eb\u3084tokenizer\u306f\u524d\u56de\u3068\u540c\u3058\uff08sacrebleu \u2013tokenize ja-mecab\uff09\u3067\u3042\u308b\u3002<\/li>\n\n\n\n<li>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306f<a href=\"https:\/\/www.mofa.go.jp\/press\/release\/index.html#archives\">\u5916\u52d9\u7701WEB\u30b5\u30a4\u30c8\u306e\u30d7\u30ec\u30b9\u30ea\u30ea\u30fc\u30b9<\/a>\uff08CC BY\u4e92\u63db\u3067\u5229\u7528\u53ef\uff09\u306e\u3046\u3061\u65e5\u672c\u8a9e\u3001\u82f1\u8a9e\u304c\u5bfe\u5fdc\u3057\u3066\u3044\u308b\u30da\u30fc\u30b8\u3092\u5229\u7528\u3057\u305f\u3002\u8a55\u4fa1\u306b\u4f7f\u7528\u3057\u305f\u5bfe\u8a33\u30da\u30a2\u306f\u524d\u56de\u3068\u540c\u3058\u30022020\u5e741\u6708\uff5e2023\u5e743\u6708\u3067\u6708\u3054\u3068\u306b5\u4ef6\u306e\u30d7\u30ec\u30b9\u30ea\u30ea\u30fc\u30b9\u3092\u9078\u629e\u3057\u5168195\u4ef6\u3092\u53d6\u5f97\u3001\u305d\u306e\u4e0a\u3067\u3055\u3089\u306b50\u4ef6\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3057\u305f[3]\u3002<\/li>\n\n\n\n<li>OpenICL\u3092\u7528\u3044\u3066\u4e8b\u4f8b\u90e8\u5206\u3092\u53d6\u5f97\u3057\u305f\u3002Retriever\u306b\u306fTopkRetriever<a href=\"https:\/\/arxiv.org\/abs\/2101.06804\">[2101.06804] What Makes Good In-Context Examples for GPT-$3$? (arxiv.org)<\/a>\u300d\u3092\u7528\u3044\u305f\u30023 shot\u3068\u66f8\u304b\u308c\u305f\u884c\u306b\u3064\u3044\u3066\u306f\u5168\u304f\u540c\u3058\u30c7\u30fc\u30bf\u3092\u4f7f\u3063\u3066\u3044\u308b\u3002<\/li>\n\n\n\n<li>\u5165\u529b\u30c7\u30fc\u30bf\u306f\u524d\u56deGPT-3.5\/4\u306e\u6bd4\u8f03\u3067\u4f7f\u7528\u3057\u305f\u3082\u306e\u3068\u53ef\u80fd\u306a\u9650\u308a\u5408\u308f\u305b\u3066\u4f5c\u6210\u3057\u305f\u3002\u30d7\u30ed\u30f3\u30d7\u30c8\u306e\u4f5c\u6210\u306f<a href=\"https:\/\/huggingface.co\/spaces\/mosaicml\/mpt-30b-chat\">MPT-30B-Chat &#8211; a Hugging Face Space by mosaicml<\/a>\u3092\u53c2\u8003\u306b\u3057\u3066\u3044\u308b[4]\u3002<\/li>\n\n\n\n<li>\u51e6\u7406\u6642\u9593\u524a\u6e1b\u306e\u305f\u3081MPT-30B-Chat\u306f\u300cload_in_8bit=True\u300d\u3067\u8aad\u307f\u8fbc\u3093\u3067\u3044\u308b\u3002<\/li>\n\n\n\n<li>MPT-30B-Chat\u306e\u51fa\u529b\u306f\u56de\u7b54\u90e8\u5206\u306e\u7279\u5b9a\u304c\u4e00\u7b4b\u7e04\u3067\u306f\u3044\u304b\u306a\u3044\u3053\u3068\u304c\u3042\u308b\u3002\u8907\u6570\u306e\u65b9\u6cd5\u3067\u56de\u7b54\u90e8\u5206\u306e\u7279\u5b9a\u3092\u884c\u3063\u305f\u3002[4][5]\n<ul class=\"wp-block-list\">\n<li>\u30e2\u30c7\u30eb\u51fa\u529b\u306e\u307e\u307e\uff1a\u51fa\u529b\u3092\u305d\u306e\u307e\u307e\u5229\u7528<\/li>\n\n\n\n<li>\u30eb\u30fc\u30eb\u3067\u56de\u7b54\u62bd\u51fa\uff1a\u6b63\u898f\u8868\u73fe\u3092\u7528\u3044&lt;|im_start|&gt;\u3084&lt;|im_end|&gt;\u3092\u3044\u304f\u3064\u304b\u306e\u30d1\u30bf\u30fc\u30f3\u3067\u6307\u5b9a\u3057\u3066\u56de\u7b54\u90e8\u5206\u3092\u53d6\u5f97<\/li>\n\n\n\n<li>\u624b\u3067\u56de\u7b54\u62bd\u51fa\uff1a\u76ee\u3067\u898b\u3066\u56de\u7b54\u90e8\u5206\u3092\u7279\u5b9a<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u7d50\u679c\u3068\u307e\u3068\u3081<\/h2>\n\n\n\n<p>MPT-30B-Chat\u306f\u9577\u3044\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u306b\u5bfe\u5fdc\u3057\u3066\u3044\u308b\u3053\u3068\u304b\u3089In-Context Learning\u304c\u6709\u52b9\u306b\u6a5f\u80fd\u3059\u308b\u3002\u51fa\u529b\u304b\u3089\u56de\u7b54\u90e8\u5206\u3092\u7279\u5b9a\u3059\u308b\u3053\u3068\u306f\u7c21\u5358\u3067\u306f\u306a\u3044\u304c\u3001\u3046\u307e\u304f\u30eb\u30fc\u30eb\u5316\u3059\u308b\u3053\u3068\u3067GPT-3.5\u306b\u8fd1\u3044\u6027\u80fd\u3092\u51fa\u3059\u3053\u3068\u304c\u3067\u304d\u308b\u3002\u3053\u308c\u3089\u306e\u51e6\u7406\u306f\u30e9\u30a4\u30d6\u30e9\u30ea\u304c\u5bfe\u5fdc\u3059\u308c\u3070\u3088\u304f\u3001\u5358\u7d14\u306a\u30eb\u30fc\u30eb\u3067\u3082\u6bd4\u8f03\u7684\u826f\u3044\u30b9\u30b3\u30a2\u304c\u51fa\u305b\u308b[4]\u3002<\/p>\n\n\n\n<p>\u30b9\u30b3\u30a2\u4e0a\u306fMPT-30B-Chat+3 shot\u306eICL\u3067zero shot\u306eGPT-3.5\u4ee5\u4e0a\u306e\u6027\u80fd\u304c\u51fa\u305b\u3066\u3044\u308b\u3002\u65e5\u672c\u8a9e\u306b\u5bfe\u5fdc\u3057\u3066\u3044\u3066\u30ed\u30fc\u30ab\u30eb\u74b0\u5883\u3067\u52d5\u304fLLM\u3068\u3057\u3066\u306f\u975e\u5e38\u306b\u512a\u79c0\u3067\u3042\u308b[6]\u3002\u30aa\u30fc\u30d7\u30f3\u306a\u30e9\u30a4\u30bb\u30f3\u30b9\u306eLLM\u3067\u3042\u308b\u3053\u3068\u304b\u3089\u65e5\u672c\u8a9e\u80fd\u529b\u306e\u5f37\u5316\u3082\u6709\u671b\u305d\u3046\u3067\u4eca\u5f8c\u304c\u975e\u5e38\u306b\u697d\u3057\u307f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u811a\u6ce8<\/h3>\n\n\n\n<p>[1] \uff08\u6700\u8fd1\u8272\u3005\u8a00\u308f\u308c\u3066\u3044\u308b\uff09GPT-3.5\u3084GPT-4\u306e\u30b5\u30a4\u30ba\u3068\u6bd4\u3079\u308b\u3068\u304b\u306a\u308a\u5c0f\u898f\u6a21\u3002<br>[2]  \u300c\u305d\u306e\u4ed6\u300d\u306b\u3082\u66f8\u3044\u3066\u3044\u308b\u901a\u308a\u3001\u691c\u8a3c\u3059\u308b\u306e\u3082\u7d50\u69cb\u306a\u8a08\u7b97\u30ea\u30bd\u30fc\u30b9\u304c\u5fc5\u8981\u3002ICL\u3092\u4f7f\u3063\u3066\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u304c\u9577\u3044\u306e\u3067\u3057\u3087\u3046\u304c\u306a\u3044\u9762\u306f\u3042\u308b\u304c\u2026\uff083 shot\u3067\u306f\u306a\u304fzero shot\u3060\u3068\u304b\u306a\u308a\u901f\u3044\uff09<br>[3] \u4e00\u5b9a\u671f\u9593\u3054\u3068\u306b\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3057\u3066\u3044\u308b\u3002<br>[4] \u7c21\u5358\u306b\u4f7f\u3048\u308b\u30e9\u30a4\u30d6\u30e9\u30ea\u304c\u6b32\u3057\u3044\u3002\u8ab0\u304b\u4f5c\u3063\u3066\u3044\u305d\u3046\u3067\u306f\u3042\u308b\u3057\u65e2\u5b58\u306e\u4f55\u304b\u306e\u30e9\u30a4\u30d6\u30e9\u30ea\u3067\u30b5\u30dd\u30fc\u30c8\u3055\u308c\u305d\u3046\u306a\u6c17\u3082\u3059\u308b\u3002<br>[5] \u30eb\u30fc\u30eb\u3092\u8ffd\u52a0\u3059\u308b\u306a\u3069\u3057\u3066\u81ea\u52d5\u5316\u3057\u305f\u304b\u3063\u305f\u304c\u6642\u9593\u306e\u95a2\u4fc2\u4e0a\u65ad\u5ff5\u3057\u305f\u3002\u3053\u308c\u3089\u306e\u51e6\u7406\u306f\u56de\u7b54\u306e\u7279\u5b9a\u3092\u76ee\u7684\u3068\u3057\u3066\u304a\u308a\u56de\u7b54\u4e2d\u306e\u304a\u304b\u3057\u306a\u51fa\u529b\u3092\u9664\u53bb\u3057\u305f\u308a\u306f\u3057\u3066\u3044\u306a\u3044\u3002\u300c\u624b\u3067\u56de\u7b54\u62bd\u51fa\u300d\u306f\u9811\u5f35\u3063\u3066\u30eb\u30fc\u30eb\u5316\u3057\u305f\u3068\u304d\u306e\u6700\u5584\u306e\u7d50\u679c\u3068\u3057\u3066\u8a18\u8f09\u3057\u3066\u3044\u308b\u3002<br>[6] \u3053\u308c\u306f\u672c\u5f53\u306b\u305d\u3046\u601d\u3046\u3002<br>[7] \u4e2d\u56fd\u8a9e\u306a\u3069\u4ed6\u306e\u8a00\u8a9e\u304c\u6df7\u3056\u308b\u3053\u3068\u304c\u3042\u308b\u3001\u65e5\u672c\u8a9e\u3068\u3057\u3066\u304a\u304b\u3057\u306a\u8868\u73fe\u304c\u51fa\u529b\u3055\u308c\u308b\u3053\u3068\u304c\u3042\u308b\u3001\u306a\u3069\u5370\u8c61\u306e\u60aa\u3044\u51fa\u529b\u304c\u6df7\u3056\u308b\u3053\u3068\u304c\u3042\u308b\u3002\u3053\u308c\u3089\u304c\u60aa\u3044\u65b9\u5411\u306b\u76ee\u7acb\u3064\u306e\u306f\u4e8b\u5b9f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u305d\u306e\u4ed6<\/h3>\n\n\n\n<p>MPT-30B-Chat\u3092\u691c\u8a3c\u3057\u305f\u3068\u3053\u308dGPT-4\u30843.5\u306b\u304b\u306a\u308a\u8feb\u3063\u3066\u3044\u308b\u611f\u3058\u304c\u3042\u308a\u9a5a\u3044\u305f\u3002\u3068\u306f\u3044\u3048\u3001\u76ee\u691c\u8a3c\u3092\u3057\u3066\u307f\u308b\u3068BLEU\u4ee5\u4e0a\u306e\u5dee\u306f\u611f\u3058\u306a\u304f\u3082\u306a\u304f\u65e5\u672c\u8a9e\u80fd\u529b\u306f\u5341\u5206\u3068\u306f\u8a00\u3044\u96e3\u3044[7]\u3002\u6a5f\u68b0\u7ffb\u8a33\u80fd\u529b\u3060\u3051\u3067\u306a\u304f\u7dcf\u5408\u7684\u306a\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3082\u3057\u3066\u307f\u305f\u3044\u3068\u3053\u308d\u3002<\/p>\n\n\n\n<p>\u691c\u8a3c\u306fCloab Pro+\u306eA100 GPU\u3001load_in_8bit=True\u3067\u5b9f\u65bd\u3057\u305f\u3002\u3053\u306e\u74b0\u5883\u3067\u3082\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u304c\u9577\u3044\uff087000\u6587\u5b57\uff09\u5834\u5408\u306f5-10\u5206\u306e\u51e6\u7406\u6642\u9593\u304c\u5fc5\u8981\u3060\u3063\u305f\u3002\u516c\u5f0f\u306e\u8aac\u660e\u901a\u308abfloat16+triton\u69cb\u6210\u3067\u306f20-30\u5206\u304b\u304b\u308b\u3002\u9ad8\u901f\u5316\u624b\u6cd5\u306f\u8272\u3005\u3042\u308b\u3068\u306f\u3044\u3048\u5b9f\u5229\u7528\u3067\u5b89\u5b9a\u52d5\u4f5c\u3055\u305b\u308b\u306e\u306f\u304b\u306a\u308a\u5927\u5909\u305d\u3046\u3067\u306f\u3042\u308b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>MPT-30B\u306e\u6027\u80fd\u304c\u9ad8\u3044\u3068\u805e\u3044\u3066\u305d\u306eChat\u7248\u3067\u3042\u308bmosaicml\/mpt-30b-chat \u00b7 Hugging Face\u306e\u6027\u80fd\u3092\u691c\u8a3c\u3057\u3066\u307f\u305f\u3002\u904e\u53bb\u306e\u4e8b\u4f8b\u901a\u308a\u6a5f\u68b0\u7ffb\u8a33\uff08\u82f1\u6587\u548c\u8a33\uff09\u3067GPT-3.5\u3001GPT-4\u3068\u6027\u80fd\u3092\u6bd4 &hellip; <a href=\"https:\/\/staka.jp\/wordpress\/?p=928\" class=\"more-link\"><span class=\"screen-reader-text\">&#8220;MPT-30B-Chat + In-Context Learning\u306e\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":[37,38,48,49],"class_list":["post-928","post","type-post","status-publish","format-standard","hentry","category-ai","category-13","tag-gpt-3-5-turbo","tag-gpt-4","tag-mpt-30b","tag-mpt-30b-chat"],"_links":{"self":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/928","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=928"}],"version-history":[{"count":0,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/928\/revisions"}],"wp:attachment":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=928"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=928"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=928"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}