{"id":6712,"date":"2024-07-04T11:00:29","date_gmt":"2024-07-04T02:00:29","guid":{"rendered":"https:\/\/highreso.jp\/edgehub\/?p=6712"},"modified":"2024-07-20T19:38:10","modified_gmt":"2024-07-20T10:38:10","slug":"sarashinafinetune","status":"publish","type":"post","link":"https:\/\/highreso.jp\/edgehub\/machinelearning\/sarashinafinetune.html","title":{"rendered":"Sarashina\u3092\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3010\u30bd\u30d5\u30c8\u30d0\u30f3\u30af\u306e\u65e5\u672c\u8a9eLLM\u3011"},"content":{"rendered":"\n<p><span class=\"swl-format-2\"><span class=\"swl-format-2\">\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001Sarashina<strong>(\u30b5\u30e9\u30b7\u30ca)<\/strong>\u306e\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002<\/span><\/span><\/p>\n\n\n\n<p>Sarashina\u306f\u30bd\u30d5\u30c8\u30d0\u30f3\u30af\u306e\u5b50\u4f1a\u793e\u306eSB Intuitions\u304c\u958b\u767a\u3057\u305f\u65e5\u672c\u8a9e\u306b\u5f37\u3044LLM\u3067\u3059\u3002<\/p>\n\n\n\n<p>Sarashina\u306e\u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb\u306b\u6307\u793a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3092\u65bd\u3059\u3053\u3068\u3067\u3001\u30c1\u30e3\u30c3\u30c8\u5f62\u5f0f\u306e\u5fdc\u7b54\u304c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Sarashina\u306e\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"600\" height=\"300\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digtalneuralnet_midashi.jpg\" alt=\"\u898b\u51fa\u3057\u753b\u50cf\" class=\"wp-image-6604\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digtalneuralnet_midashi.jpg 600w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digtalneuralnet_midashi-300x150.jpg 300w\" sizes=\"(max-width: 600px) 100vw, 600px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">Sarashina\u3068\u306f<\/h3>\n\n\n\n<p><span class=\"swl-format-2\">Sarashina(\u30b5\u30e9\u30b7\u30ca)\u306f\u3001\u30bd\u30d5\u30c8\u30d0\u30f3\u30af\u306e\u5b50\u4f1a\u793e\u306eSB Intuitions\u304c\u958b\u767a\u3057\u305f\u65e5\u672c\u8a9e\u306b\u5f37\u3044LLM\u3067\u3059\u3002<\/span><\/p>\n\n\n\n<p>Sarashina\u306f\u3001\u65e5\u672c\u8a9e\u30c7\u30fc\u30bf\u30921\u5146\u30c8\u30fc\u30af\u30f3\u3082\u5b66\u7fd2\u3057\u3066\u304a\u308a\u3001\u3053\u308c\u306fMeta\u304c\u958b\u767a\u3057\u305fLlama2\u306e\u65e5\u672c\u8a9e\u5b66\u7fd2\u91cf\u306e\u6570\u767e\u500d\u306b\u76f8\u5f53\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<p>Sarashina\u306b\u3064\u3044\u3066\u8a73\u3057\u304f\u77e5\u308a\u305f\u3044\u65b9\u306f\u3001\u4ee5\u4e0b\u306e\u8a18\u4e8b\u3092\u3054\u89a7\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n<div class=\"swell-block-postLink\">\t\t\t<div class=\"p-blogCard -internal\" data-type=\"type1\" data-onclick=\"clickLink\">\n\t\t\t\t<div class=\"p-blogCard__inner\">\n\t\t\t\t\t<span class=\"p-blogCard__caption\">\u3042\u308f\u305b\u3066\u8aad\u307f\u305f\u3044<\/span>\n\t\t\t\t\t<div class=\"p-blogCard__thumb c-postThumb\"><figure class=\"c-postThumb__figure\"><img decoding=\"async\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/sarashinatoha-eyecatch-300x158.jpg\" alt=\"\" class=\"c-postThumb__img u-obf-cover\" width=\"320\" height=\"180\"><\/figure><\/div>\t\t\t\t\t<div class=\"p-blogCard__body\">\n\t\t\t\t\t\t<a class=\"p-blogCard__title\" href=\"https:\/\/highreso.jp\/edgehub\/machinelearning\/sarashinatoha.html\">\u30bd\u30d5\u30c8\u30d0\u30f3\u30af\u306e\u65e5\u672c\u8a9eLLM\u300cSarashina\u300d\u306e\u4f7f\u3044\u65b9\u3092\u89e3\u8aac\uff01<\/a>\n\t\t\t\t\t\t<span class=\"p-blogCard__excerpt\">Sarashina\u306f\u3001\u30bd\u30d5\u30c8\u30d0\u30f3\u30af\u306e\u5b50\u4f1a\u793e\u306eSB Intuitions\u304c\u958b\u767a\u3057\u305f\u65e5\u672c\u8a9e\u306b\u5f37\u3044LLM\u3067\u3059\u3002 70\u5104\u30fb130\u5104\u30fb650\u5104\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u8907\u6570\u30e2\u30c7\u30eb\u3092\u516c\u958b\u3057\u3001\u3044\u305a\u308c\u3082\u7121\u6599\u3067\u5546\u7528\u5229\u7528\u53ef\u80fd&#8230;<\/span>\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\n\n<h3 class=\"wp-block-heading\">\u6307\u793a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3068\u306f<\/h3>\n\n\n\n<p>Sarashina\u306e\u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb\u3067\u306f\u3001\u30c1\u30e3\u30c3\u30c8\u5f62\u5f0f\u306e\u5fdc\u7b54\u304c\u3067\u304d\u306a\u3044\u305f\u3081\u3001\u5225\u9014\u6307\u793a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u304c\u5fc5\u8981\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">\u6307\u793a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\uff08Instruction Tuning\uff09\u3068\u306f\u3001\u30e2\u30c7\u30eb\u3092\u6307\u793a\u306b\u5fdc\u3058\u305f\u5fdc\u7b54\u304c\u3067\u304d\u308b\u3088\u3046\u306b\u5b66\u7fd2\u3055\u305b\u3001\u9069\u5fdc\u3055\u305b\u308b\u3053\u3068\u3067\u3059\u3002<\/span><\/p>\n\n\n\n<p>\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001HuggingFace\u306eSFTTrainer\u3092\u4f7f\u3063\u3066\u3001Sarashina\u306e\u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb\u306b\u5bfe\u3057\u3066\u30d5\u30eb\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3092\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<div class=\"swell-block-capbox cap_box is-style-onborder_ttl2\" data-colset=\"col3\"><div class=\"cap_box_ttl\"><span>\u7528\u8a9e\u306e\u89e3\u8aac<\/span><\/div><div class=\"cap_box_content\">\n<p><span class=\"swl-format-2\">SFTTrainer<\/span><\/p>\n\n\n\n<p>SFTTrainer\u306f\u3001\u6559\u5e2b\u3042\u308a\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\uff08Supervised Fine-tuning\uff09\u3067\u5b66\u7fd2\u3059\u308b\u305f\u3081\u306e\u30e9\u30a4\u30d6\u30e9\u30ea\u3067\u3059\u3002<\/p>\n\n\n\n<p>Transformers\u30e9\u30a4\u30d6\u30e9\u30ea\u306b\u7d71\u5408\u3055\u308c\u3066\u304a\u308a\u3001\u5c11\u306a\u3044\u30b3\u30fc\u30c9\u3067\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u306e\u5b9f\u88c5\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n<h3 class=\"wp-block-heading\">\u4f7f\u7528\u3059\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8<\/h3>\n\n\n\n<p><span class=\"swl-format-2\">\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u306b\u306f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u300ckunishou\/databricks-dolly-15k-ja\u300d\u3092\u4f7f\u7528\u3057\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<p>15,000\u4ee5\u4e0a\u306e\u6307\u793a\u3068\u5fdc\u7b54\u3067\u69cb\u6210\u3055\u308c\u305f\u65e5\u672c\u8a9e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3067\u3059\u3002<\/p>\n\n\n\n<p>\u82f1\u8a9e\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u65e5\u672c\u8a9e\u306b\u81ea\u52d5\u7ffb\u8a33\u3057\u3066\u4f5c\u6210\u3055\u308c\u305f\u305f\u3081\u3001\u7ffb\u8a33\u8abf\u306a\u65e5\u672c\u8a9e\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<figure class=\"wp-block-embed\"><div class=\"wp-block-embed__wrapper\">\nhttps:\/\/huggingface.co\/datasets\/kunishou\/databricks-dolly-15k-ja\n<\/div><\/figure>\n\n\n\n<div class=\"p-blogParts post_content\" data-partsID=\"4485\"><p><div class=\"p-adBox -ranking -border-on\" data-id=\"4507\" data-ad=\"ranking\"><div class=\"p-adBox__title -rank1\">1\u6642\u959350\u5186\u304b\u3089\u4f7f\u3048\u308bGPU\u30af\u30e9\u30a6\u30c9\u300cGPUSOROBAN\u300d<\/div><div class=\"p-adBox__body\"><div class=\"p-adBox__img\"><div class=\"centerw250\"><a href=\"https:\/\/soroban.highreso.jp\/?utm_medium=edgehub&utm_source=edgehub&utm_campaign=highend-boxtop&adid=edgehub\" rel=\"nofollow\" referrerpolicy=\"no-referrer-when-downgrade\"><img decoding=\"async\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/gpusorobanlllm_banner.jpg\" alt=\"GPUSOROBAN\"><\/a><\/div><\/div><div class=\"p-adBox__details\"><div class=\"p-adBox__star c-reviewStars\"><i class=\"icon-star-full\"><\/i><i class=\"icon-star-full\"><\/i><i class=\"icon-star-full\"><\/i><i class=\"icon-star-full\"><\/i><i class=\"icon-star-full\"><\/i><\/div><div class=\"p-adBox__price u-thin u-fz-s\">GPU\u306e\u30b3\u30b9\u30c8\u3067\u304a\u60a9\u307f\u306e\u65b9\u306b\uff01<\/div><div class=\"p-adBox__desc\">\u751f\u6210AI\u30fbLLM\u5411\u3051\u306e\u9ad8\u901fGPU\u3092\u696d\u754c\u6700\u5b89\u7d1a\u306e1\u6642\u959350\u5186\u304b\u3089\u5229\u7528\u3067\u304d\u307e\u3059\u3002\r\n\u30a4\u30f3\u30bf\u30fc\u30cd\u30c3\u30c8\u74b0\u5883\u3055\u3048\u3042\u308c\u3070\u3001GPU\u30b5\u30fc\u30d0\u30fc\u3092\u3059\u3050\u306b\u4f7f\u3048\u307e\u3059\u3002\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u3092\u8d77\u52d5\u3057\u3066\u3044\u308b\u3068\u304d\u306e\u307f\u306e\u8ab2\u91d1\u3067\u3001\u30b3\u30b9\u30c8\u3092\u6700\u5c0f\u5316\u3067\u304d\u307e\u3059\u3002\r\n<\/div><\/div><\/div><div class=\"p-adBox__btns\"><a href=\"https:\/\/soroban.highreso.jp\/compute\/?utm_medium=edgehub&utm_source=edgehub&utm_campaign=boxtop&adid=edgehub\" class=\"p-adBox__btn -btn1\" target=\"_blank\" rel=\"noopener nofollow\">\u516c\u5f0f\u30b5\u30a4\u30c8\u3092\u898b\u308b<\/a><a href=\"https:\/\/soroban.highreso.jp\/download\/compute-catalog\/?utm_medium=edgehub&utm_source=edgehub&utm_campaign=boxtop&adid=edgehub\" class=\"p-adBox__btn -btn2\" target=\"_blank\" rel=\"noopener nofollow\">\u8cc7\u6599\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9<\/a><\/div><\/div><\/p>\n<\/div>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8b\u524d\u6e96\u5099<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"600\" height=\"300\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digitalflower_midashi.jpg\" alt=\"\u898b\u51fa\u3057\u753b\u50cf\" class=\"wp-image-6603\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digitalflower_midashi.jpg 600w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digitalflower_midashi-300x150.jpg 300w\" sizes=\"(max-width: 600px) 100vw, 600px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">\u5fc5\u8981\u306a\u30b9\u30da\u30c3\u30af\u30fb\u5b9f\u884c\u74b0\u5883<\/h3>\n\n\n\n<p>Sarashina\u306e\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u306f\u3001\u5927\u5bb9\u91cf\u306eGPU\u30e1\u30e2\u30ea\u3092\u5fc5\u8981\u3068\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<p>\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001<span class=\"swl-format-2\">GPU\u30e1\u30e2\u30ea80GB\u3092\u642d\u8f09\u3057\u305fNVIDIA A100 80GB\u306e\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<p>\u5b9f\u884c\u74b0\u5883\u306e\u8a73\u7d30\u306f\u4ee5\u4e0b\u306e\u3068\u304a\u308a\u3067\u3059\u3002<\/p>\n\n\n\n<div class=\"wp-block-group is-style-big_icon_good\"><div class=\"wp-block-group__inner-container\">\n<ul class=\"wp-block-list\">\n<li>GPU\uff1aNVIDIA A100 80GB<\/li>\n\n\n\n<li>GPU\u30e1\u30e2\u30ea\uff08VRAM\uff09\uff1a80GB<\/li>\n\n\n\n<li>OS\uff1aUbuntu22.04<\/li>\n\n\n\n<li>Docker<\/li>\n<\/ul>\n<\/div><\/div>\n\n\n\n<h2 class=\"wp-block-heading\">Docker\u3067\u74b0\u5883\u69cb\u7bc9<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"600\" height=\"300\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/penguin_midashi.jpg\" alt=\"\u898b\u51fa\u3057\u753b\u50cf\" class=\"wp-image-6476\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/penguin_midashi.jpg 600w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/penguin_midashi-300x150.jpg 300w\" sizes=\"(max-width: 600px) 100vw, 600px\" \/><\/figure>\n\n\n\n<p><span class=\"swl-format-2\">Docker\u3092\u4f7f\u7528\u3057\u3066Sarashina\u306e\u74b0\u5883\u69cb\u7bc9\u3092\u3057\u3066\u3044\u304d\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<p>Docker\u306e\u4f7f\u3044\u65b9\u306f\u4ee5\u4e0b\u306e\u8a18\u4e8b\u3092\u3054\u89a7\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n<div class=\"swell-block-postLink\">\t\t\t<div class=\"p-blogCard -external\" data-type=\"type3\" data-onclick=\"clickLink\">\n\t\t\t\t<div class=\"p-blogCard__inner\">\n\t\t\t\t\t<span class=\"p-blogCard__caption\">EdgeHUB<\/span>\n\t\t\t\t\t<div class=\"p-blogCard__thumb c-postThumb\"><figure class=\"c-postThumb__figure\"><img decoding=\"async\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/ubuntudocker-eyecatch.jpg\" alt=\"\" class=\"c-postThumb__img u-obf-cover\" width=\"320\" height=\"180\"><\/figure><\/div>\t\t\t\t\t<div class=\"p-blogCard__body\">\n\t\t\t\t\t\t<a class=\"p-blogCard__title\" href=\"https:\/\/highreso.jp\/edgehub\/machinelearning\/ubuntudocker.html\" target=\"_blank\" rel=\"noopener noreferrer\">Ubuntu\u306bDocker\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\uff01GPU\u3092\u30b3\u30f3\u30c6\u30ca\u304b\u3089\u4f7f\u3046\u65b9\u6cd5 | EdgeHUB<\/a>\n\t\t\t\t\t\t<span class=\"p-blogCard__excerpt\">\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001Ubuntu\u306bDocker\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3059\u308b\u65b9\u6cd5\u304b\u3089\u3001\u30b3\u30f3\u30c6\u30ca\u304b\u3089GPU\u3092\u4f7f\u3046\u65b9\u6cd5\u307e\u3067\u7d39\u4ecb\u3057\u307e\u3059\u3002 GPU\u306e\u74b0\u5883\u69cb\u7bc9\u3067\u306f\u3001\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u3001\u30c9\u30e9\u30a4\u30d0\u3001\u30e9\u30a4\u30d6\u30e9\u30ea\u306e\u30d0\u30fc\u30b8&#8230;<\/span>\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\n\n<div class=\"swell-block-step\" data-num-style=\"circle\">\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">Dockerfile\u306e\u4f5c\u6210<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">Dockerfile\u306b\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3059\u308b\u30d1\u30c3\u30b1\u30fc\u30b8\u3092\u8a18\u8ff0\u3057\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<div class=\"swell-block-capbox cap_box is-style-onborder_ttl2\" data-colset=\"col3\"><div class=\"cap_box_ttl\"><span>\u30d1\u30c3\u30b1\u30fc\u30b8\u4e00\u89a7<\/span><\/div><div class=\"cap_box_content\">\n<ul class=\"wp-block-list\">\n<li>CUDA\uff1a12.1<\/li>\n\n\n\n<li>Python\uff1a3.10<\/li>\n\n\n\n<li>PyTorch\uff1a2.2.2<\/li>\n\n\n\n<li>transformers\uff1a4.41.2<\/li>\n\n\n\n<li>bitsandbytes<\/li>\n\n\n\n<li>ninja<\/li>\n\n\n\n<li>packaging<\/li>\n\n\n\n<li>accelerate<\/li>\n\n\n\n<li>datasets<\/li>\n\n\n\n<li>bitsandbytes<\/li>\n\n\n\n<li>evaluate<\/li>\n\n\n\n<li>trl<\/li>\n\n\n\n<li>peft<\/li>\n\n\n\n<li>wheel<\/li>\n\n\n\n<li>sentencepiece<\/li>\n\n\n\n<li>flash-attn<\/li>\n\n\n\n<li>JupyterLab<\/li>\n\n\n\n<li>huggingface_hub[cli]<\/li>\n\n\n\n<li>wandb<\/li>\n<\/ul>\n<\/div><\/div>\n\n\n\n<p class=\"is-style-icon_info\">CUDA\u3001PyTorch\u3001Transformer\u7b49\u306f\u3001\u30d0\u30fc\u30b8\u30e7\u30f3\u4f9d\u5b58\u95a2\u4fc2\u306b\u3088\u308b\u30a8\u30e9\u30fc\u304c\u8d77\u304d\u3084\u3059\u3044\u306e\u3067\u3001\u52d5\u4f5c\u691c\u8a3c\u6e08\u306e\u30d0\u30fc\u30b8\u30e7\u30f3\u6307\u5b9a\u3057\u3066\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p>Ubuntu\u306e\u30b3\u30de\u30f3\u30c9\u30e9\u30a4\u30f3\u304b\u3089\u3001Dockerfile\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>mkdir sarashina_instruct\ncd sarashina_instruct\nnano Dockerfile<\/code><\/pre>\n\n\n\n<p>\u6b21\u306e\u8a18\u8ff0\u3092\u30b3\u30d4\u30fc\u3057\u3066Dockerfile\u306b\u8cbc\u308a\u4ed8\u3051\u307e\u3059\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code># \u30d9\u30fc\u30b9\u30a4\u30e1\u30fc\u30b8(CUDA)\u306e\u6307\u5b9a\nFROM nvidia\/cuda:12.1.0-cudnn8-devel-ubuntu22.04\n\n# \u5fc5\u8981\u306a\u30d1\u30c3\u30b1\u30fc\u30b8\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\nRUN apt-get update &amp;&amp; apt-get install -y python3-pip python3-venv git nano\n\n# \u4f5c\u696d\u30c7\u30a3\u30ec\u30af\u30c8\u30ea\u3092\u8a2d\u5b9a\nWORKDIR \/app\n\n# \u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u30b3\u30fc\u30c9\u3092\u30b3\u30d4\u30fc\nCOPY . \/app\n\n# Python\u4eee\u60f3\u74b0\u5883\u306e\u4f5c\u6210\nRUN python3 -m venv \/app\/.venv\n\n# \u4eee\u60f3\u74b0\u5883\u3092\u30a2\u30af\u30c6\u30a3\u30d9\u30fc\u30c8\u3059\u308b\u30b3\u30de\u30f3\u30c9\u3092.bashrc\u306b\u8ffd\u52a0\nRUN echo \"source \/app\/.venv\/bin\/activate\" &gt;&gt; \/root\/.bashrc\n\n# JupyterLab,HuggingFaceHub,WandB\u306e\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\nRUN \/app\/.venv\/bin\/pip install Jupyter \\\n    jupyterlab \\\n    huggingface_hub&#091;cli] \\\n    wandb\n\n# PyTorch\u306e\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\nRUN \/app\/.venv\/bin\/pip install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https:\/\/download.pytorch.org\/whl\/cu121\n\n# Transformer\u95a2\u9023\u306e\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\nRUN \/app\/.venv\/bin\/pip install \\\n    transformers==4.41.2 \\\n    ninja \\\n    packaging \\\n    accelerate \\\n    datasets \\\n    bitsandbytes \\\n    evaluate \\\n    trl \\\n    peft \\\n    wheel \\\n    sentencepiece\n\n# Flash attention\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\nRUN \/app\/.venv\/bin\/pip install flash-attn --no-build-isolation\n\n# \u30b3\u30f3\u30c6\u30ca\u306e\u8d77\u52d5\u6642\u306bbash\u3092\u5b9f\u884c\nCMD &#091;\"\/bin\/bash\"]<\/code><\/pre>\n\n\n\n<p class=\"is-style-icon_announce\">[Ctrl + S]\u30ad\u30fc\u3067\u5909\u66f4\u5185\u5bb9\u3092\u4fdd\u5b58\u3057\u3001[Ctrl + X]\u30ad\u30fc\u3067\u7de8\u96c6\u30e2\u30fc\u30c9\u304b\u3089\u629c\u3051\u307e\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">docker-compose.yml\u30d5\u30a1\u30a4\u30eb\u306e\u4f5c\u6210<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">docker-compose.yml\u30d5\u30a1\u30a4\u30eb\u3092\u4f7f\u3063\u3066Docker\u30b3\u30f3\u30c6\u30ca\u306e\u8a2d\u5b9a\u3092\u3057\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<p>docker-compose.yml\u30d5\u30a1\u30a4\u30eb\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>nano docker-compose.yml<\/code><\/pre>\n\n\n\n<p>\u6b21\u306e\u8a18\u8ff0\u3092\u30b3\u30d4\u30fc\u3057\u3066docker-compose.yml\u306b\u8cbc\u308a\u4ed8\u3051\u307e\u3059\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>services:\n&nbsp;&nbsp;sarashina_instruct:\n&nbsp;&nbsp;&nbsp;&nbsp;build:\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;context: .\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;dockerfile: Dockerfile\n&nbsp;&nbsp;&nbsp;&nbsp;image: sarashina_instruct\n&nbsp;&nbsp;&nbsp;&nbsp;runtime: nvidia\n&nbsp;&nbsp;&nbsp;&nbsp;container_name: sarashina_instruct\n&nbsp;&nbsp;&nbsp;&nbsp;ports:\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;- \"8888:8888\"\n&nbsp;&nbsp;&nbsp;&nbsp;volumes:\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;- .:\/app\/sarashina_instruct\n&nbsp;&nbsp;&nbsp;&nbsp;deploy:\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;resources:\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;reservations:\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;devices:\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;- driver: nvidia\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;count: 1\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;capabilities: &#091;gpu]\n&nbsp;&nbsp;&nbsp;&nbsp;command: &gt;\n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;bash -c '\/app\/.venv\/bin\/jupyter lab --ip=\"*\" --port=8888 --NotebookApp.token=\"\" --NotebookApp.password=\"\" --no-browser --allow-root'<\/code><\/pre>\n\n\n\n<p class=\"is-style-icon_announce\">[Ctrl + S]\u30ad\u30fc\u3067\u5909\u66f4\u5185\u5bb9\u3092\u4fdd\u5b58\u3057\u3001[Ctrl + X]\u30ad\u30fc\u3067\u7de8\u96c6\u30e2\u30fc\u30c9\u304b\u3089\u629c\u3051\u307e\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">Docker\u30b3\u30f3\u30c6\u30ca\u3092\u8d77\u52d5<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">Dockerfile\u304b\u3089\u30d3\u30eb\u30c9\u3057\u3066\u30b3\u30f3\u30c6\u30ca\u3092\u8d77\u52d5\u3057\u307e\u3059\u3002&nbsp;&nbsp;&nbsp;<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>docker compose up<\/code><\/pre>\n\n\n\n<p>Docker\u306e\u8d77\u52d5\u5f8c\u306b<span class=\"swl-format-2\">\u30d6\u30e9\u30a6\u30b6\u306e\u691c\u7d22\u7a93\u306b&#8221;localhost:8888&#8243;\u3092\u5165\u529b<\/span>\u3059\u308b\u3068\u3001Jupyter Lab\u3092\u30d6\u30e9\u30a6\u30b6\u3067\u8868\u793a\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>localhost:8888<\/code><\/pre>\n<\/div><\/div>\n<\/div>\n\n\n\n<h2 class=\"wp-block-heading\">Sarashina\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u306e\u5b9f\u88c5<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"600\" height=\"300\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/roma_midashi.jpg\" alt=\"\u898b\u51fa\u3057\u753b\u50cf\" class=\"wp-image-6397\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/roma_midashi.jpg 600w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/roma_midashi-300x150.jpg 300w\" sizes=\"(max-width: 600px) 100vw, 600px\" \/><\/figure>\n\n\n\n<p><span class=\"swl-format-2\">Dokcer\u30b3\u30f3\u30c6\u30ca\u4e0a\u3067\u8d77\u52d5\u3057\u305fJupyter Lab\u3092\u4f7f\u3063\u3066\u3001Sarashina\u306e\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3092\u5b9f\u88c5\u3057\u3066\u3044\u304d\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<p>Jupyter Lab\u306e\u30b3\u30fc\u30c9\u30bb\u30eb\u3067\u4ee5\u4e0b\u306e\u30b3\u30de\u30f3\u30c9\u3092\u5b9f\u884c\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<div class=\"swell-block-step\" data-num-style=\"circle\">\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">\u30e9\u30a4\u30d6\u30e9\u30ea\u306e\u30a4\u30f3\u30dd\u30fc\u30c8<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">\u5fc5\u8981\u306a\u30e9\u30a4\u30d6\u30e9\u30ea\u3092\u30a4\u30f3\u30dd\u30fc\u30c8\u3057\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>import torch\nimport transformers\nfrom transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, logging\nfrom datasets import load_dataset\nfrom trl import SFTTrainer, SFTConfig\nimport wandb<\/code><\/pre>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">\u5916\u90e8\u9023\u643a\u306e\u8a2d\u5b9a<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">\u5b66\u7fd2\u30ed\u30b0\u306e\u7ba1\u7406\u3092\u3059\u308b\u305f\u3081\u306b\u3001WandB\u306b\u30ed\u30b0\u30a4\u30f3\u3057\u307e\u3059\u3002<\/span>\uff08WandB\u3092\u4f7f\u7528\u3057\u306a\u3044\u5834\u5408\u306f\u7701\u7565\u3057\u3066\u304f\u3060\u3055\u3044\u3002\uff09<\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>API_KEY = \"*************\"\nwandb.login(key=API_KEY)<\/code><\/pre>\n\n\n\n<p class=\"is-style-icon_info\">API_KEY = &#8220;*************&#8221;\u306b\u306f\u3001Wandb\u3067\u767a\u884c\u3057\u305fAPI\u30ad\u30fc\u304c\u5165\u308a\u307e\u3059\u3002<\/p>\n\n\n\n<p>Wandb\u3067API\u30ad\u30fc\u3092\u767a\u884c\u3059\u308b\u65b9\u6cd5\u3092\u4ee5\u4e0b\u306e\u8a18\u4e8b\u3067\u89e3\u8aac\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n<div class=\"swell-block-postLink\">\t\t\t<div class=\"p-blogCard -external\" data-type=\"type3\" data-onclick=\"clickLink\">\n\t\t\t\t<div class=\"p-blogCard__inner\">\n\t\t\t\t\t<span class=\"p-blogCard__caption\">EdgeHUB<\/span>\n\t\t\t\t\t<div class=\"p-blogCard__thumb c-postThumb\"><figure class=\"c-postThumb__figure\"><img decoding=\"async\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/wandbtsukaikata_eyecatch.jpg\" alt=\"\" class=\"c-postThumb__img u-obf-cover\" width=\"320\" height=\"180\"><\/figure><\/div>\t\t\t\t\t<div class=\"p-blogCard__body\">\n\t\t\t\t\t\t<a class=\"p-blogCard__title\" href=\"https:\/\/highreso.jp\/edgehub\/machinelearning\/wandbstartup.html\" target=\"_blank\" rel=\"noopener noreferrer\">WandB\u306e\u4f7f\u3044\u65b9\uff01\u30ed\u30b0\u30a4\u30f3\u65b9\u6cd5\u304b\u3089API\u30ad\u30fc\u767a\u884c\u3001\u30ed\u30b0\u53d6\u5f97\u307e\u3067\u7d39\u4ecb | EdgeHUB<\/a>\n\t\t\t\t\t\t<span class=\"p-blogCard__excerpt\">\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001WandB\u306e\u30a2\u30ab\u30a6\u30f3\u30c8\u4f5c\u6210\u304b\u3089\u4f7f\u3044\u65b9\u307e\u3067\u89e3\u8aac\u3057\u3066\u3044\u304d\u307e\u3059\u3002 WandB\u306f\u6a5f\u68b0\u5b66\u7fd2\u306e\u30ed\u30b0\u7ba1\u7406\u306b\u4fbf\u5229\u306a\u30c4\u30fc\u30eb\u3067\u3059\u3002 WandB\u3092\u5229\u7528\u3059\u308b\u3053\u3068\u3067\u3001\u30e2\u30c7\u30eb\u5b66\u7fd2\u306e\u30ed\u30b0\u53ce\u96c6\u3084&#8230;<\/span>\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div><\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">\u30e2\u30c7\u30eb\u306e\u8a2d\u5b9a<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">\u30e2\u30c7\u30eb\u3068\u30c8\u30fc\u30af\u30ca\u30a4\u30b6\u30fc\u306e\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3092\u3057\u3066\u8aad\u307f\u8fbc\u307f\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>model_id = \"sbintuitions\/sarashina2-7b\"\n\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_id,\n    torch_dtype=torch.bfloat16,\n    device_map=\"auto\",\n    attn_implementation=\"flash_attention_2\"\n)\n\ntokenizer = AutoTokenizer.from_pretrained(\n    model_id,\n    padding_side=\"right\",\n    add_eos_token=True\n)<\/code><\/pre>\n\n\n\n<div class=\"swell-block-capbox cap_box is-style-onborder_ttl2\" data-colset=\"col3\"><div class=\"cap_box_ttl\"><span>\u30b3\u30fc\u30c9\u306e\u89e3\u8aac<\/span><\/div><div class=\"cap_box_content\">\n<p><span class=\"swl-format-2\">sbintuitions\/sarashina2-7b<\/span><\/p>\n\n\n\n<p>Sarashina 7B\u306e\u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb\u3092\u6307\u5b9a\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">AutoModelForCausalLM.from_pretrained<\/span><\/p>\n\n\n\n<p>\u30e2\u30c7\u30eb\u306e\u8aad\u307f\u8fbc\u3080\u8a2d\u5b9a\u3092\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">torch.bfloat16<\/span><\/p>\n\n\n\n<p>Bfloat16\u306f\u3001FP32\u3068\u540c\u3058\u6570\u5024\u7bc4\u56f2\u3092\u6301\u3061\u306a\u304c\u3089\u3001GPU\u30e1\u30e2\u30ea\u3092\u7bc0\u7d04\u3067\u304d\u3001\u8a08\u7b97\u901f\u5ea6\u3082\u5411\u4e0a\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">attn_implementation=&#8221;flash_attention_2&#8243;<\/span><\/p>\n\n\n\n<p>Transformer\u306e\u51e6\u7406\u3092\u52b9\u7387\u5316\u3057\u3066\u5b66\u7fd2\u3092\u9ad8\u901f\u3059\u308b\u30e9\u30a4\u30d6\u30e9\u30ea\u3092\u6307\u5b9a\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">AutoTokenizer.from_pretrained<\/span><\/p>\n\n\n\n<p>\u30c8\u30fc\u30af\u30ca\u30a4\u30b6\u30fc\u306e\u8aad\u307f\u8fbc\u307f\u8a2d\u5b9a\u3092\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n<p class=\"is-style-icon_info\">\u30e2\u30c7\u30eb\u306e\u8aad\u307f\u8fbc\u307f\u6642\u70b9\u3067\u3001GPU\u30e1\u30e2\u30ea\u3092\u6d88\u8cbb\u3057\u307e\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u524d\u306e\u30e2\u30c7\u30eb\u3092\u30c6\u30b9\u30c8<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u524d\u306e\u30e2\u30c7\u30eb\u3092\u4f7f\u3063\u3066\u3001\u30c6\u30ad\u30b9\u30c8\u751f\u6210\u306e\u30c6\u30b9\u30c8\u3092\u3057\u3066\u307f\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>prompt = \"\u9999\u5ddd\u770c\u306e\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u3092\u6559\u3048\u3066\u4e0b\u3055\u3044\"\n\npipeline = transformers.pipeline(\n    \"text-generation\",\n    model=model,\n    tokenizer=tokenizer\n)\n\noutput = pipeline(\n    prompt,\n    max_length=256,\n    truncation=True,\n    do_sample=True,\n    pad_token_id=tokenizer.pad_token_id,\n    num_return_sequences=3,\n)\n\nprint(output&#091;0]&#091;'generated_text'])<\/code><\/pre>\n\n\n<div class=\"swell-block-balloon\"><div class=\"c-balloon -bln-right\" data-col=\"green\"><div class=\"c-balloon__body -speaking -border-none\"><div class=\"c-balloon__text\">\n<p>\u9999\u5ddd\u770c\u306e\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u3092\u6559\u3048\u3066\u4e0b\u3055\u3044\u3002<\/p>\n<span class=\"c-balloon__shapes\"><span class=\"c-balloon__before\"><\/span><span class=\"c-balloon__after\"><\/span><\/span><\/div><\/div><\/div><\/div>\n\n<div class=\"swell-block-balloon\"><div class=\"c-balloon -bln-left\" data-col=\"gray\"><div class=\"c-balloon__body -speaking -border-none\"><div class=\"c-balloon__text\">\n<p>\u9999\u5ddd\u770c\u306e\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u3092\u6559\u3048\u3066\u4e0b\u3055\u3044\u3002\u3088\u308d\u3057\u304f\u304a\u9858\u3044\u81f4\u3057\u307e\u3059\u3002 \u9999\u5ddd\u770c\u306e\u540d\u7269\u6599\u7406 \u3053\u3093\u3070\u3093\u306f\u3001\u79c1\u306f\u9999\u5ddd\u770c\u306b\u4f4f\u3093\u3067\u3044\u308b\u306e\u3067\u3059\u304c\u3001\u9999\u5ddd\u770c\u306e\u6709\u540d\u306a\u6599\u7406\u3063\u3066\u4f55\u3067\u3057\u3087\u3046\u304b\uff1f \u3054\u5b58\u77e5\u306e\u65b9\u306f\u6559\u3048\u3066\u4e0b\u3055\u3044\u3002 \u3088\u308d\u3057\u304f\u304a\u9858\u3044\u3057\u307e\u3059\u3002 \u9999\u5ddd\u770c\u306e\u540d\u7269 \u3053\u3093\u306b\u3061\u306f\u3001\u9999\u5ddd\u770c\u306e\u540d\u7269\u3063\u3066\u4f55\u304c\u3042\u308a\u307e\u3059\u304b\uff1f \u56de\u7b54\u3092\u3088\u308d\u3057\u304f\u304a\u9858\u3044\u3057\u307e\u3059\u3002 \u9999\u5ddd\u770c\u306e\u304a\u30b9\u30b9\u30e1\u30b0\u30eb\u30e1\u3092\u6559\u3048\u3066\u304f\u3060\u3055\u3044\uff01 \uff14\u6708\u4e2d\u65ec\u306b\u9999\u5ddd\u770c\u306b\u884c\u304f\u3053\u3068\u306b\u306a\u308a\u3001\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u3092\u98df\u3079\u308b\u3053\u3068\u3092\u697d\u3057\u307f\u306b\u3057\u3066\u304a\u308a\u307e\u3059\u3002 \u9999\u5ddd\u770c\u306f\u3046\u3069\u3093\u4ee5\u5916\u306b\u3082\u3001\u9aa8\u4ed8\u304d\u9ce5\u3084\u30aa\u30ea\u30fc\u30d6\u725b\u3001\u3042\u3093\u3082\u3061\u96d1\u716e\u306a\u3069\u3001\u305f\u304f\u3055\u3093\u306e\u30b0\u30eb\u30e1\u304c\u3042\u308b\u3068\u805e\u3044\u305f\u306e\u3067\u3059\u304c\u3001\u3053\u306e\u4e2d\u304b\u3089\u79c1\u304c\u98df\u3079\u305f\u3053\u3068\u304c\u3042\u308b\u3082\u306e\u306f\u3001\u9aa8\u4ed8\u304d\u9ce5\u3068\u3001\u3042\u3093\u3082\u3061\u96d1\u716e\u3001\u3042\u3068\u6c34\u3088\u3046\u304b\u3093\u3067\u3059\u3002 \u3053\u306e\u4e2d\u3067\u3001\u79c1\u304c\u98df\u3079\u305f\u3053\u3068\u304c\u306a\u3044\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u306e\u30aa\u30b9\u30b9\u30e1\u304c\u3042\u308a\u307e\u3057\u305f\u3089\u3001\u305c\u3072\u6559\u3048\u3066\u304f\u3060\u3055\u3044\uff01 \u9999\u5ddd\u30b0\u30eb\u30e1 \u9999\u5ddd\u306e\u540d\u7269\u30b0\u30eb\u30e1\u3068\u3044\u3048\u3070\u3046\u3069\u3093\u304c\u7b46\u982d\u306b\u3042\u304c\u308a\u307e\u3059\u304c\uff0c \u305d\u308c\u4ee5\u5916\u3067\uff0c\u9999\u5ddd\u306e\u304a\u3044\u3057\u3044\u3082\u306e\u3068\u3044\u3048\u3070\u4f55\u3067\u3057\u3087\u3046\uff1f \u8b83\u5c90\u725b\uff0c\u9aa8\u4ed8\u304d\u9ce5\uff0c\u3057\u3087\u3046\u3086\u8c46\uff0c\u3042\u3093\u9905\u96d1\u716e\uff08\u3053\u308c\u306f\u5ca1\u5c71\u306e\u90f7\u571f\u6599\u7406\uff1f\uff09\uff0c\u6817\u6797\u306e\u30a2\u30a4\u30b9\u30af\u30ea\u30f3\uff0c\u4e00\u9db4\u306e\u3068\u308a\u3081\u3057\u3050\u3089\u3044\u3057\u304b\u77e5\u3089\u306a\u3044\u306e\u3067\u3059\u3002 \u4f55\u304b\u304a\u3044\u3057\u3044\u3082\u306e\u3042\u308a\u307e\u305b\u3093\u304b\uff1f \u9999\u5ddd\u770c\u306e\u540d\u7269\u3000\u9aa8\u4ed8\u9ce5\u30fb\u3042\u3093\u3082\u3061\u96d1\u716e\u3000\u306b\u3064\u3044\u3066\u8cea\u554f \u9999\u5ddd\u770c\u3067\u6709\u540d\u306a \u9aa8\u4ed8\u9ce5 \u3042\u3093\u3082\u3061\u96d1\u716e \u306e\u5e97\u3067 \u304a\u52e7\u3081\u306e\u3068\u3053\u304c\u3042\u3063\u305f\u3089\u6559\u3048\u3066\u4e0b\u3055\u3044\u3002 \u9999\u5ddd\u306e\u540d\u7269\u3068\u3044\u3048\u3070\u30fb\u30fb\u30fb \u9999\u5ddd\u306e\u540d\u7269\u3068\u3044\u3048\u3070\u3001\u306a\u3093\u3068\u3044\u3063\u3066\u3082\u8b83\u5c90\u3046\u3069\u3093\uff01\u3067\u3059\u3088\u306d\u3002\u3067\u3082\u8b83\u5c90\u3046\u3069\u3093\u3057\u304b<\/p>\n<span class=\"c-balloon__shapes\"><span class=\"c-balloon__before\"><\/span><span class=\"c-balloon__after\"><\/span><\/span><\/div><\/div><\/div><\/div>\n\n\n<p class=\"is-style-big_icon_point\">Sarashina\u306e\u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb\u3067\u306f\u3001\u30c1\u30e3\u30c3\u30c8\u5f62\u5f0f\u306e\u5fdc\u7b54\u304c\u3067\u304d\u3066\u3044\u306a\u3044\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002<br>\u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb\u306f\u4eba\u9593\u306e\u610f\u56f3\u306b\u6cbf\u3063\u305f\u5fdc\u7b54\u304c\u3067\u304d\u306a\u3044\u305f\u3081\u3001\u6307\u793a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u304c\u5fc5\u8981\u3067\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u8a2d\u5b9a<\/div><div class=\"swell-block-step__body\">\n<p>\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u306b\u4f7f\u7528\u3059\u308b<span class=\"swl-format-2\">\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u300ckunishou\/databricks-dolly-15k-ja\u300d\u3092\u8aad\u307f\u8fbc\u307f\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>dataset = load_dataset(\"kunishou\/databricks-dolly-15k-ja\", split=\"train\")<\/code><\/pre>\n\n\n\n<p>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304b\u3089<span class=\"swl-format-2\">\u300cinstruction\uff08\u6307\u793a\uff09\u300d\u3068\u300coutput\uff08\u5fdc\u7b54\uff09\u300d\u3092\u62bd\u51fa\u3057\u3066\u3001\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u306b\u5909\u63db<\/span>\u3057\u3066\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u66f4\u65b0\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>template = \"{% for message in messages %}{{'&lt;|im_start|&gt;' + message&#091;'role'] + '\\n' + message&#091;'content'] + '&lt;|im_end|&gt;' + '\\n'}}{% endfor %}\"\n\ntokenizer.chat_template = template\n\ndef formatting_func(example):\n    messages = &#091;\n        {'role': \"system\", 'content': \"\u3042\u306a\u305f\u306f\u65e5\u672c\u8a9e\u3067\u56de\u7b54\u3059\u308b\u30a2\u30b7\u30b9\u30bf\u30f3\u30c8\u3067\u3059\u3002\"},\n        {'role': \"user\", 'content': example&#091;\"instruction\"]},\n        {'role': \"assistant\", 'content': example&#091;\"output\"]}\n    ]\n    return tokenizer.apply_chat_template(messages, tokenize=False)\n\ndef update_dataset(example):\n    example&#091;\"text\"] = formatting_func(example)\n    for field in &#091;\"index\", \"category\", \"instruction\", \"input\", \"output\"]:\n        example.pop(field, None)\n    return example\n\ndataset = dataset.map(update_dataset)\n\nprint(dataset&#091;312]&#091;\"text\"])<\/code><\/pre>\n\n\n\n<p class=\"is-style-balloon_box\">\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u5909\u63db\u5f8c\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u4f8b<\/p>\n\n\n\n<pre class=\"wp-block-code has-swl-gray-background-color has-background\"><code>&lt;|im_start|&gt;system\n\u3042\u306a\u305f\u306f\u65e5\u672c\u8a9e\u3067\u56de\u7b54\u3059\u308b\u30a2\u30b7\u30b9\u30bf\u30f3\u30c8\u3067\u3059\u3002&lt;|im_end|&gt;\n&lt;|im_start|&gt;user\n\u826f\u3044\u5199\u771f\u3092\u64ae\u308b\u305f\u3081\u306e\u30dd\u30a4\u30f3\u30c8\u3068\u306f\uff1f&lt;|im_end|&gt;\n&lt;|im_start|&gt;assistant\n\u826f\u3044\u5199\u771f\u306b\u306f\u3001\u5149\u3001\u88ab\u5199\u4f53\u3001\u69cb\u56f3\u3068\u3044\u30463\u3064\u306e\u91cd\u8981\u306a\u8981\u7d20\u304c\u3042\u308a\u307e\u3059\u3002\u512a\u308c\u305f\u5199\u771f\u306f\u3001\u30d4\u30f3\u30c8\u304c\u5408\u3063\u3066\u3044\u3066\u3001\u30b7\u30e3\u30fc\u30d7\u3067\u3001\u9732\u51fa\uff08\u5149\uff09\u304c\u3088\u304f\u3001\u69cb\u56f3\u304c\u6c7a\u307e\u3063\u3066\u3044\u307e\u3059\u3002  \u5b8c\u74a7\u306a\u5199\u771f\u3092\u64ae\u308b\u305f\u3081\u306b\u306f\u3001\u6b63\u3057\u3044\u7126\u70b9\u8ddd\u96e2\u3067\u64ae\u5f71\u3057\u3001\u30d5\u30ec\u30fc\u30e0\u5185\u306e\u88ab\u5199\u4f53\u306b\u9069\u3057\u305f\u7d5e\u308a\u3092\u8a2d\u5b9a\u3057\u3001\u64ae\u5f71\u3059\u308b\u30a2\u30af\u30b7\u30e7\u30f3\u306b\u5408\u308f\u305b\u3066\u30b7\u30e3\u30c3\u30bf\u30fc\u30b9\u30d4\u30fc\u30c9\u3092\u8a2d\u5b9a\u3057\uff08\u30b9\u30dd\u30fc\u30c4\u3067\u306f\u9ad8\u3044\u30b7\u30e3\u30c3\u30bf\u30fc\u30b9\u30d4\u30fc\u30c9\u3001\u30dd\u30fc\u30c8\u30ec\u30fc\u30c8\u3067\u306f\u4f4e\u3044\u30b7\u30e3\u30c3\u30bf\u30fc\u30b9\u30d4\u30fc\u30c9\uff09\u3001\u30b7\u30e3\u30c3\u30bf\u30fc\u30b9\u30d4\u30fc\u30c9\u3068\u7d5e\u308a\u306e\u4e21\u65b9\u306b\u5bfe\u5fdc\u3059\u308b\u3088\u3046\u306bISO\u304c\u6b63\u3057\u304f\u8a2d\u5b9a\u3055\u308c\u3066\u3044\u308b\u3053\u3068\u3092\u78ba\u8a8d\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002  ISO\u3001\u7d5e\u308a\u3001\u30b7\u30e3\u30c3\u30bf\u30fc\u30b9\u30d4\u30fc\u30c9\u304c\u3069\u306e\u3088\u3046\u306b\u9023\u52d5\u3057\u3066\u3044\u308b\u304b\u3092\u7406\u89e3\u3059\u308b\u3053\u3068\u304c\u3001\u5b8c\u74a7\u306a\u5199\u771f\u3092\u64ae\u308b\u305f\u3081\u306e\u9375\u3068\u306a\u308a\u307e\u3059\u3002&lt;|im_end|&gt;<\/code><\/pre>\n\n\n\n<p class=\"is-style-icon_info\">\u30e2\u30c7\u30eb\u5b66\u7fd2\u306e\u524d\u306b\u306f\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u306b\u5909\u63db\u3057\u3066\u304a\u304f\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n<div class=\"swell-block-capbox cap_box is-style-onborder_ttl2\" data-colset=\"col3\"><div class=\"cap_box_ttl\"><span>\u30b3\u30fc\u30c9\u89e3\u8aac<\/span><\/div><div class=\"cap_box_content\">\n<p><span class=\"swl-format-2\">tokenizer.chat_template = template<\/span><\/p>\n\n\n\n<p>Tokenizer\u306b\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u304c\u5b9a\u7fa9\u3055\u308c\u3066\u3044\u306a\u3044\u305f\u3081\u3001\u65b0\u3057\u304f\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u3092\u4f5c\u6210\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">def formatting_func(example):<\/span><\/p>\n\n\n\n<p>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u300cinstruction\u300d\u3068\u300coutput\u300d\u3092\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u306b\u5909\u63db\u3059\u308b\u95a2\u6570\u3092\u5b9a\u7fa9\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">def update_dataset(example):<\/span><\/p>\n\n\n\n<p>\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u306b\u5909\u63db\u3057\u305f\u30c7\u30fc\u30bf\u3092\u300ctext\u300d\u30d5\u30a3\u30fc\u30eb\u30c9\u306b\u4ee3\u5165\u3057\u3001\u305d\u306e\u4ed6\u306e\u4e0d\u8981\u306a\u30d5\u30a3\u30fc\u30eb\u30c9\u3092\u524a\u9664\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8<\/span><\/p>\n\n\n\n<p>\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u306e\u8a18\u53f7\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u610f\u5473\u3092\u6301\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code has-swl-gray-background-color has-background\"><code>&lt;|im_start|&gt;system...&lt;|im_end|&gt;:\u30b7\u30b9\u30c6\u30e0\u30e1\u30c3\u30bb\u30fc\u30b8\u304c\u5165\u308a\u307e\u3059\u3002\n&lt;|im_start|&gt;user...&lt;|im_end|&gt;:\u30d7\u30ed\u30f3\u30d7\u30c8\uff08\u6307\u793a\uff09\u304c\u5165\u308a\u307e\u3059\u3002\n&lt;|im_start|&gt;assistant...&lt;|im_end|&gt;\uff1a\u30ec\u30b9\u30dd\u30f3\u30b9(\u5fdc\u7b54)\u304c\u5165\u308a\u307e\u3059\u3002  <\/code><\/pre>\n<\/div><\/div>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u8a2d\u5b9a<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u8a2d\u5b9a\u3057\u3066\u3044\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>sft_config = SFTConfig(\n    output_dir=\".\/results\",\n    bf16=True,\n    per_device_train_batch_size=4,\n    gradient_accumulation_steps=16,\n    num_train_epochs=2,\n    optim=\"adamw_torch_fused\",\n    learning_rate=2e-4,\n    lr_scheduler_type=\"cosine\",\n    warmup_steps=100,\n    weight_decay=0.01,\n    logging_steps=10,\n    save_strategy= \"epoch\", \n    group_by_length=True,\n    dataset_text_field=\"text\",\n    max_seq_length=512,\n    packing=True,\n    report_to=\"wandb\" \n)<\/code><\/pre>\n\n\n\n<div class=\"swell-block-capbox cap_box is-style-onborder_ttl2\" data-colset=\"col3\"><div class=\"cap_box_ttl\"><span>\u30b3\u30fc\u30c9\u306e\u89e3\u8aac<\/span><\/div><div class=\"cap_box_content\">\n<p><span class=\"swl-format-2\">per_device_train_batch_size<\/span><\/p>\n\n\n\n<p>\u4e00\u5ea6\u306b\u51e6\u7406\u3059\u308b\u30d0\u30c3\u30c1\u306e\u30b5\u30a4\u30ba\u3092\u6307\u5b9a\u3057\u307e\u3059\u3002GPU\u30e1\u30e2\u30ea\u304c\u4e0d\u8db3\u3059\u308b\u5834\u5408\u306f\u3001\u5024\u3092\u5c0f\u3055\u304f\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">gradient_accumulation_steps<\/span><\/p>\n\n\n\n<p>\u52fe\u914d\u7d2f\u7a4d\u3002\u3053\u306e\u5024\u3092\u5927\u304d\u304f\u3059\u308b\u3053\u3068\u3067\u3001\u64ec\u4f3c\u7684\u306b\u30df\u30cb\u30d0\u30c3\u30c1\u306e\u30b5\u30a4\u30ba\u3092\u5927\u304d\u304f\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">group_by_length=True<\/span><\/p>\n\n\n\n<p>\u540c\u3058\u9577\u3055\u306e\u30b7\u30fc\u30b1\u30f3\u30b9\u3092\u307e\u3068\u3081\u3066\u30d0\u30c3\u30c1\u5316\u3057\u3066\u3001\u30e1\u30e2\u30ea\u3092\u7bc0\u7d04\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">report_to=&#8221;wandb&#8221;<\/span><\/p>\n\n\n\n<p>WandB\u306b\u30ed\u30b0\u3092\u4e3b\u529b\u3057\u307e\u3059\u3002WandB\u3092\u4f7f\u7528\u3057\u306a\u3044\u5834\u5408\u306f\u3001\u30b3\u30e1\u30f3\u30c8\u30a2\u30a6\u30c8\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<\/div><\/div>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">SFTTrainer\u306e\u8a2d\u5b9a\u3001\u5b66\u7fd2\u306e\u5b9f\u884c<\/div><div class=\"swell-block-step__body\">\n<p><span class=\"swl-format-2\">SFTTrainer\u3092\u4f7f\u3063\u3066\u6559\u5e2b\u3042\u308a\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\uff08Supervised Fine-tuning\uff09\u3092\u5b9f\u884c\u3057\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>new_model=\"sarashina2-7b-instruct\"\n\ntrainer = SFTTrainer(\n    model=model,\n    tokenizer=tokenizer,\n    train_dataset=dataset,\n    args=sft_config,\n)\n\nwandb.init(project=\"sarashina2-7b-instruct\")\ntrainer.train()\ntrainer.model.save_pretrained(new_model)<\/code><\/pre>\n\n\n\n<div class=\"swell-block-capbox cap_box is-style-onborder_ttl2\" data-colset=\"col3\"><div class=\"cap_box_ttl\"><span>\u30b3\u30fc\u30c9\u306e\u89e3\u8aac<\/span><\/div><div class=\"cap_box_content\">\n<p><span class=\"swl-format-2\">SFTTrainer(&#8230;<\/span><\/p>\n\n\n\n<p>\u4e8b\u524d\u306b\u8a2d\u5b9a\u3057\u305f\u30e2\u30c7\u30eb\u3084\u30c8\u30fc\u30af\u30ca\u30a4\u30b6\u30fc\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3001\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u7b49\u3092SFTTrainer\u306b\u6e21\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">wandb.init(&#8230;<\/span><\/p>\n\n\n\n<p>WandB\u306e\u8a18\u9332\u3092\u958b\u59cb\u3057\u307e\u3059\u3002WandB\u3068\u9023\u643a\u3057\u306a\u3044\u5834\u5408\u306f\u30b3\u30e1\u30f3\u30c8\u30a2\u30a6\u30c8\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<\/div><\/div>\n<\/div><\/div>\n\n\n\n<div class=\"swell-block-step__item\"><div class=\"swell-block-step__number u-bg-main\"><span class=\"__label\">STEP<\/span><\/div><div class=\"swell-block-step__title u-fz-l\">\u30e1\u30c8\u30ea\u30af\u30b9\u306e\u78ba\u8a8d<\/div><div class=\"swell-block-step__body\">\n<p>WandB\u306b\u4fdd\u5b58\u3057\u305f\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u5b9f\u884c\u4e2d\u306e\u30e1\u30c8\u30ea\u30af\u30b9\u3092\u78ba\u8a8d\u3057\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">Loss\uff08\u640d\u5931\uff09\u306f\u5b66\u7fd2\u304c\u9032\u3080\u306b\u3064\u308c\u3066\u5c0f\u3055\u304f\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"800\" height=\"342\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_3.jpg\" alt=\"loss\" class=\"wp-image-6894\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_3.jpg 800w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_3-300x128.jpg 300w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_3-768x328.jpg 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/><figcaption class=\"wp-element-caption\">(\u51fa\u5178\uff1ahttps:\/\/www.wandb.jp\/)<\/figcaption><\/figure>\n\n\n\n<p><span class=\"swl-format-2\">GPU\u30e1\u30e2\u30ea\u306f\u3001\u30d4\u30fc\u30af91%(72GB)\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img decoding=\"async\" width=\"800\" height=\"443\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_4.jpg\" alt=\"(\u51fa\u5178\uff1ahttps:\/\/www.wandb.jp\/)\" class=\"wp-image-6897\" style=\"width:783px;height:auto\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_4.jpg 800w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_4-300x166.jpg 300w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/07\/sarashinafinetune_4-768x425.jpg 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/><figcaption class=\"wp-element-caption\">(\u51fa\u5178\uff1ahttps:\/\/wandb.ai\/)<\/figcaption><\/figure>\n\n\n\n<p class=\"is-style-icon_announce\">\u5b66\u7fd2\u306f30\u5206\u7a0b\u5ea6\u3067\u5b8c\u4e86\u3057\u307e\u3057\u305f<\/p>\n<\/div><\/div>\n<\/div>\n\n\n\n<h2 class=\"wp-block-heading\">\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u5f8c\u306e\u30e2\u30c7\u30eb\u3067\u30c6\u30ad\u30b9\u30c8\u751f\u6210<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"600\" height=\"300\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digitaltown_midashi.jpg\" alt=\"\u898b\u51fa\u3057\u753b\u50cf\" class=\"wp-image-6277\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digitaltown_midashi.jpg 600w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/digitaltown_midashi-300x150.jpg 300w\" sizes=\"(max-width: 600px) 100vw, 600px\" \/><\/figure>\n\n\n\n<p>\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u5f8c\u306e\u30e2\u30c7\u30eb\u3067\u30c6\u30ad\u30b9\u30c8\u751f\u6210\u3092\u3057\u3066\u3044\u304d\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u5f8c\u306e\u30e2\u30c7\u30eb\u3092\u8aad\u307f\u8fbc\u307f\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>torch.cuda.empty_cache()\n\nft_model = AutoModelForCausalLM.from_pretrained(\n    new_model, \n    device_map='auto',\n    torch_dtype=torch.bfloat16,\n    attn_implementation=\"flash_attention_2\"\n)\n\ntokenizer = AutoTokenizer.from_pretrained(\n    model_id\n)\n\ntokenizer.chat_template = template<\/code><\/pre>\n\n\n\n<div class=\"swell-block-capbox cap_box is-style-onborder_ttl2\" data-colset=\"col3\"><div class=\"cap_box_ttl\"><span>\u30b3\u30fc\u30c9\u306e\u89e3\u8aac<\/span><\/div><div class=\"cap_box_content\">\n<p><span class=\"swl-format-2\">torch.cuda.empty_cache()<\/span><\/p>\n\n\n\n<p>\u30c6\u30ad\u30b9\u30c8\u751f\u6210\u3092\u5b9f\u884c\u3059\u308b\u524d\u306b\u3001\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u4f7f\u7528\u3057\u3066\u3044\u305fGPU\u30e1\u30e2\u30ea\u3092\u30ea\u30bb\u30c3\u30c8\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n<p><span class=\"swl-format-2\">AutoModelForCausalLM.from_pretrained(new_model,&#8230;<\/span><\/p>\n\n\n\n<p>\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u5f8c\u306e\u30e2\u30c7\u30eb\u3092\u8aad\u307f\u8fbc\u3093\u3067\u3044\u307e\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n<p><span class=\"swl-format-2\">\u300c\u9999\u5ddd\u770c\u306e\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u3092\u6559\u3048\u3066\u4e0b\u3055\u3044\u3002\u300d\u3068\u3044\u3046\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u5b9f\u884c\u3057\u307e\u3059\u3002<\/span><\/p>\n\n\n\n<pre class=\"wp-block-code has-black-background-color has-background\"><code>messages = &#091;\n    {'role': \"system\",'content': \"\u3042\u306a\u305f\u306f\u65e5\u672c\u8a9e\u3067\u56de\u7b54\u3059\u308b\u30a2\u30b7\u30b9\u30bf\u30f3\u30c8\u3067\u3059\u3002\"},\n    {\"role\": \"user\", \"content\": \"\u9999\u5ddd\u770c\u306e\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u3092\u6559\u3048\u3066\u4e0b\u3055\u3044\"}\n]\n\ninput_ids = tokenizer.apply_chat_template(\n    messages,\n    add_generation_prompt=True,\n    return_tensors=\"pt\"\n).to(ft_model.device)\n\nterminators = &#091;\n    tokenizer.eos_token_id,\n    tokenizer.convert_tokens_to_ids(\"&lt;|im_end|&gt;\")\n]\n\noutputs = ft_model.generate(\n    input_ids,\n    max_new_tokens=512,\n    eos_token_id=terminators,\n    pad_token_id=tokenizer.pad_token_id,\n    do_sample=True,\n    repetition_penalty=1.2,\n    temperature=0.6,\n    top_p=0.9,\n)\n\nresponse = outputs&#091;0]&#091;input_ids.shape&#091;-1]:]\nprint(tokenizer.decode(response, skip_special_tokens=True))<\/code><\/pre>\n\n\n<div class=\"swell-block-balloon\"><div class=\"c-balloon -bln-right\" data-col=\"green\"><div class=\"c-balloon__body -speaking -border-none\"><div class=\"c-balloon__text\">\n<p>\u9999\u5ddd\u770c\u306e\u3054\u5f53\u5730\u30b0\u30eb\u30e1\u3092\u6559\u3048\u3066\u4e0b\u3055\u3044\u3002<\/p>\n<span class=\"c-balloon__shapes\"><span class=\"c-balloon__before\"><\/span><span class=\"c-balloon__after\"><\/span><\/span><\/div><\/div><\/div><\/div>\n\n<div class=\"swell-block-balloon\"><div class=\"c-balloon -bln-left\" data-col=\"gray\"><div class=\"c-balloon__body -speaking -border-none\"><div class=\"c-balloon__text\">\n<p>&lt;|im_start|&gt;assistant<br>\u9999\u5ddd\u770c\u306e\u30ed\u30fc\u30ab\u30eb\u30d5\u30fc\u30c9\u3068\u3044\u3048\u3070\u3001\u300c\u8b83\u5c90\u3046\u3069\u3093\u300d\u3092\u601d\u3044\u6d6e\u304b\u3079\u308b\u4eba\u304c\u591a\u3044\u3068\u601d\u3044\u307e\u3059\u3002\u3057\u304b\u3057\u3001\u305d\u308c\u3060\u3051\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u9aa8\u4ed8\u304d\u9ce5\u3084\u3057\u3087\u3046\u3086\u8c46\u3001\u30aa\u30ea\u30fc\u30d6\u3084\u30ec\u30e2\u30f3\u3092\u4f7f\u3063\u305f\u6599\u7406\u306a\u3069\u3001\u3055\u307e\u3056\u307e\u306a\u6599\u7406\u304c\u3042\u308a\u307e\u3059\u3002&lt;|im_end|&gt;<\/p>\n<span class=\"c-balloon__shapes\"><span class=\"c-balloon__before\"><\/span><span class=\"c-balloon__after\"><\/span><\/span><\/div><\/div><\/div><\/div>\n\n\n<p class=\"is-style-icon_good\">\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u3001\u30c1\u30e3\u30c3\u30c8\u5f62\u5f0f\u3067\u306e\u5fdc\u7b54\u304c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\uff01<\/p>\n\n\n\n<p class=\"is-style-icon_info\">&lt;|im_start|&gt;assistant&#8230;&lt;|im_end|&gt;\u306f\u3001\u5fdc\u7b54\u306e\u30c1\u30e3\u30c3\u30c8\u30c6\u30f3\u30d7\u30ec\u30fc\u30c8\u69cb\u6587\u3067\u3059\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u751f\u6210AI\u30fbLLM\u306e\u30b3\u30b9\u30c8\u3067\u304a\u56f0\u308a\u306a\u3089<\/h2>\n\n\n\n<p>GPU\u306e\u30b9\u30da\u30c3\u30af\u4e0d\u8db3\u3067\u751f\u6210AI\u306e\u958b\u767a\u304c\u601d\u3046\u3088\u3046\u306b\u9032\u307e\u306a\u3044\u3053\u3068\u306f\u3042\u308a\u307e\u305b\u3093\u304b\uff1f<\/p>\n\n\n\n<p>\u305d\u3093\u306a\u3068\u304d\u306b\u306f\u3001\u9ad8\u6027\u80fd\u306aGPU\u3092\u30ea\u30fc\u30ba\u30ca\u30d6\u30eb\u306a\u4fa1\u683c\u3067\u4f7f\u3048\u308bGPU\u30af\u30e9\u30a6\u30c9\u30b5\u30fc\u30d3\u30b9\u304c\u304a\u3059\u3059\u3081\u3067\u3059\uff01<\/p>\n\n\n\n<div class=\"p-blogParts post_content\" data-partsID=\"4543\">\n<div class=\"swell-block-capbox cap_box\" data-colset=\"col2\"><div class=\"cap_box_ttl\"><span>GPUSOROBAN<\/span><\/div><div class=\"cap_box_content\">\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"800\" height=\"283\" src=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/gpusorobanpr_1.jpg\" alt=\"GPUSOROBAN\" class=\"wp-image-4552\" srcset=\"https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/gpusorobanpr_1.jpg 800w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/gpusorobanpr_1-300x106.jpg 300w, https:\/\/highreso.jp\/edgehub\/wp-content\/uploads\/2024\/06\/gpusorobanpr_1-768x272.jpg 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/><\/figure>\n\n\n\n<p>GPUSOROBAN\u306f\u3001<span class=\"swl-format-1\"><span class=\"swl-marker mark_yellow\">\u751f\u6210AI\u30fbLLM\u5411\u3051\u306e\u9ad8\u901fGPU\u3092\u696d\u754c\u6700\u5b89\u7d1a\u306e\u6599\u91d1\u3067\u4f7f\u7528\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/span><\/span><\/p>\n\n\n\n<p>\u30a4\u30f3\u30bf\u30fc\u30cd\u30c3\u30c8\u74b0\u5883\u3055\u3048\u3042\u308c\u3070\u3001\u30af\u30e9\u30a6\u30c9\u74b0\u5883\u306eGPU\u30b5\u30fc\u30d0\u30fc\u3092\u3059\u3050\u306b\u5229\u7528\u53ef\u80fd\u3067\u3059\u3002<\/p>\n\n\n\n<p>\u5927\u898f\u6a21\u306a\u8a2d\u5099\u6295\u8cc7\u306e\u5fc5\u8981\u304c\u306a\u304f\u3001\u7169\u96d1\u306a\u30b5\u30fc\u30d0\u30fc\u7ba1\u7406\u304b\u3089\u3082\u89e3\u653e\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n<div class=\"swell-block-columns\"><div class=\"swell-block-columns__inner\">\n<div class=\"swell-block-column swl-has-mb--s\">\n<div class=\"swell-block-button blue_ -size-l is-style-btn_normal\"><a href=\"https:\/\/soroban.highreso.jp\/?utm_medium=edgehub&amp;utm_source=edgehub&amp;utm_campaign=highend_boxa&amp;adid=edgehub\" class=\"swell-block-button__link\"><span>\u516c\u5f0f\u30b5\u30a4\u30c8\u3092\u307f\u308b<\/span><\/a><\/div>\n<\/div>\n\n\n\n<div class=\"swell-block-column swl-has-mb--s\">\n<div class=\"swell-block-button red_ -size-l is-style-btn_normal\"><a href=\"https:\/\/soroban.highreso.jp\/download\/compute-catalog\/?utm_medium=edgehub&amp;utm_source=edgehub&amp;utm_campaign=highend_boxa&amp;adid=edgehub\" class=\"swell-block-button__link\"><span>\u8cc7\u6599\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9<\/span><\/a><\/div>\n<\/div>\n<\/div><\/div>\n<\/div><\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001Sarashina(\u30b5\u30e9\u30b7\u30ca)\u306e\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002 Sarashina\u306f\u30bd\u30d5\u30c8\u30d0\u30f3\u30af\u306e\u5b50\u4f1a\u793e\u306eSB Intuitions\u304c\u958b\u767a\u3057\u305f\u65e5\u672c\u8a9e\u306b\u5f37\u3044LLM\u3067\u3059\u3002 Sarashina\u306e\u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":6757,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"swell_btn_cv_data":"","footnotes":""},"categories":[20],"tags":[],"class_list":["post-6712","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-machinelearning"],"_links":{"self":[{"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/posts\/6712","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/comments?post=6712"}],"version-history":[{"count":3,"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/posts\/6712\/revisions"}],"predecessor-version":[{"id":7932,"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/posts\/6712\/revisions\/7932"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/media\/6757"}],"wp:attachment":[{"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/media?parent=6712"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/categories?post=6712"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/highreso.jp\/edgehub\/wp-json\/wp\/v2\/tags?post=6712"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}