{"id":959,"date":"2023-07-18T06:00:00","date_gmt":"2023-07-17T21:00:00","guid":{"rendered":"https:\/\/staka.jp\/wordpress\/?p=959"},"modified":"2023-07-18T06:00:00","modified_gmt":"2023-07-17T21:00:00","slug":"damo-polylm-13b%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=959","title":{"rendered":"DAMO PolyLM-13B\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd"},"content":{"rendered":"\n

Alibaba\u3001DAMO ACADEMY\u3088\u308aApache-2\u30e9\u30a4\u30bb\u30f3\u30b9\u3067\u30de\u30eb\u30c1\u30ea\u30f3\u30ac\u30eb\u306aLLM\u3001[2307.06018v1] PolyLM: An Open Source Polyglot Large Language Model (arxiv.org)<\/a>\u304c\u51fa\u3066\u304a\u308a\u3001\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u3092\u8a55\u4fa1\u3057\u3066\u307f\u305f\u3002<\/p>\n\n\n\n

HuggingFace\u304b\u3089\u5229\u7528\u53ef\u80fd\u3067DAMO-NLP-MT\/polylm-13b \u00b7 Hugging Face<\/a>\u306e\u4ed6\u3001DAMO-NLP-MT\/polylm-multialpaca-13b \u00b7 Hugging Face<\/a>\u3082\u63d0\u4f9b\u3055\u308c\u3066\u3044\u308b\u3002<\/p>\n\n\n\n

\u8ad6\u6587\u306b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3084\u69cb\u7bc9\u65b9\u6cd5\u306e\u8a73\u7d30\u306a\u60c5\u5831\u304c\u3042\u308b\u306e\u304c\u975e\u5e38\u306b\u3042\u308a\u304c\u305f\u3044\u300213B\u30e2\u30c7\u30eb\u69cb\u7bc9\u306b\u8981\u3057\u305f\u8a08\u7b97\u30ea\u30bd\u30fc\u30b9\u306f32 A100 GPU (8\u00d780G) servers\u306729 days\u3068\u306e\u3053\u3068\u3002\u4eca\u5f8c\u306e\u62e1\u5f35\u3068\u3057\u3066\u4e0b\u8a18\u304c\u4e88\u5b9a\u3055\u308c\u3066\u3044\u308b\u3068\u306e\u3053\u3068\u3067\u975e\u5e38\u306b\u697d\u3057\u307f\u3067\u3042\u308b\u3002<\/p>\n\n\n\n

\n

We are continuously enhancing the capabilities of PolyLM by focusing on the following aspects:<\/p>\n\n\n\n

    \n
  1. Replacement of absolute position embeddings with RoPE, as outlined in the research paper here<\/a>.<\/li>\n\n\n\n
  2. Expansion of window size to more than 10,000.<\/li>\n\n\n\n
  3. Verification of lightweight techniques to quickly enhance multilingual quality, especially for low-resource languages.<\/li>\n<\/ol>\nhttps:\/\/huggingface.co\/DAMO-NLP-MT\/polylm-13b<\/cite><\/blockquote>\n\n\n\n

    \u8ad6\u6587\u3067\u306fpolylm-multialpaca-13b<\/a>\u306e\u65b9\u304c\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u304c\u9ad8\u305d\u3046\u3067\u3042\u3063\u305f\u304c\u5236\u5fa1\u304c\u96e3\u3057\u304f[1]\u672c\u4ef6\u8a66\u884c\u306fpolylm-13b<\/a>\u3067\u884c\u3063\u3066\u3044\u308b\u3002\u74b0\u5883\u306fColab Pro+\u3092\u7528\u3044\u305f\u3002bfloat16\u3067\u8aad\u307f\u8fbc\u3080\u3053\u3068\u3068GPU\u5074\u3067\u8a08\u7b97[2]\u3055\u305b\u306a\u3044\u3068\u30a8\u30e9\u30fc\u306b\u306a\u308b\u70b9\u306b\u6ce8\u610f\u304c\u5fc5\u8981\u3002\u307e\u305f\u3001A100 VRAM 40GB\u3067\u306f\u30e1\u30e2\u30ea\u5236\u7d04\u304c\u53b3\u3057\u3044\u305f\u3081\u7279\u306b\u9577\u3044\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3067\u306f\u30a8\u30e9\u30fc\u304c\u767a\u751f\u3059\u308b\u4e8b\u304c\u3042\u308b[3]\u3002\u672c\u4ef6\u3067\u3082Out of memory\u767a\u751f\u6642\u306fnum_beams\u306a\u3069\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u8abf\u6574\u3057\u3066\u3044\u308b\u3002<\/p>\n\n\n\n

    \u6027\u80fd\u8a55\u4fa1\u306b\u4f7f\u7528\u3057\u305f\u30c7\u30fc\u30bf\u306f\u524d\u56de\uff08MPT-30B-Chat + In-Context Learning\u306e\u6027\u80fd | \u3077\u308b\u30fc\u3075\u304a\u3076\u3053\u3093\u305b\u3077\u3068 (staka.jp)<\/a>\uff09\u3068\u540c\u3058\u3067\u3042\u308a50\u4ef6\u3092\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3057\u3066\u4f7f\u3063\u3066\u3044\u308b\u3002MPT-30B\u540c\u69d8\u3001\u56de\u7b54\u90e8\u5206\u306e\u62bd\u51fa\u304c\u96e3\u3057\u3044\u3053\u3068\u304c\u3042\u308a\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

    \u7d50\u679c\u306f\u4e0b\u8868\u306e\u901a\u308a\u3002OSS\u306a\u30de\u30eb\u30c1\u30ea\u30f3\u30ac\u30eb\u30e2\u30c7\u30eb\u3068\u3057\u3066\u306f\u304b\u306a\u308a\u6027\u80fd\u304c\u9ad8\u3044\u3002<\/p>\n\n\n\n

    \u30e2\u30c7\u30eb\u30fb\u6761\u4ef6<\/strong><\/td>zero shot<\/strong><\/td>1 shot [4]<\/strong><\/td><\/tr>
    GPT-3.5 [5]<\/td>26.7<\/td>37.0<\/strong><\/td><\/tr>
    PolyLM-13B<\/td>10.6<\/td>21.5<\/td><\/tr>
    PolyLM-13B
    \u203b\u56de\u7b54\u90e8\u3092\u624b\u52d5\u62bd\u51fa<\/td>    		14.1<\/td>25.8<\/strong><\/td><\/tr><\/tbody><\/table>
    PolyLM-13B, GPT-3.5-TURBO-16K\u306eBLEU<\/figcaption><\/figure>\n\n\n\n

    BLEU\u7684\u306b\u306f1 shot\u3067\u56de\u7b54\u3092\u3046\u307e\u304f\u62bd\u51fa\u3067\u304d\u308c\u3070GPT-3.5\u306e\u30bc\u30ed\u30b7\u30e7\u30c3\u30c8\u3068\u540c\u30ec\u30d9\u30eb\u306e\u54c1\u8cea\u3068\u306a\u3063\u3066\u3044\u308b\u3002\u76ee\u691c\u3059\u308b\u3068\u91cd\u8981\u90e8\u5206\u3067\u306e\u8a33\u30df\u30b9\u304c\u591a\u304fGPT-3.5\u3068\u6bd4\u8f03\u3057\u3066\u3082BLEU\u4ee5\u4e0a\u306b\u5dee\u304c\u5927\u304d\u3044\u5370\u8c61\u3002<\/p>\n\n\n\n

    \u30d1\u30e9\u30e1\u30fc\u30bf\u657013B\u30681 GPU\u3067\u52d5\u4f5c\u53ef\u80fd\uff06\u30de\u30eb\u30c1\u30ea\u30f3\u30ac\u30eb\uff06\u65e5\u672c\u8a9e\u5bfe\u5fdc\u306e\u30e2\u30c7\u30eb\u304cApache-2\u3067\u5229\u7528\u53ef\u80fd\u3068\u3044\u3046\u306e\u306f\u3068\u3066\u3082\u3042\u308a\u304c\u305f\u3044\u3002LLaMA v2\u306e\u5642\u3082\u51fa\u3066\u3044\u3066\u30aa\u30fc\u30d7\u30f3\u306aLLM\u958b\u767a\u306f\u76db\u308a\u4e0a\u304c\u3063\u3066\u3044\u308b\u3002\u3053\u306e\u72b6\u6cc1\u304c\u7d9a\u304d\u826f\u3044\u3082\u306e\u304c\u51fa\u3066\u307b\u3057\u3044\u3068\u5f37\u304f\u601d\u3046\u3002<\/p>\n\n\n\n

    \u811a\u6ce8<\/h4>\n\n\n\n

    [1] \u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u3046\u307e\u304f\u8abf\u6574\u3067\u304d\u308c\u3070\u826f\u3044\u306e\u304b\u3082\u3060\u304c\u3001\u4e2d\u56fd\u8a9e\u306e\u51fa\u529b\u304c\u3055\u308c\u308b\u3001hallucination\u304c\u51c4\u3044\u306a\u3069\u6b63\u3057\u3044\u7ffb\u8a33\u3078\u306e\u8a98\u5c0e\u304c\u96e3\u3057\u3044\u3053\u3068\u304c\u591a\u304b\u3063\u305f\u3002\u3002
    [2] \u672c\u4ef6\u3067\u306fset_default_tensor_type\u3092\u4f7f\u7528\u3057\u305f\u304c\u3001to.(‘cuda’)\u307f\u305f\u304f\u660e\u793a\u7684\u306b\u98db\u3070\u3057\u3066\u3082\u3088\u3044\u306f\u305a\u3002
    [3] torch.cuda.empty_cache()\u3082\u547c\u3073\u307e\u304f\u3063\u3066\u3044\u308b\u3002
    [4] OpenICL\u3001TopkRetriever\u306b\u3088\u308a\u53d6\u5f97
    [5] gpt-3.5-turbo-16k-0613<\/p>\n","protected":false},"excerpt":{"rendered":"

    Alibaba\u3001DAMO ACADEMY\u3088\u308aApache-2\u30e9\u30a4\u30bb\u30f3\u30b9\u3067\u30de\u30eb\u30c1\u30ea\u30f3\u30ac\u30eb\u306aLLM\u3001[2307.06018v1] PolyLM: An Open Source Polyglot Large Language … “DAMO PolyLM-13B\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd” \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":[37,58,59],"class_list":["post-959","post","type-post","status-publish","format-standard","hentry","category-ai","category-13","tag-gpt-3-5-turbo","tag-polylm","tag-polylm-13b"],"_links":{"self":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/959","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=959"}],"version-history":[{"count":0,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/959\/revisions"}],"wp:attachment":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=959"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=959"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=959"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}