{"id":5011,"date":"2024-06-21T02:55:00","date_gmt":"2024-06-20T17:55:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=5011"},"modified":"2024-06-21T02:55:00","modified_gmt":"2024-06-20T17:55:00","slug":"an-empirical-study-of-mamba-based-language-models","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=5011","title":{"rendered":"An Empirical Study of Mamba-based Language Models"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>An Empirical Study of Mamba-based Language Models\u00a0<\/strong>[69.7]<br>Mamba\u306e\u3088\u3046\u306a\u9078\u629e\u7684\u306a\u72b6\u614b\u7a7a\u9593\u30e2\u30c7\u30eb(SSM)\u306f\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\u306e\u6b20\u70b9\u3092\u514b\u670d\u3059\u308b\u3002 \u540c\u3058\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u4e0a\u3067\u8a13\u7df4\u3055\u308c\u305f8B-context Mamba, Mamba-2, Transformer \u30e2\u30c7\u30eb\u3092\u76f4\u63a5\u6bd4\u8f03\u3059\u308b\u3002 8B\u306eMamba-2-Hybrid\u306f\u300112\u306e\u6a19\u6e96\u30bf\u30b9\u30af\u30678B\u306eTransformer\u3092\u4e0a\u56de\u3063\u3066\u3044\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2406.07887v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2406.07887v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 12 Jun 2024 05:25:15 GMT)<\/li>\n\n\n\n<li>Mamba\u306e\u5b9f\u9a13\u7684\u691c\u8a3c\u30028B\u30013.5T tokens\u3067mamba\u3001mamba2\u3001transformer\u3092\u6bd4\u8f03\u3002\u300cOur results show that while pure SSM-based models match or exceed Transformers on many tasks, both Mamba and Mamba-2 models lag behind Transformer models on tasks which require strong copying or in-context learning abilities (e g , five-shot MMLU, Phonebook Lookup) or long-context reasoning.\u300d\u3001\u300cwe find that the 8B-parameter Mamba2-Hybrid exceeds the 8B-parameter Transformer on all 12 standard tasks we evaluated (+2.65 points on average) and is predicted to be up to 8\u00d7 faster when generating tokens at inference time.\u300d\u3068\u306e\u7d50\u679c\u3002\u4eca\u307e\u3067\u306e\u8ad6\u6587\u3068\u6bd4\u3079\u3066\u610f\u5916\u6027\u306f\u306a\u3044\u304c\u3001\u5305\u62ec\u7684\u306a\u691c\u8a3c\u306f\u3068\u3066\u3082\u53c2\u8003\u306b\u306a\u308b\u3002\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u69cb\u6210\u306f\u3068\u3066\u3082\u6709\u52b9\u306a\u9078\u629e\u80a2\u306b\u898b\u3048\u305f\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/NVIDIA\/Megatron-LM\/tree\/ssm\/examples\/mamba\">Megatron-LM\/examples\/mamba at ssm \u00b7 NVIDIA\/Megatron-LM \u00b7 GitHub<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[235],"class_list":["post-5011","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-mamba"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/5011","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5011"}],"version-history":[{"count":0,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/5011\/revisions"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5011"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5011"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5011"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}