{"id":7132,"date":"2025-07-22T05:54:00","date_gmt":"2025-07-21T20:54:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7132"},"modified":"2025-07-20T15:00:58","modified_gmt":"2025-07-20T06:00:58","slug":"language-models-improve-when-pretraining-data-matches-target-tasks","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7132","title":{"rendered":"Language Models Improve When Pretraining Data Matches Target Tasks\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Language Models Improve When Pretraining Data Matches Target Tasks\u00a0<\/strong>[8.9]<br>BETR\u306f\u3001\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306e\u4f8b\u3068\u985e\u4f3c\u6027\u306b\u57fa\u3065\u3044\u3066\u3001\u4e8b\u524d\u5b66\u7fd2\u3057\u305f\u6587\u66f8\u3092\u9078\u629e\u3059\u308b\u65b9\u6cd5\u3067\u3042\u308b\u3002 \u30c7\u30fc\u30bf\u9078\u629e\u306e\u65b9\u6cd5\u306f10^19\u304b\u308910^22FLOP\u306b\u307e\u305f\u304c\u308b500\u4ee5\u4e0a\u306e\u30e2\u30c7\u30eb\u3092\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3057\u3001\u305d\u308c\u3089\u3092\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u6cd5\u5247\u306b\u9069\u5408\u3055\u305b\u308b\u3053\u3068\u3067\u6bd4\u8f03\u3059\u308b\u3002 BETR\u306fDCLM-Baseline\u4e0a\u30672.1\u500d\u306e\u8a08\u7b97\u4e57\u7b97\u3092\u5b9f\u73fe\u3057,\u5168\u30b9\u30b1\u30fc\u30eb\u306710\u30bf\u30b9\u30af\u4e2d9\u30bf\u30b9\u30af\u306e\u6027\u80fd\u5411\u4e0a\u3092\u5b9f\u73fe\u3057\u3066\u3044\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.12466v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.12466v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 16 Jul 2025 17:59:45 GMT)<\/li>\n\n\n\n<li>\u300cWe tested whether language models improve when pretraining data matches target tasks. This hypothesis seems almost self-evident: training on relevant data should naturally improve relevant capabilities.\u300d\u306f\u3067\u3059\u3088\u306d\u30fc\u3068\u3057\u3066\u3001\u300cAlthough explicit targeting might seem at odds with pretraining\u2019s traditional emphasis on generality, our scaling analysis offers a reconciling insight: as compute increases, optimal filtering becomes predictably less strict. Smaller models perform best when trained on narrowly filtered datasets, while larger models benefit from more diverse data.\u300d\u307e\u3067\u5206\u6790\u3059\u308b\u3068\u8208\u5473\u6df1\u3044\u3002<\/li>\n\n\n\n<li>\u8ad6\u6587\u306b\u3082\u66f8\u304b\u308c\u3066\u3044\u305f\u304c\u3001\u591a\u8a00\u8a9e\u3067\u3069\u3046\u306a\u308b\u304b\u306f\u3068\u3066\u3082\u8208\u5473\u304c\u3042\u308b\u3002<\/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":[349,541],"class_list":["post-7132","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-scaling-law","tag-541"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7132","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=7132"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7132\/revisions"}],"predecessor-version":[{"id":7133,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7132\/revisions\/7133"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7132"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7132"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7132"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}