[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-target-distribution":3},{"tag":4,"articles":10},{"id":5,"name":6,"slug":7,"article_count":8,"description_zh":9,"description_en":9},"16fc1f9f-ac63-4884-a9dd-9ada03e64132","target distribution","target-distribution",0,null,[11],{"id":12,"slug":13,"title":14,"summary":15,"category":16,"image_url":17,"cover_image":17,"language":18,"created_at":19},"12ecefe1-00ea-4c54-8c7f-b71646f5dba3","unifying-sft-target-distribution-design-zh","SFT 不只看 loss，先設計目標分布","這篇論文把 SFT 重新定義成目標分布設計，主張先決定該信任多少標註，再分配剩餘機率，並在十組推理資料與模型設定中一致優於基線。","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781072297687-gtyc.png","zh","2026-06-10T06:17:32.859647+00:00"]