[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-fine-tuning-beats-rag-style-not-facts-zh":3,"article-related-fine-tuning-beats-rag-style-not-facts-zh":30,"series-ai-agent-7d860405-aca6-486b-8de0-1c5193a3b06d":81},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"7d860405-aca6-486b-8de0-1c5193a3b06d","fine-tuning-beats-rag-style-not-facts-zh","當目標是文風不是事實時，微調比 RAG 更有效","\u003Cp data-speakable=\"summary\">如果你要訓練 \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> 學會特定文風，微調才是正解；RAG 比較適合補事實，不適合改寫作風格。\u003C\u002Fp>\u003Cp>如果你要 LLM 寫出 1990 年代技術文件的味道，答案不是檢索，而是微調。Fabrizio Ferri Benedetti 的實驗把這件事講得很清楚：他用數千萬字的舊版 \u003Ca href=\"\u002Ftag\u002Fmicrosoft\">Microsoft\u003C\u002Fa> 手冊訓練本地 adapter，最後得到的輸出不只會引用老文件，而是連章節節奏、詞彙選擇、段落開頭都像那個年代的文風。\u003C\u002Fp>\u003Ch2>第一個論點\u003C\u002Fh2>\u003Cp>RAG 的設計目標是「回答問題時抓取相關資料」，所以它擅長的是事實正確、來源可追溯、內容可更新。這很適合產品手冊、法規問答、內部知識庫，\u003Ca href=\"\u002Fnews\u002Fbitcoin-defi-will-grow-but-not-by-copying-ethereum-zh\">但不\u003C\u002Fa>適合文風轉換。Benedetti 的測試提示刻意偏向風格任務，例如要模型用 1990 年代 Microsoft 的口吻解釋 \u003Ccode>malloc()\u003C\u002Fcode>，或替虛構的 \u003Ccode>ConnectWifi()\u003C\u002Fcode> \u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa> 寫文件；微調後的模型真的寫出 man page 式標頭、synopsis \u003Ca href=\"\u002Fnews\u002Fai-blockchain-projects-need-real-utility-not-token-theater-zh\">區塊\u003C\u002Fa>與年代感措辭。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780924689232-5elu.png\" alt=\"當目標是文風不是事實時，微調比 RAG 更有效\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這裡的差別不是「有沒有看到範例」，而是「有沒有改變行為」。基準模型多半只會輸出現代 Markdown、親切解說，甚至直接跑題；微調後的 \u003Ca href=\"\u002Ftag\u002Fqwen\">Qwen\u003C\u002Fa> 變體卻能在不同提示下維持相同的註解格式與敘述節奏。文風不是一段可被撈出的文字，而是一組輸出分佈。你要的是分佈偏移，微調才做得到。\u003C\u002Fp>\u003Ch2>第二個論點\u003C\u002Fh2>\u003Cp>小型 adapter 已經足夠捕捉強烈風格，而且成本低到個人也能做。Benedetti 用的是 QLoRA，凍結底層權重，只在上面加輕量 adapter，整個實驗大約花了 50 美元、一天左右就完成。這代表不是只有大公司能做風格模型，獨立開發者也能在租來的 GPU 上完成一次可用的文風訓練。\u003C\u002Fp>\u003Cp>資料品質比花俏提示更重要，這在他的結果裡很明顯。作者從 Bitsavers 蒐集、清理 OCR 噪音、過濾可讀段落，最後做出超過 192,000 筆 instruction 範例。這種語料反覆灌輸同一套文件慣例，像是標題、回傳值、程式碼區塊、交叉引用與範例。單靠 \u003Ca href=\"\u002Ftag\u002Fprompt-engineering\">prompt engineering\u003C\u002Fa> 不會長出這種深度模式；只有把這些模式寫進模型，才會真的穩定重現。\u003C\u002Fp>\u003Ch2>反方可能怎麼說\u003C\u002Fh2>\u003Cp>RAG 支持者的論點其實很強：如果你在乎的是正確性、來源、以及最新資訊，檢索比把知識烘進權重更安全。文件庫比重新訓練更容易更新，答案也更容易附上引用、稽核與修正。對企業系統來說，RAG 常常是更穩的預設方案。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780924676978-bl0p.png\" alt=\"當目標是文風不是事實時，微調比 RAG 更有效\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>還有成本問題。微調需要整理語料、跑訓練、做評估，還要維持部署流程；RAG 常常能直接接到既有架構上，前期阻力更小。對那些這季就要交付產品的人來說，檢索看起來確實比較務實。\u003C\u002Fp>\u003Cp>但這個論點\u003Ca href=\"\u002Fnews\u002Frust-worth-the-hype-2026-right-jobs-zh\">只適\u003C\u002Fa>用於事實任務，不適用於文風任務。當你的目標是讓模型寫出特定年代、特定團隊或特定媒體的語氣，RAG 只是把更多文字放進上下文，並沒有改變模型的預設輸出習慣。Benedetti 的結果證明，模型必須內化模式，而不是旁邊放一堆範例。要風格就用微調，要事實就用 RAG，兩者不是替代品。\u003C\u002Fp>\u003Ch2>你能做什麼\u003C\u002Fh2>\u003Cp>如果你是工程師或 PM，先把問題拆成兩層：用 RAG 管來源真實性，用微調管語氣、結構與格式；如果你是創辦人，不要賣「RAG 解決一切」，而是賣一個知道何時該檢索、何時該模仿的系統。這才是文件助手、品牌寫作工具與內部知識產品真正的分工。風格靠訓練，事實靠檢索，懂這條線的人，會做出更好的產品。\u003C\u002Fp>","如果你要訓練 LLM 學會特定文風，微調才是正解；RAG 比較適合補事實，不適合改寫作風格。","passo.uno","https:\u002F\u002Fpasso.uno\u002Ffine-tuning-docs-llm\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780924689232-5elu.png","ai-agent","zh","b413d484-6786-4c32-abdc-77f010ac7eba",[17,18,19,20,21],"LLM 微調","RAG","文風轉移","QLoRA","技術文件",[23,24,25],"微調會改變模型的輸出習慣，RAG 主要只是補上下文。","要學特定文風，資料品質與一致性比提示詞更重要。","實務上應分工：RAG 管事實，微調管語氣與格式。",2,"2026-06-08T13:17:25.235242+00:00","2026-06-08T13:17:25.225+00:00","e3b68196-9e64-4c18-a3b6-a73e73bfb367",{"tags":31,"relatedLang":40,"relatedPosts":44},[32,34,36,37,38],{"name":20,"slug":33},"qlora",{"name":18,"slug":35},"rag",{"name":21,"slug":21},{"name":19,"slug":19},{"name":17,"slug":39},"llm-微調",{"id":15,"slug":41,"title":42,"language":43},"fine-tuning-beats-rag-style-not-facts-en","Fine-tuning beats RAG when the goal is style, not facts","en",[45,51,57,63,69,75],{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"3d1e5ef7-8f31-4e57-b286-306825d7f38e","openclaw-small-business-ai-staff-zh","OpenClaw把AI變成夜班員工","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780904888882-6w0v.png","2026-06-08T07:47:27.229503+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"0cd44c8d-6ba8-4e6c-851b-d040a5c1a9bd","litellm-rust-minimal-ai-gateway-zh","LiteLLM 推出 Rust 版輕量網關","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780899485895-xavr.png","2026-06-08T06:17:32.954118+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"e5195a2a-110f-450d-97f7-298edd173747","claurst-terminal-coding-agents-open-local-zh","Claurst 證明終端編碼代理應該開源且本地化","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780888681781-75z2.png","2026-06-08T03:17:22.236957+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"823f413f-0504-425c-a8de-956a60754b9d","how-to-set-up-agentscope-java-harness-zh","怎麼設定 AgentScope Java Harness","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780877895461-376p.png","2026-06-08T00:17:46.417304+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"066e5903-7569-402f-a397-f01468ffecd4","reid-hoffman-leaves-microsoft-board-manus-ai-zh","霍夫曼離開微軟董事會，轉向 Manus","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780827469679-fpyr.png","2026-06-07T10:17:21.250923+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":13},"fb5eb422-6013-43fe-bdc4-26c57eee0d9e","how-to-understand-codex-chatgpt-merge-zh","怎麼理解 Codex 與 ChatGPT 合併","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780704172974-ypaa.png","2026-06-06T00:02:25.891496+00:00",[82,87,92,97,102,107,112,117,122,127],{"id":83,"slug":84,"title":85,"created_at":86},"4ae1e197-1d3d-4233-8733-eafe9cb6438b","claude-now-uses-your-pc-to-finish-tasks-zh","Claude 開始幫你操作電腦","2026-03-26T07:20:48.457387+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"5bede67f-e21c-413d-9ab8-54a3c3d26227","googles-2026-ai-agent-report-decoded-zh","Google 2026 AI Agent 報告解讀","2026-03-26T11:15:22.651956+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"2987d097-563f-46c7-b76f-b558d8ef7c2b","kimi-k25-review-stronger-still-not-legend-zh","Kimi K2.5 評測：更強，但還不是神作","2026-03-27T07:15:55.277513+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"95c9053b-e3f4-4cb5-aace-5c54f4c9e044","claude-code-controls-mac-desktop-zh","Claude Code 也能操控 Mac 了","2026-03-28T03:01:58.58121+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"dc58e153-e3a8-4c06-9b96-1aa64eabbf5f","cloudflare-100x-faster-ai-agent-sandbox-zh","Cloudflare 的 AI 沙箱跑超快","2026-03-28T03:09:44.142236+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"1c8afc56-253f-47a2-979f-1065ff072f2a","openai-backs-isara-agent-swarm-bet-zh","OpenAI 挺 Isara 的 agent swarm …","2026-03-28T03:15:27.513155+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"7379b422-576e-45df-ad5a-d57a0d9dd467","openai-plan-automated-ai-researcher-zh","OpenAI 想做自動化 AI 研究員","2026-03-28T03:17:42.090548+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"48c9889e-86df-450b-a356-e4a4b7c83c5b","harness-engineering-ai-agent-reliability-2026-zh","駕馭工程：從「馬具」到「作業系統」，AI Agent 可靠性的終極密碼","2026-03-31T06:42:53.556721+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"96d8e8c8-1edd-475d-9145-b1e7a1b02b65","mcp-explained-from-prompts-to-production-zh","MCP 怎麼把提示詞變工作流","2026-04-01T09:24:39.321274+00:00",{"id":128,"slug":129,"title":130,"created_at":131},"f2ca7720-b471-4ce5-9336-2a9ac2a876fd","amazon-bedrock-agents-multi-agent-workflows-zh","Amazon Bedrock Agents 進入多代理工作流","2026-04-01T09:30:29.945429+00:00"]