[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-5-steps-fine-tune-local-llm-zh":3,"article-related-5-steps-fine-tune-local-llm-zh":33,"series-industry-546dc8aa-3e25-49c5-aa9c-f06841c6827f":85},{"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":25,"views":29,"created_at":30,"published_at":31,"topic_cluster_id":32},"546dc8aa-3e25-49c5-aa9c-f06841c6827f","5-steps-fine-tune-local-llm-zh","5 個本地 LLM 微調步驟","\u003Cp data-speakable=\"summary\">這篇整理 5 個步驟，讓你在週末完成本地 \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> 微調，從環境、資料到訓練、評估與匯出一次走完。\u003C\u002Fp>\u003Cp>如果你想在 2 天內做出可本機部署的微調模型，這份清單會把流程拆成 5 步，並給你一個具體參考：27B 的 \u003Ca href=\"\u002Ftag\u002Fqwen\">Qwen\u003C\u002Fa> 3.5 最後可壓到約 18 GB 的 GGUF 檔。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>步驟\u003C\u002Fth>\u003Cth>時間窗口\u003C\u002Fth>\u003Cth>主要產出\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>1. 週五環境\u003C\u002Ftd>\u003Ctd>2-3 小時\u003C\u002Ftd>\u003Ctd>可用 GPU、驅動、基座模型\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>2. 週六資料\u003C\u002Ftd>\u003Ctd>4 小時\u003C\u002Ftd>\u003Ctd>Prompt-response 資料集\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>3. 週六訓練\u003C\u002Ftd>\u003Ctd>3-4 小時\u003C\u002Ftd>\u003Ctd>LoRA adapter\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>4. 週日評估\u003C\u002Ftd>\u003Ctd>2 小時\u003C\u002Ftd>\u003Ctd>對照測試與品質檢查\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>5. 週日匯出\u003C\u002Ftd>\u003Ctd>2 小時\u003C\u002Ftd>\u003Ctd>可本機使用的 GGUF 模型\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. 週五先把環境架好\u003C\u002Fh2>\u003Cp>先處理機器與訓練堆疊，因為環境出問題，整個週末都會被拖垮。最實際的做法是先確認 GPU、驅動、Python 環境與基座模型都能正常運作，再開始碰資料。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779988670307-nfq9.png\" alt=\"5 個本地 LLM 微調步驟\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>單卡週末流程裡，\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Funslothai\u002Funsloth\">Unsloth\u003C\u002Fa> 很適合追求快速 LoRA 訓練，\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Faxolotl-ai-cloud\u002Faxolotl\">Axolotl\u003C\u002Fa> 則適合想保留更多控制權的人。硬體上，\u003Ca href=\"\u002Ftag\u002Fnvidia\">NVIDIA\u003C\u002Fa> \u003Ca href=\"\u002Ftag\u002Fcuda\">CUDA\u003C\u002Fa> 最穩，AMD ROCm 可當備案，\u003Ca href=\"\u002Ftag\u002Fapple\">Apple\u003C\u002Fa> silicon \u003Ca href=\"\u002Fnews\u002Fbest-prompt-engineering-tools-2026-zh\">比較\u003C\u002Fa>適合推理，不適合拿來做微調。\u003C\u002Fp>\u003Cul>\u003Cli>預留 2 到 3 小時做環境檢查。\u003C\u002Fli>\u003Cli>安裝 CUDA 驅動與乾淨的 Python 環境。\u003C\u002Fli>\u003Cli>先載入基座模型，確認能正常推理。\u003C\u002Fli>\u003Cli>模型回應正確前，不要開始訓練。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. 週六把資料整理成可訓練格式\u003C\u002Fh2>\u003Cp>微調不會從零散筆記或原始對話中學到好結果，它需要格式固定的 prompt-response 配對。這一步的重點不是資料多，而是資料乾淨、風格一致、和你之後要使用的聊天格式相符。\u003C\u002Fp>\u003Cp>做法通常是先清理來源文字，再切段、改寫，最後整理成訓練樣本。你可以用小模型輔助產生問題，但最後仍要人工確認答案是否符合你的語氣、領域與輸出規則。\u003C\u002Fp>\u003Cul>\u003Cli>8B 模型可先抓 1 到 2 百萬原始 tokens 當起點。\u003C\u002Fli>\u003Cli>先修正拼字與格式錯誤，再進訓練。\u003C\u002Fli>\u003Cli>把長篇內容轉成 prompt-response 配對。\u003C\u002Fli>\u003Cli>訓練格式要和推理格式保持一致。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>3. 週六用 LoRA 開始訓練\u003C\u002Fh2>\u003Cp>LoRA 讓週末微調變得可行。它不是更新整個模型，而是只訓練少量 adapter，通常只佔總參數的 0.5% 到 1.5%，所以消費級 GPU 也有機會跑完。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779988672442-wnav.png\" alt=\"5 個本地 LLM 微調步驟\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這一步最常見的失誤，是學習率設錯、rank 不合適，或把推理型模型放在不對的模式下。實務上，27B 模型至少需要 14 GB VRAM，留更多餘量會更\u003Ca href=\"\u002Fnews\u002Fhow-to-secure-ai-assistants-end-to-end-zh\">安全\u003C\u002Fa>。\u003C\u002Fp>\u003Ccode>週末訓練檢查表：\n- 1 張 GPU\n- 1 套框架\n- 1 份資料集\n- 1 次失敗嘗試\n- 1 次修正後重跑\u003C\u002Fcode>\u003Ch2>4. 週日用固定題目做評估\u003C\u002Fh2>\u003Cp>評估是確認模型真的學到東西的關鍵。做一組你已經知道標準答案的測試題，然後把基座模型與微調後模型並排比較，這樣最容易看出差異。\u003C\u002Fp>\u003Cp>判斷重點不只在對錯，還包括語氣、長度與格式。如果基座模型回答得很泛、很長，而微調後模型更像你的風格，回覆更直接、結構更穩定，這次訓練就算有價值。\u003C\u002Fp>\u003Cul>\u003Cli>使用固定的測試題組。\u003C\u002Fli>\u003Cli>把原模型與微調模型並排看。\u003C\u002Fli>\u003Cli>同時檢查正確性、語氣與格式。\u003C\u002Fli>\u003Cli>如果結果怪，先回頭看資料而不是先怪參數。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. 週日把模型匯出成 GGUF\u003C\u002Fh2>\u003Cp>確認 LoRA adapter 表現穩定後，就把它合併回基座模型，再匯出成 GGUF。這是 \u003Ca href=\"https:\u002F\u002Follama.com\">Ollama\u003C\u002Fa> 和 \u003Ca href=\"https:\u002F\u002Flmstudio.ai\">LM Studio\u003C\u002Fa> 等本機\u003Ca href=\"\u002Fnews\u002F7-ai-code-review-tools-zh\">工具\u003C\u002Fa>常用的格式，最後的量化通常也在這一步完成。\u003C\u002Fp>\u003Cp>實際好處是部署簡單。原始素材提到，一個微調後的 27B Qwen 3.5 模型，最後大約可落在 18 GB 的 GGUF 檔，並且能在同一台機器上順利跑起來。\u003C\u002Fp>\u003Cul>\u003Cli>先把 adapter 合併進 base weights。\u003C\u002Fli>\u003Cli>輸出成 GGUF。\u003C\u002Fli>\u003Cli>需要時再做 4-bit 或 5-bit 量化。\u003C\u002Fli>\u003Cli>用 Modelfile 或本機 runner 註冊模型。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>怎麼挑\u003C\u002Fh2>\u003Cp>如果你是第一次做本地微調，先把時間放在環境與資料品質上，這兩步最決定成敗。若你的目標是讓模型說話像你本人，就把更多精力放在 prompt-response 格式與評估，而不是一直調參。\u003C\u002Fp>\u003Cp>如果你要的是可長期維護的本機 AI，微調適合穩定語氣與固定知識，RAG 則更適合會變動的資訊，例如新聞、政策或產品資料。兩者搭配，通常比單靠微調更實用。\u003C\u002Fp>","5 個步驟帶你在週末完成本地 LLM 微調，從環境、資料到訓練、評估與 GGUF 匯出。","zenvanriel.com","https:\u002F\u002Fzenvanriel.com\u002Fai-engineer-blog\u002Ffine-tune-local-llm-single-weekend-home-lab\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779988670307-nfq9.png","industry","zh","6bcb38e2-63e6-4de1-898f-92976aaf003f",[17,18,19,20,21,22,23,24],"本地 LLM","微調","LoRA","GGUF","Unsloth","Axolotl","Ollama","LM Studio",[26,27,28],"先確認環境與基座模型可用，再開始訓練。","資料要整理成一致的 prompt-response 格式。","LoRA 與 GGUF 讓週末完成訓練與本機部署成為可能。",5,"2026-05-28T17:17:22.855882+00:00","2026-05-28T17:17:22.835+00:00","da242733-a19a-4cb7-b706-05f8699aa19e",{"tags":34,"relatedLang":44,"relatedPosts":48},[35,37,39,41,43],{"name":21,"slug":36},"unsloth",{"name":19,"slug":38},"lora",{"name":17,"slug":40},"本地-llm",{"name":20,"slug":42},"gguf",{"name":18,"slug":18},{"id":15,"slug":45,"title":46,"language":47},"5-steps-fine-tune-local-llm-en","5 steps to fine tune a local LLM","en",[49,55,61,67,73,79],{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"0d604500-3a70-40ec-a70e-370f972a66ab","korea-nvidia-talks-ai-factory-push-zh","韓國與 Nvidia 對話，重點是 AI 工廠","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781057871797-7uxx.png","2026-06-10T02:17:21.099824+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"173b8876-1867-4e0b-948f-27891d6b6364","openai-should-not-rush-its-ipo-just-to-win-the-ai-race-zh","OpenAI 不該為了搶 AI 賽道而急著 IPO","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781053365610-1hko.png","2026-06-10T01:02:19.886627+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"3d7ff80a-4045-4b66-9e21-b6a8eb3b6f6d","openai-europe-privacy-policy-zh","OpenAI 歐洲隱私政策更新重點","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781052479369-yomr.png","2026-06-10T00:47:31.176745+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"69002c63-177a-4723-9e63-d28506f08edd","openai-ads-sensitive-chats-policy-zh","OpenAI把廣告擋在敏感對話外是對的","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781051578409-en02.png","2026-06-10T00:32:23.404084+00:00",{"id":74,"slug":75,"title":76,"cover_image":77,"image_url":77,"created_at":78,"category":13},"ea98a8c9-ebe1-4258-8a2b-b0d82b25deed","ai-bootlegs-streaming-royalties-stick-figure-zh","AI bootlegs 正在抽走串流版稅","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781050681742-3rdh.png","2026-06-10T00:17:31.017287+00:00",{"id":80,"slug":81,"title":82,"cover_image":83,"image_url":83,"created_at":84,"category":13},"20d0b5fc-a363-481d-86b2-e30276a49e92","amd-microsoft-windows-ml-acceleration-zh","AMD 與 Microsoft 把 Windows ML 推進 GPU 與 N…","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781047980407-vd5p.png","2026-06-09T23:32:31.304436+00:00",[86,91,96,101,106,111,116,121,126,131],{"id":87,"slug":88,"title":89,"created_at":90},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":92,"slug":93,"title":94,"created_at":95},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":97,"slug":98,"title":99,"created_at":100},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":102,"slug":103,"title":104,"created_at":105},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":107,"slug":108,"title":109,"created_at":110},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":112,"slug":113,"title":114,"created_at":115},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":117,"slug":118,"title":119,"created_at":120},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":122,"slug":123,"title":124,"created_at":125},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":127,"slug":128,"title":129,"created_at":130},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":132,"slug":133,"title":134,"created_at":135},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]