https://docs.keinmal.top/notes/llm-lab/ Recent content in LLM Lab on Docs Hugo --0.165.0 zh-cnwyz2016zxc@outlook.com(Hiraeth)wyz2016zxc@outlook.com(Hiraeth)Thu, 08 Oct 2026 19:25:46 +0800 https://docs.keinmal.top/notes/llm-lab/07-minimind-train/ Thu, 08 Oct 2026 00:00:00 +0000wyz2016zxc@outlook.com(Hiraeth)https://docs.keinmal.top/notes/llm-lab/07-minimind-train/ 训练顺序：每步做什么、要什么、产出什么、多久 https://docs.keinmal.top/notes/llm-lab/06-minimind/ Thu, 08 Oct 2026 00:00:00 +0000wyz2016zxc@outlook.com(Hiraeth)https://docs.keinmal.top/notes/llm-lab/06-minimind/ 上游项目的结构：trainer 阶梯、model 与 dataset 的取舍 https://docs.keinmal.top/notes/llm-lab/05-measurement/ Thu, 08 Oct 2026 00:00:00 +0000wyz2016zxc@outlook.com(Hiraeth)https://docs.keinmal.top/notes/llm-lab/05-measurement/ 四个测量 bug、累计 tok/s 的陷阱与有效算力 https://docs.keinmal.top/notes/llm-lab/04-data/ Thu, 08 Oct 2026 00:00:00 +0000wyz2016zxc@outlook.com(Hiraeth)https://docs.keinmal.top/notes/llm-lab/04-data/ 从 JSONL 语料到 (input_ids, labels) 张量，含实测数字 https://docs.keinmal.top/notes/llm-lab/03-model/ Thu, 08 Oct 2026 00:00:00 +0000wyz2016zxc@outlook.com(Hiraeth)https://docs.keinmal.top/notes/llm-lab/03-model/ MiniMind-3 的规格、一次前向的顺序与五个静默出错的坑 https://docs.keinmal.top/notes/llm-lab/02-tokenizer/ Thu, 08 Oct 2026 00:00:00 +0000wyz2016zxc@outlook.com(Hiraeth)https://docs.keinmal.top/notes/llm-lab/02-tokenizer/ 词表 6400 的来源、两套 API 的分工与实测更正 https://docs.keinmal.top/notes/llm-lab/01-plan/ Thu, 08 Oct 2026 00:00:00 +0000wyz2016zxc@outlook.com(Hiraeth)https://docs.keinmal.top/notes/llm-lab/01-plan/ 分阶段计划：环境、预训练、后训练与 RL