Sebastian Raschka releases an 11-hour video course on building a reasoning model, using…
English text is being prepared.
Sebastian Raschka's accompanying video course for "Build a Reasoning Model (From Scratch)" is now available: 11 hours across eight lectures, starting from the pretrained small Qwen3 0.6B and using Python/PyTorch to hand-code a verifier, inference-time scaling (sampling, self-consistency, scoring, self-refinement), a simplified GRPO version of RLVR, and distillation toward a stronger teacher. The course skips training the base model from scratch; the early generation and evaluation work can be done on a CPU, while the later training stages are recommended to use a GPU, with results evaluated on MATH-500. Shao Meng relayed the course design in a tweet and used an engine analogy to distinguish regular LLMs, reasoning models, and the Agent harness.
Why read this
AI score 67
The course starts from the pretrained Qwen3 0.6B and strings together a verifier, inference-time scaling, a GRPO version of RLVR, and distillation into a hands-on, reproducible path.
Topics
View original postThe complete reasoning-from-scratch video series is now up (and linked) here: https://t.co/5IOHLTCIGt (Taking a short break but will be back with LLM model and research updates in November!)
View original
The complete reasoning-from-scratch video series is now up (and linked) here: https://t.co/5IOHLTCIGt (Taking a short break but will be back with LLM model and research updates in November!)
Body
从零构建推理模型：11 小时视频课，把评估、推理时扩展、RLVR 与蒸馏写进 PyTorch
@rasbt 「Build a Reasoning Model (From Scratch)」系列绝对是系统学习模型构建原理和实践最佳教程，可能没有之一，这次开放的是与新书配套的完整编程课：从已预训练的小 Qwen3 起步，用 Python/PyTorch 手写评估 verifier、推理时扩展（采样 / 自洽 / 评分 / 自我精炼）、GRPO 版 RLVR，以及向更强教师蒸馏；主线是理解「推理模型」本身。 https://t.co/9HfdL9ePwo
推理模型、普通 LLM、Agent 怎么放 文中用发动机类比：常规 LLM 是引擎；推理模型是在其上加更多可验证奖励的强化学习后训练（可选再加推理时扩展），更擅长数学与代码等难题；Agent harness 则是外层工具与上下文管理，让引擎去做更大任务。Agent 可用普通 LLM 或推理模型，但当代主流多是带可调推理力度的推理模型。系列资源对应三本/三块：LLM from Scratch → Reasoning Model from Scratch → Coding Agent 组件文。
课程怎么设计 · 起点：小预训练 Qwen3 0.6B，跳过「从零训底座」，专攻评估、推理时扩展、RL、蒸馏 · 前置：熟悉 Python 与 PyTorch；上一门 LLM from Scratch 推荐但不强制 · 算力：前期生成与评估可 CPU；后期训练强烈建议 GPU；视频 1 与仓库有环境建议 · 学习路径建议：先读对应章节 → 跟视频写代码 → 做练习；另有约 90 分钟 LinkedIn Learning 概念课可当预习
八讲内容（按工作流） 1. 动机与环境：LLM / 推理 / Agent 关系；uv 装依赖；CUDA / Apple Silicon / 远程 GPU 2. 加载底座与生成：载入 Qwen3 权重；流式生成；KV cache + torch.compile 3. Verifier：抽答案 → 规范化 → 数学等价；在 MATH-500 上评估；并为后续 RLVR 奖励 复用 4. 推理时扩展 I：temperature / multinomial / top-p；self-consistency（多数票）；精度 vs 额外算力 5. 推理时扩展 II：token/序列 logprob 打分破平；self-refinement（批评→改写）并评估 6. RL：简化 GRPO：RLHF vs RLVR；组内相对优势；简化 GRPO loss；训练循环与 checkpoint 7. RL：稳定与分析：entropy / advantage、clip、KL、format reward；防 reward hacking；再评 MATH-500 8. 蒸馏：soft/hard、on/off-policy；用 DeepSeek R1 等教师数据；监督训小模型并对比