Learn MiniMind: A 24-Lesson Course for Training a Small LLM from Scratch

What Learn MiniMind Is and Who It Is For

MiniMind is a separate open-source project that trains a 64-million-parameter GPT-style model for roughly three yuan in about two hours. Learn MiniMind is a companion tutorial repository, not the MiniMind code itself. It exists to explain how MiniMind works, from the first principles of what a large language model is through pretraining, supervised fine-tuning, LoRA, DPO, and deployment, and to prepare the reader to discuss the project in a job interview.

The README describes three intended audiences. Zero-base learners with no Python or deep learning background can follow the course with the help of illustrated manga explanations. Developers with Python knowledge but no ML experience can take the seven-day systematic path. People with existing deep learning background who need to prepare for interviews quickly can follow the three-day accelerated track.

The README notes that each lesson contains runnable PyTorch experiment code. The primary course language is Chinese; all lesson files in the docs/ directory and the interview question files are written in Chinese.

Five Phases: From First Principles to Offer Letter

The course is organized into five phases covering 24 lessons. The README shows the full progression in a single diagram:

你现在在这里
    ↓
Phase 1                    Phase 2                    Phase 3                    Phase 4                 Phase 5
零基础入门                 模型核心组件               训练全流程                 高级特性 & 面试          求职冲刺
============              ==============             ============              ================        ============

L01 什么是LLM         →   L05 Tokenizer分词器    →   L11 数据处理流水线    →   L17 DPO偏好优化     →  L23 简历撰写指南
 |                         |                          |                         |                       |
L02 Transformer全景    →   L06 词嵌入Embedding    →   L12 预训练Pretra

Phase 1 (Lessons 1 to 4) covers what a large language model is, the Transformer architecture, PyTorch basics, and a walkthrough of the MiniMind project structure.

Phase 2 (Lessons 5 to 10) goes through each component of the Transformer: Tokenizer, word embeddings, RMSNorm normalization, RoPE positional encoding, self-attention with GQA, and the FFN with SwiGLU. Each lesson maps to the corresponding MiniMind source code and includes a Doraemon-style manga illustration.

Phase 3 (Lessons 11 to 16) covers the full training pipeline: data processing, pretraining, supervised fine-tuning, LoRA parameter-efficient fine-tuning, knowledge distillation, and assembling a complete Transformer block into a full model.

Phase 4 (Lessons 17 to 22) adds DPO preference optimization, PPO and GRPO reinforcement learning, mixture-of-experts (MoE) models, inference optimization with KV-Cache and YaRN, and deployment. Lesson 22 is an interview guide.

Phase 5 (Lessons 23 and 24) is explicitly for job search: a resume writing guide with four template versions for four target roles, and a complete STAR interview script with self-introduction templates and 12 rounds of simulated interview.

190+ Interview Questions Organized by Topic

The interview question bank in the interview/ directory is a major part of what distinguishes this repository from a generic LLM tutorial. The questions are split into two tiers.

The basic interview set covers project introductions (three-length templates: 30 seconds, one minute, three minutes), model architecture questions (28 questions on Transformer, GQA, RoPE, and RMSNorm), training pipeline questions (30 or more questions on pretrain through DPO and PPO), optimization and deployment (20 or more on KV-Cache, MoE, quantization, and inference acceleration), and comprehensive follow-up questions (15 or more on advanced topics).

The deep interview set includes five additional collections with a combined focus on mathematical depth. Transformer deep questions (30 questions, each with mathematical derivations), training pipeline questions (50 questions framing pretraining through GRPO in a unified preference optimization framework), inference optimization (30 questions on Flash Attention, quantization, and vLLM, with precise VRAM calculations), engineering practice (30 questions on AMP, gradient accumulation, DDP, and checkpoint recovery, with code implementations), and MiniMind-specific questions (50 questions that focus on what a hiring manager would ask about this specific project).

The interview quick-reference table in the README maps interview scenarios directly to which files to read, so a reader can prepare for a specific interview type without reading everything.

Manga Illustrations and Output Formats

Fifteen original manga-style illustrations in the assets/comics/ directory explain concepts such as self-attention, RoPE encoding, LoRA fine-tuning, DPO, and MoE. These illustrations are referenced from the lesson files and are one of the repository's distinguishing features for visual learners who struggle with mathematical notation before seeing a concrete analogy.

The repository also provides the curriculum in three output formats: Markdown (the source files in docs/), HTML, and PDF. The web/ directory suggests a browser-based reading experience as well. Having the material in PDF and HTML means a learner can study on mobile or a PDF reader without needing a Markdown viewer.

Three Learning Paths and Their Time Estimates

The README gives concrete time estimates for each path. The three-day fast-track path is for learners with deep learning background: Day 1 covers the introductory and attention lessons plus the pretraining lesson, Day 2 covers SFT, DPO, and LoRA plus the core interview questions on training, and Day 3 is resume writing and STAR practice. The focus is on interview readiness rather than comprehensive understanding.

The seven-day systematic path adds proper time with each component: two days on foundations and PyTorch, two days on model components, two days on the full training pipeline, and a final day combining DPO, PPO, MoE, inference optimization, and all interview materials.

The fourteen-day zero-base path is designed for readers with no prior Python or deep learning background. The README frames the goal as: write the code yourself, pass interviews, and complete the resume. The manga illustrations and basic Python explanations in this path assume nothing.

Limitations: Language, Recency, and Scope

The repository material is written in Chinese. Lesson files, interview questions, STAR scripts, and the resume guide are all in Chinese. An English-speaking learner looking for an explanation of MiniMind will find almost none of the prose readable without translation. The repository does link to the parent MiniMind project, which has its own English-language README, but Learn MiniMind's tutorial content is not available in English.

The last push to the repository was on 2026-04-01. The MiniMind parent project may have received updates since then; if the MiniMind codebase has changed, some lesson files or interview questions may reference an older version. The repository has no GitHub releases, so there is no tagged stable version to compare against.

The course is specifically oriented toward the Chinese AI job market. The resume templates, STAR scripts, and 'offer letter' language in the README reflect Chinese tech industry hiring practices. Learners outside that context will find the concepts transferable but the interview and resume material less directly applicable.

Comparison with Other Small LLM Learning Resources

The most widely cited comparable effort is Andrej Karpathy's nanoGPT and his 'Let's build GPT from scratch' video, which covers similar territory: building a GPT-style model from first principles with PyTorch, starting from tokenization and ending with a trained model. The principal difference is scope: nanoGPT is a codebase and a video lecture without an interview preparation component. Learn MiniMind adds the job-search layer explicitly.

Another alternative is the official MiniMind repository itself, which includes a detailed README explaining the architecture and training steps. Learn MiniMind restructures that information into a lesson sequence with more granular explanations for each component, adds the manga illustrations, and appends the full interview question bank. A reader who can follow the MiniMind README directly does not need this repository; Learn MiniMind is for those who find the source code and architecture notes insufficient without more scaffolded explanation.

Editorial conclusion

Learn MiniMind is well-suited for engineers in China's job market who want to use the MiniMind project as a resume item and need a structured path from concept to interview-ready explanation. The material is primarily in Chinese; readers who need English-language resources will find almost none of the prose usable directly. The last recorded push to the repository was on 2026-04-01, so recent additions to the MiniMind parent project may not be reflected in the course material yet. Before starting, check whether the lesson files in the docs/ directory cover the specific MiniMind version you intend to train, and verify that the interview question bank reflects the MiniMind codebase you have cloned.

Frequently asked questions

Official sources

1. bcefghj/learn-minimind on GitHub
2. Issues
3. License: MIT
4. Project website
5. README