Hands-On Large Language Models · Course Overview
This course has 12 chapters + bonus readings, written chapter-by-chapter alongside the official notebooks from the O'Reilly book Hands-On Large Language Models by Jay Alammar & Maarten Grootendorst: chapter structure, code, images, and the repository are kept identical; only the explanatory text is translated into Chinese.
What You Will Learn
| Stage | Chapters | Content |
|---|---|---|
| Foundations | 01–03 | Language model intro, tokens & embeddings, inside the LLM |
| Applications | 04–06 | Text classification, clustering & topic modeling, prompt engineering |
| Advanced | 07–09 | Advanced text generation techniques & tools, semantic search, multimodal LLMs |
| Training | 10–12 | Building text embedding models, fine-tuning BERT, fine-tuning generative models |
| Extensions | Bonus | Agent, diffusion models, MoE, Mamba/SSM, quantization & inference — extended readings on frontier topics |
Each chapter corresponds to the official notebook in the repository's chapter01–chapter12 directories (.ipynb). Code is preserved as-is and can be run directly in Colab.
Bonus Material
The author has written 8 visual deep-dive guides alongside the book; all are included here with illustrations:
How Transformer LLMs Work (DeepLearning.AI course) · A Visual Guide to Quantization · Mamba and State Space Models · Mixture of Experts (MoE) · The Illustrated Stable Diffusion · A Visual Guide to Reasoning LLMs · The Illustrated DeepSeek-R1 · A Visual Guide to LLM Agents
Official Reference Sources
| Source | Link |
|---|---|
| GitHub repository | HandsOnLLM/Hands-On-Large-Language-Models |
| Book website | huggingface.co/learn |
| Environment | Repository root environment.yml / requirements.txt |
Content version baseline: latest main branch of the repository.
Learning Suggestions
- Chapters 1–3 are the foundation of the entire book: tokenizers, embeddings, and the internal Transformer mechanism run through all subsequent chapters;
- Application chapters (4–8) can be read selectively: classification/clustering/semantic search are independent of each other;
- The training trilogy (10–12) builds progressively: build an embedding model first, then fine-tune BERT, and finally fine-tune a generative model;
- Notebooks default to running on GPU (Colab's free tier is sufficient for the vast majority of examples).