Skip to main content
Back

Hands-on LLM Hits 2.5k Stars Teaching the Full Open-Source AI Stack

A hands-on Chinese-language LLM curriculum bundles video, docs, and runnable code across fine-tuning, RAG, agents, and deployment into one path.

Pro

Repo, Full Access
  • cgft-llm hit 2.5k stars as a hands-on Chinese-language curriculum spanning the full modern LLM stack.
  • Section 2 packs 21 lessons on fine-tuning, RAG, agents, and deployment tools.
  • Covers llama-factory, llama.cpp, vLLM, Ollama, Dify, Milvus, Langfuse, and fastmcp with runnable code.
  • Includes DeepSeek-R1 fine-tuning, graph-rag local deployment, and minimal browser-use style tool calling.
  • Teaches production hygiene with uv, ruff, pre-commit, pytest, Docker, and Gradio.
  • MIT licensed, all free, but video lectures are Mandarin-only.

A GitHub repository called cgft-llm has been quietly building itself into one of the more complete hands-on LLM curriculums available, and it just crossed the 2.5k star mark. It bundles video tutorials, written docs, and runnable code into a single learning path that walks through the full modern LLM stack, from fine-tuning a base model to shipping an agent behind a Docker container.

The repo is authored by echonoshy and organized around a series called Hands-on LLM. Every topic ships with a Bilibili or YouTube video, a markdown doc, and code you can actually run.

What's actually inside

The material is split into six numbered folders, with the density of the middle sections carrying most of the value. Section 2, LLM core technology, alone contains 21 lessons covering the tools most teams end up touching.

The fine-tuning and deployment track walks through the workhorse open-source stack:

  • llama-factory for LoRA and full fine-tuning of Llama-3 and DeepSeek-R1, including a dedicated lesson on constructing training data for reasoning models