liguodongiot/llm-action

What it is

llm-action is a collection of hands-on tutorials, articles, and sample code for learning how to train, fine-tune, run, compress, and evaluate large language models (LLMs). It is written mainly in Chinese and is maintained by an author known as “吃果冻不吐果冻皮”. The repository is essentially a guided tour through the entire lifecycle of an LLM project, from setting up a server to putting a model into production.

What's inside

The README is organized like a full curriculum. It covers:

• Training – practical walkthroughs for models like Alpaca, ChatGLM, Vicuna, and LLaMA, using techniques such as full fine-tuning, LoRA, QLoRA, and RLHF (reinforcement learning from human feedback). Each walkthrough links to a detailed article (mostly on Zhihu) and often includes the matching code.
• Fine-tuning theory – a seven-part series explaining what “parameter‑efficient” fine‑tuning is and how methods like BitFit, P‑Tuning, Adapter, LoRA, and QLoRA work.
• Distributed training – articles on splitting training across many GPUs (data parallelism, pipeline parallelism, tensor parallelism, etc.) and how frameworks like PyTorch, Megatron-LM, and DeepSpeed do this.
• Inference – how to make a trained model answer quickly, including guides for famous inference engines (vLLM, TensorRT‑LLM, Triton) and optimization tricks like KV Cache or continuous batching.
• Compression – making models smaller and cheaper to run through quantization (GPTQ, AWQ, QLoRA, etc.), pruning, knowledge distillation, and low‑rank decomposition — each with explanatory articles.
• Evaluation – benchmarks to test how good a model is (C‑Eval, CMMLU, LongBench, SWE‑bench, etc.) and tools like OpenCompass.
• Big‑picture topics – data engineering, prompt engineering, applying LLMs to real products, adapting models for Chinese domestic hardware, AI compilers, and interview questions.

How it is organized

A detailed table of contents splits everything into clear sections. Each section contains either:

• a list of articles (with links), or
• a table that pairs a tutorial with its accompanying code folder in the repository, or
• a brief note like “待更新…” (to be updated) when there is nothing yet.

So you can jump straight to the topic you care about and follow the links exactly as the author prepared them.

Who it is for

It is aimed at people who want practical, guided learning about LLMs — developers, students, or researchers who want to go from reading theory to actually running training or fine-tuning on their own hardware. The author also maintains a WeChat group and a public account, suggesting a community for learners to ask questions.

Code and community

The repository contains ready‑to‑use code (e.g., llm-train/alpaca, llm-train/chatglm-lora, llm-train/peft) that matches the tutorials. The badge links point to the author’s pages on Zhihu, Juejin, CSDN — where most of the articles live — and a WeChat contact for community discussions.

Note: Everything above is taken directly from the README. The project does not claim to be an official framework or library; it is a curated set of learning materials and companion code.