Repolode
Submit
changyeyu

changyeyu/LLM-RL-Visualized

A bilingual collection of 100+ original diagrams explaining LLM, RL, RLHF, DPO, GRPO and related large-model algorithms.

View on GitHubWebsite
Stars
4,958
Forks
478
Open issues
3
Last commit
last month

What is it?

This repository is a visual guide to large language model and reinforcement learning techniques. It contains more than 100 original architecture diagrams covering LLM and VLM structure, decoding strategies, pre-training and post-training pipelines, and fine-tuning methods such as LoRA, Prefix-Tuning and SFT.

The second half of the material focuses on alignment and reasoning: DPO and RLHF training layouts, PPO and GRPO, reward models, chain-of-thought distillation, MCTS and Best-of-N sampling. Classic RL foundations such as Markov decision processes, Monte Carlo and TD methods, DQN, policy gradient and actor-critic are also illustrated.

Diagrams are provided as high-resolution PNG and editable SVG files, with Chinese and English versions of the text. Each section is a self-contained figure, so readers can jump directly to a topic such as LoRA or GRPO.

Key features

  • 01Over 100 original diagrams covering LLM, VLM and reinforcement learning topics
  • 02Covers fine-tuning methods including LoRA, Prefix-Tuning and SFT
  • 03Explains alignment algorithms such as RLHF, DPO, PPO and GRPO
  • 04Includes reasoning-oriented topics like chain-of-thought distillation, MCTS and Best-of-N
  • 05Provides SVG vector files and high-resolution PNGs in Chinese and English

Use cases

  • Building a visual mental model of how LLM training pipelines fit together
  • Studying the difference between PPO, GRPO and DPO before implementing one
  • Reviewing reinforcement learning fundamentals for someone entering LLM alignment work
  • Preparing slides or lecture material that needs clear algorithm diagrams

Things to know

  • The license could not be identified (listed as NOASSERTION), so reuse terms are unclear.

Topics

#llm#reinforcement-learning#rlhf#dpo#fine-tuning#diagrams#deep-learning#nlp#ai#algorithm

FAQ

What is LLM-RL-Visualized?
A bilingual collection of 100+ original diagrams explaining LLM, RL, RLHF, DPO, GRPO and related large-model algorithms. This repository is a visual guide to large language model and reinforcement learning techniques. It contains more than 100 original architecture diagrams covering LLM and VLM structure, decoding strategies, pre-training and post-training pipelines, and fine-tuning methods such as LoRA, Prefix-Tuning and SFT.
Is LLM-RL-Visualized free to use? What license does it have?
LLM-RL-Visualized is open source under the Other license. Check the license terms before using it in commercial products.
Is LLM-RL-Visualized still maintained?
Updated within 6 months. The last commit to LLM-RL-Visualized was on Sep 10, 2026.
What can I use LLM-RL-Visualized for?
Building a visual mental model of how LLM training pipelines fit together Studying the difference between PPO, GRPO and DPO before implementing one Reviewing reinforcement learning fundamentals for someone entering LLM alignment work Preparing slides or lecture material that needs clear algorithm diagrams
Where can I get LLM-RL-Visualized?
The source code and installation instructions are on GitHub: https://github.com/changyeyu/LLM-RL-Visualized

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects