A visualized compendium of LLM/RL/VLM principles with architecture diagrams and micro-tutorials. It lists many topics and references, with the repository having 4737 stars and 454 forks.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
The repository presents original architecture diagrams and explanations focused on large language models (LLM), reinforcement learning (RL), and multimodal models (VLM/MLLM). It covers model structures, training workflows, and various fine-tuning methods, complemented by links to related resources.
How it works
The README outlines several core topics and diagrams that illustrate model components (input layer, decoder stacks, output layer, LM head), decoding strategies, RL-based training concepts (PPO, GRPO, DPO, RLHF), and SFT methods (LoRA, Prefix-Tuning, TokenID mappings, packing). It includes references to PDFs and SVG/PNG figures for visual explanations. Specific processes described include:
- Decoding strategies and their impact on text generation
- SFT and related fine-tuning techniques
- Prefix-tuning and LoRA implementations
- Packing for fixed-length inputs with position/id mask resets
Getting started
The README provides navigational structure and links to various diagrams and the English version. There is no explicit installation or usage instruction in the truncated excerpt, but it points to the repository root and image assets that can be browsed online. The repository is Python-language and has no license listed in the provided metadata.
Recent releases
RELEASES (latest 0): - none
Traction
stars_7d: none stars_1d: none
Behind the repo
No startup/company link is present in the provided facts.
Caveats
license: none listed created: 2025-04-26 last_push: 2026-07-27 open_issues: 3