TL;DR
An educational framework that implements a Transformer LLM from scratch in PyTorch, covering the entire process from pretraining to alignment.
Key features
Direct implementation of Transformer model in PyTorch (based on Attention Is All You Need)
Pretraining support using the Pile dataset
Pure PyTorch implementation of alignment algorithms: SFT, Reward Model, PPO, DPO, GRPO
Capable of training models with millions to billions of parameters on a single GPU
Supports DDP + bf16 multi-GPU training
Utilizes real public datasets (Alpaca, Dolly, HH-RLHF, UltraFeedback, GSM8K)
When to use it
When you want to learn the internal workings of LLMs
When you want to implement the entire pipeline from pretraining to alignment yourself
When you want to experiment with pure PyTorch without relying on existing libraries (trl, peft, transformers)