karpathy/minGPT ? reverse-engineered prompt

Build me a small educational Python project that reimplements a GPT style language model in PyTorch, focused on being easy to read and learn from rather than production ready.

I want the core model code to be compact and understandable, with a transformer that takes token ids and predicts the next token. Include a simple trainer, a byte pair encoder compatible with GPT style text tokens, and examples that show how to train from scratch on a tiny task like adding numbers or character level text. Also include a notebook or script that can load a pretrained GPT 2 style model and generate text from a prompt.

Please keep the code clean, minimal, and well commented so someone studying transformers can trace what’s happening. Add basic unit tests and simple install instructions so I can import the library in another project. If you need exact PyTorch details, look up current docs online.