younissk/nanoBeard — reverse-engineered prompt

Build me a small Python project for a pirate themed text generator called nanoBeard. I want it to feel like a learning friendly tiny GPT style model, trained from scratch on pirate versions of simple story data, then tuned a bit more so it answers in pirate flavored language. It should let me build the dataset, train a model version like Sloop, generate sample text from a prompt, and run basic evaluation so I can see loss and a little sample gallery.

Please keep it set up so more model versions can be added later under the same project, not just one fixed model. I also want a simple browser playground where I can try the published versions and switch between them. Include good docs, a clean way to publish a trained model and tokenizer, and tests that make sure new model versions still behave correctly.

This is meant to be educational and lightweight, not a serious chat product. If you need details, look up current docs online.