Ahmad-Al-Dibo/ArcLM ? reverse-engineered prompt

Build me a small local first Python toolkit for causal language models that feels simple to use but covers the basics end to end. I want a compact GPT style model in PyTorch, tokenizers for both plain word based use and SentencePiece, training loops, checkpoint save and load, dataset preprocessing, evaluation helpers, experiment tracking, and easy inference utilities. It should also be able to inspect where a model came from and adapt outside checkpoints when possible.

Please include a very simple web interface I can launch from the library or the command line. In that app, I want to load a model from a local path or a model source, run text generation, choose a save path, pick a save mode, decide whether to overwrite, and save the loaded model locally. If saving works, update the model source to the new local path so it is easy to reload later.

Make it runnable out of the box with examples, docs, and basic tests. Look up current docs online if you need to.