Katsukuten/Harmonia-LM ? reverse-engineered prompt

Build me a MIDI music generation project that can train and sample a causal language model on a MIDI dataset.

I want it to take raw MIDI files, convert them into tokens with a MIDI tokenizer, split them into fixed length chunks, and cache the processed data so training is fast. Then train a Hugging Face compatible causal LM with a PyTorch Lightning style workflow, save the best checkpoint, and support simple inference where I can give a short primer MIDI and get a continuation back as a new MIDI file.

It should work with models like Qwen3 or any other causal model from Transformers, and it should include sensible generation settings for music, like low temperature, a small repetition penalty, and top p sampling. If possible, add a basic notebook or demo script so I can try training and generation end to end with the included sample MIDI files or the MAESTRO dataset. Look up current docs online if you need to.