midi-gpt

IsoFLOP scaling-law checkpoints for MIDI language models. Each subfolder is one (compute budget, model size) run. Load with:

from transformers import AutoModelForCausalLM, AutoTokenizer

subfolder = "C3e15-d64"
model = AutoModelForCausalLM.from_pretrained("wmatejuk/midi-gpt", subfolder=subfolder)
tokenizer = AutoTokenizer.from_pretrained("wmatejuk/midi-gpt", subfolder=subfolder)

Subfolder names are {budget}-{width}, e.g. C3e15-d64. Each folder also stores midi_codec.json (how notes were tokenized) and summary.json (N, D, FLOPs, val loss).