Pretrained weights for MIDIT, a BERT-style masked language model for symbolic music infilling (REMI+ tokenization, ~42.5M parameters).

Code, usage and full results: https://github.com/SimoneZanetti00/MIDIT

Weights contain the Lightning state_dict + hyper_parameters only (optimizer state removed). Load with the model class from the repository.

Trained on the Lakh MIDI Dataset; released for research purposes.