This talk presents a methodology for working with very large, GPT-like deep learning models trained on (open and ethically sourced) MIDI data. This approach promotes nuanced, musical interfacing with the model, requiring practice and skill development instead of one-shot text-based prompting.
The full machine-learning pipeline is presented, including data pre-processing, tokenization, model training and inference. The presented system will be used to demonstrate multiple generative examples created through musical interaction with Large Piano Models.