Show HN: I trained a 125M model to autocomplete piano on-device

A developer trained a 125M-parameter transformer model to autocomplete piano performances in real time on iPhone, achieving ~108 notes per second through careful MIDI representation design, aggressive data cleaning, and direct preference optimization (DPO) post-training. The resulting app, RollTab, functions like GitHub Copilot for piano by taking MIDI input from a connected keyboard and generating musical continuations. Key technical innovations included using a factored tokenization grammar to reduce vocabulary size and enforce syntactically valid outputs during generation.