Show HN: I trained a 125M model to autocomplete piano on-device
A developer shares their experience training a 125M-parameter transformer model to perform real-time piano accompaniment on an iPhone. The project highlights the importance of MIDI data representation and post-training optimization for on-device performance.
Why it matters
It demonstrates the growing feasibility of running sophisticated generative AI models locally on mobile hardware.
TL;DR: I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The biggest improvements came from finding the right MIDI representation, cleaning the training data aggressively, and adding DPO post-training.
Almost a year ago, I started tinkering with an idea: connect my MIDI piano to my phone, play something, and have AI autocomplete the song for me. Think GitHub Copilot, but for piano.
It turned out to be a deeper rabbit hole than I expected. Fourteen experiments later, it is finally at a point where I am happy enough with it to write about.
The app, RollTab, is available for free here if you have a MIDI keyboard and an iPhone/iPad. 1
Each audio starts with a short prompt, followed by the model's continuation.
Your browser does not support the audio tag.
Final Fantasy VI, Terra's Theme (16-note prompt)
Get smarter about the news
Sign up free for a feed built around what you actually care about, Dive Deeper research on any story, and the full text of every article.
Create free accountAlready have an account? Sign in