Show HN: I Trained a 125M Model to Autocomplete Piano On-Device

Key points

1. The key design choice was a unified note token format encoding pitch, timing, duration, and velocity together, so the model advances by complete notes rather than individual attributes.
2. Post-training with Direct Preference Optimization using Gemini-based pairwise evaluation boosted preference scores from 24.5% to 69%, and the model powers the RollTab iOS app.