GigaMIDI

Summary

The Extended GigaMIDI Dataset is the largest symbolic music collection to date, comprising over 2.1 million unique MIDI files with detailed annotations for music loop detection and expressive performance characteristics. It introduces a novel expressive loop detection method using the Note Onset Median Metric Level (NOMML) heuristic, identifying 9.2 million non-expressive loops and 2.3 million expressive loops across all General MIDI instruments. The dataset encompasses several major symbolic MIDI resources including MetaMIDI, Lakh MIDI, XMIDI, top-MAGD, and MASD, and includes a curated human-annotated genre/style subset. It was used to train an expressive multitrack symbolic music loop generation model via the MIDI-GPT system.

Keywords

midi symbolic-music music-generation loop-detection expressive-performance music-information-retrieval genre-classification

Topics

Music / Symbolic Music AI

Research notes

• Requires agreeing to share contact information to access. Published in Transactions of the International Society for Music Information Retrieval (TISMIR). Extended version under review at NeurIPS 2025 Dataset Track for Creative AI. GitHub: github.com/Metacreation-Lab/GigaMIDI-Dataset. Point of contact: Keon Ju Maverick Lee (keon_maverick@sfu.ca). Train/Val/Test split: 80%/10%/10%.