marisa0v0/BEAT-ICML2026-piano-dataset

BEAT ICML 2026 Piano Dataset

This repository hosts the piano-score dataset release for BEAT: Tokenizing and Generating Symbolic Music by Uniform Temporal Steps (ICML 2026).

Released archive

BEAT-ICML2026-piano-v3-dedup-musicxml.tar.gz contains 169,283 deduplicated piano scores in score-partwise MusicXML format. This is the v3 deduplicated release of the MuseScore-derived XML portion of the BEAT piano corpus. The paper reports a broader piano training collection of approximately 208K pieces: about 193K MuseScore pieces before release filtering/deduplication, plus about 15K piano pieces from the Lakh MIDI Dataset. The archive here does not include that separate Lakh MIDI subset.

• —Files: 169,283 .musicxml
• —Archive size: 1,645,907,562 bytes
• —SHA-256: b39d2bb79fae10f96e60e782170b7af783215d8b942089b2eccbef354d650781

Privacy cleanup

The released copy removes document-level MusicXML identity and provenance fields, including work/movement titles, creators, rights, source and encoding information, uploader/miscellaneous fields, page credits, comments, and processing instructions. Notes, score structure, part and instrument names, lyrics, directions, harmony, tempo, and dynamics are preserved. The original source files and metadata cache are not included.

Using the data with BEAT

Clone BEAT-code, extract the archive into data/piano_xml/, and run:

python -m data_prep.xml2pianonpz --workers 32
accelerate launch --multi_gpu --num_processes 4 -m scripts.train_piano

The v3 preprocessor expands repeat bars and voltas by default, creates the NPZ representation, and writes stable 80/10/10 train/eval/test manifests. See the BEAT repository README for the full data layout and options.

Rights notice

Removing embedded document metadata does not alter any rights that may apply to the underlying musical works or source scores. Users are responsible for ensuring that their access and use comply with applicable terms and law.

Citation

@article{qian2026beat,
  title   = {BEAT: Tokenizing and Generating Symbolic Music by Uniform Temporal Steps},
  author  = {Qian, Lekai and Gu, Haoyu and Zhao, Jingwei and Wang, Ziyu},
  journal = {arXiv preprint arXiv:2604.19532},
  year    = {2026}
}