OLiMPiC — OpenScore Lieder Linearized MusicXML Piano Corpus

A HuggingFace-formatted mirror of the OLiMPiC dataset for end-to-end optical music recognition of pianoform music.

Dataset description

OLiMPiC provides system-level (one staff row) crops of piano scores paired with ground-truth annotations in Linearized MusicXML (LMX) format. Each sample is one system — the smallest unit that makes musical sense for training sequence models.

• Synthetic variant: 17,945 rendered systems (train/dev/test)
• Scanned variant: ~2,900 real IMSLP scans (dev/test only)
• Source: OpenScore Lieder corpus (1,356 manually verified scores)

Format

{
    "image":       PIL.Image,   # system-level crop (one row of grand staff)
    "lmx":         str,         # Linearized MusicXML token sequence
    "musicxml":    str,         # full MusicXML for this system
    "score_id":    str,         # OpenScore score identifier
    "page_system": str,         # e.g. "p2-s3" (page 2, system 3)
    "source":      str,         # "synthetic" or "scanned"
    "split":       str,         # "train", "dev", or "test"
}

Usage

from datasets import load_dataset

ds = load_dataset("zzsi/olimpic")
example = ds["train"][0]
print(example["lmx"][:200])
example["image"].show()

License

CC BY-SA 4.0

Attribution

Please cite the original work:

@inproceedings{OLiMPiC,
  title     = {Practical End-to-End Optical Music Recognition for Pianoform Music},
  author    = {Fier, Jiří and Hajič, Jan},
  booktitle = {International Conference on Document Analysis and Recognition (ICDAR)},
  year      = {2024}
}

Original dataset: https://github.com/ufal/olimpic-icdar24 Original authors: Jiří Fier, Jan Hajič (UFAL, Charles University)