guangyangmusic/OpenScore-Lieder

OpenScore Lieder (OMR Evaluation, Piano-Only)

This dataset is derived from the OpenScore Lieder corpus (Gotham & Jonas, 2022), a collection of 19th-century songs in MuseScore / MusicXML form. It is built for evaluating Optical Music Recognition (OMR) systems under piano-only, full-page conditions.

We follow the same broad recipe as our OpenScore String Quartets evaluation set: we use source PDFs to obtain camera page images and render clean page images from the ground-truth MusicXML with MuseScore. To compare fairly with systems that target typeset / piano notation (e.g. SMT++), we remove the vocal staff by masking it with white rectangles, leaving only the piano staves visible. This is conceptually similar to the masking strategy used in the OLiMPiC dataset (Mayer et al., 2024), but here we use full pages rather than single-system crops.

Evaluation settings

• Camera (image_imslp): page images derived from PDFs / scans (real-world appearance).
• Rendered (image_mscore): clean page images rendered from the aligned MusicXML via MuseScore, with the same piano-only masking applied.

Dataset structure

Column	Type	Description
image_imslp	Image	Camera-style page image (piano staves only; vocal staff masked)
image_mscore	Image	Rendered page image from MuseScore (same masking)
musicxml	string	Ground-truth MusicXML content
filename	string	Composition and page identifier

The dataset contains 64 page-level examples.

Versions

v1.2 (current) — In v1.1, the annotated source PDFs (white rectangles covering the vocal staff) were converted to PNG without rendering the annotations, so 12 pages still showed the voice part in image_imslp, contrary to the piano-only design described above. Those pages are re-rendered with PyMuPDF at 200 dpi with annots=True, matching how the other pages were produced. The affected pages are Barnby — The Beggar Maid 1-4, Beach — Op. 44 No. 3, I Send My Heart up to Thee 1-5, and Abbott — Just for Today 1-3. musicxml, image_mscore and filename are unchanged, and the other 52 pages are byte-identical to v1.1.

If you evaluated on v1.1, note that results on those 12 camera pages are not comparable across versions. Earlier revisions remain accessible through the repository's commit history.

Usage

from datasets import load_dataset

ds = load_dataset("guangyangmusic/OpenScore-Lieder", split="train")

example = ds[0]
example["image_imslp"].show()
example["image_mscore"].show()
print(example["musicxml"][:200])
print(example["filename"])

Source

The original OpenScore Lieder corpus is mirrored at https://github.com/OpenScore/Lieder and hosted on MuseScore.

License

The underlying OpenScore Lieder scores are in the public domain under CC0 1.0.

This repository is a derived resource (subset selection, PDF-based camera pages, MuseScore renders, vocal-staff masking, aligned MusicXML strings, and packaging). We release it under Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and adapt it if you give appropriate credit, including this dataset and the original OpenScore Lieder corpus (see citation below).

Citation

If you use this dataset, please cite both this release (Legato evaluation subset) and the original OpenScore Lieder corpus.

This dataset / Legato paper:

@inproceedings{yang2026legato,
  title={{LEGATO}: Large-scale End-to-end Generalizable Approach to Typeset {OMR}},
  author={Guang Yang and Victoria Ebert and Nazif Can Tamer and Brian Siyuan Zheng and Luiza Amador Pozzobon and Noah A. Smith},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  eprint={2506.19065},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2506.19065}
}

Original OpenScore Lieder corpus:

@inproceedings{gotham_jonas_openscore_lieder_2022,
  author    = {Gotham, Mark R. H. and Jonas, Peter},
  title     = {The {OpenScore} {Lieder} Corpus},
  booktitle = {Proceedings of the Music Encoding Conference 2021},
  year      = {2022},
  pages     = {131--136},
  address   = {Alicante, Spain},
  publisher = {Universidad de Alicante},
  isbn      = {978-84-1302-173-7},
  url       = {http://hdl.handle.net/10045/123677},
}