Datasets:

guangyangmusic
/

OpenScore-StringQuartets

license: cc-by-4.0
task_categories:
  - image-to-text
language:
  - en
tags:
  - music
  - optical-music-recognition
  - OMR
  - MusicXML
  - sheet-music
pretty_name: OpenScore String Quartets (OMR Evaluation)
size_categories:
  - n<1K
dataset_info:
  features:
    - name: image_imslp
      dtype: image
    - name: image_mscore
      dtype: image
    - name: musicxml
      dtype: string
    - name: filename
      dtype: string
  splits:
    - name: train
      num_examples: 252

OpenScore String Quartets (OMR Evaluation)

This dataset is derived from the OpenScore String Quartets corpus (Gotham et al., 2023), a collection of string quartets by "long 19th century" composers. It is designed for evaluating Optical Music Recognition (OMR) systems.

We extract a subset of the OpenScore String Quartets that contains both scanned images of real scores and the corresponding MusicXML ground truth. We also render clean images from the MusicXML files using MuseScore. This gives two evaluation settings:

Camera (image_imslp): scanned page images from IMSLP, reflecting real-world conditions (noise, aging, varying scan quality).
Rendered (image_mscore): clean page images rendered from the ground-truth MusicXML via MuseScore.

Dataset Structure

Column	Type	Description
`image_imslp`	`Image`	Scanned page image from IMSLP (camera)
`image_mscore`	`Image`	Clean rendered page image from MuseScore
`musicxml`	`string`	Ground-truth MusicXML content
`filename`	`string`	Composition and page identifier

The dataset contains 252 page-level examples.

Usage

from datasets import load_dataset

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

# access an example
example = ds[0]
example["image_imslp"].show()   # scanned image
example["image_mscore"].show()  # rendered image
print(example["musicxml"][:200])
print(example["filename"])

Source

The original OpenScore String Quartets corpus is available at https://github.com/OpenScore/StringQuartets and on MuseScore.

License

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

This repository is a derived resource (subset selection, IMSLP page images, MuseScore renders, 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 corpus (see citation below).

Citation

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

This dataset / Legato paper:

@misc{yang2025legatolargescaleendtoendgeneralizable,
      title={LEGATO: Large-scale End-to-end Generalizable Approach to Typeset OMR},
      author={Guang Yang and Victoria Ebert and Nazif Tamer and Brian Siyuan Zheng and Luiza Pozzobon and Noah A. Smith},
      year={2025},
      eprint={2506.19065},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.19065},
}

Original OpenScore String Quartets corpus:

@inproceedings{gotham_openscore_2023,
    address = {Milan, Italy},
    title = {The "{OpenScore} {String} {Quartet}" {Corpus}},
    isbn = {9798400708336},
    url = {https://dl.acm.org/doi/10.1145/3625135.3625155},
    doi = {10.1145/3625135.3625155},
    booktitle = {Proceedings of the 10th International Conference on Digital Libraries for Musicology},
    publisher = {ACM},
    author = {Gotham, Mark R. H. and Redbond, Maureen and Bower, Bruno and Jonas, Peter},
    month = nov,
    year = {2023},
    pages = {49--57},
}