btrkeks/polish-scores

Polish Historical-Scan OMR Benchmark

A page-level Optical Music Recognition (OMR) evaluation benchmark of 112 real historical score scans, paired with both **kern (Humdrum) and MusicXML ground-truth transcriptions.

Derived from the PRAIG/polish-scores dataset, with kern normalization and a manual-fix pass applied. Released as the real-scan half of the Transcoda evaluation suite alongside btrkeks/verovio-synth-omr and the btrkeks/transcoda-59M-zeroshot-v1 model.

Intended Use

Evaluation only. Do not include in training data.

In the Transcoda paper and benchmarks, no Polish samples are used for model training or selection — this dataset exists exclusively to measure out-of-distribution OMR performance on real historical scans.

Dataset Structure

A single unsplit table.

Column	Type	Description
image	Image (RGB)	Scan of a Polish historical score page, resized to 1485 × 1050 pixels (width × height).
transcription_kern	string	Ground-truth **kern (Humdrum) transcription.
transcription_xml	string	Ground-truth MusicXML transcription of the same content.

Image contract: each scan is resized to width 1050 preserving aspect ratio, then bottom-padded with white or top-cropped to height 1485.

Provenance

1. Upstream: PRAIG/polish-scores (curated by the Pattern Recognition and Artificial Intelligence Group, University of Alicante).
2. Raw .ekern transcriptions extracted from the upstream dataset.
3. Converted to standard **kern by stripping the extra .ekern annotation symbols.
4. Multi-pass normalization (spine cleanup, interpretation spacing repair, rest folding, etc.).
5. Manual curation pass to fix transcription errors discovered during benchmark validation.
6. Images resized to the 1485 × 1050 page contract.

Loading

from datasets import load_dataset

ds = load_dataset("btrkeks/polish-scores")
sample = ds["train"][0]
sample["image"]                 # PIL.Image, 1050 × 1485
sample["transcription_kern"]    # str (**kern)
sample["transcription_xml"]     # str (MusicXML)

Benchmark Numbers

OMR-NED on this dataset (lower is better):

Model	Params	OMR-NED ↓
SMT++	11M	80.16%
Legato	943M	86.73%
Transcoda 59M (beam search)	59M	63.97%
Transcoda FCMAE + ConvNeXt-V2-Base	120M	60.7%

See scripts/benchmark/README.md in the project repository for reproduction commands and full ablations.

License

The upstream PRAIG/polish-scores dataset does not declare a license at the time of this writing. This derivative is therefore released under license: other until upstream terms are clarified.

If you intend to redistribute or use this data commercially, please contact PRAIG (University of Alicante) directly about the original scans, and treat this card's other license as a placeholder rather than a grant.

Limitations

• Small (112 samples) — high per-sample variance in metrics; use as a qualitative OOD probe, not a high-precision leaderboard.
• Historical scans include genuine degradation (paper aging, ink bleed, faint staves, hand-written annotations).
• Manual curation was applied; transcriptions reflect the curator's reading, not necessarily the upstream PRAIG ground truth.

Citation

Please cite both the upstream PRAIG dataset and the Transcoda paper when using this benchmark:

@misc{dratschuk2026transcodaendtoendzeroshotoptical,
      title={Transcoda: End-to-End Zero-Shot Optical Music Recognition via Data-Centric Synthetic Training},
      author={Daniel Dratschuk and Paul Swoboda},
      year={2026},
      eprint={2605.10835},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.10835},
}

@misc{praig_polish_scores,
  title  = {Polish Scores Dataset},
  author = {Pattern Recognition and Artificial Intelligence Group (PRAIG)},
  howpublished = {\url{https://huggingface.co/datasets/PRAIG/polish-scores}}
}