Polish Historical-Scan OMR Benchmark: 112 Scored Pages

Polish Historical-Scan OMR Benchmark is a page-level Optical Music Recognition evaluation dataset containing 112 real historical score scans. It provides paired **kern (Humdrum) and MusicXML ground-truth transcriptions. The dataset was derived from the PRAIG/polish-scores dataset, with kern normalization and manual fixes applied.

Use Cases

• Benchmarking Optical Music Recognition (OMR) model performance based on real historical score scans.
• Training OMR systems on paired image and symbolic music data based on the **kern and MusicXML ground-truth.
• Evaluating the robustness of music transcription models on historical documents based on the manually-fixed ground truth.

Strengths

• Contains 112 real historical score scans, providing a realistic testbed.
• Offers paired ground-truth transcriptions in both **kern (Humdrum) and MusicXML formats.
• Includes a manual-fix pass applied to the transcriptions, suggesting improved accuracy.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Derived from the PRAIG/polish-scores dataset.

Collection Method

Scans of historical scores with applied kern normalization and manual-fix pass.

Freshness

Last updated 2026-05-12 14:01:40; freshness should be verified.

Geography

Likely Poland, inferred from the dataset title and source.