HugoSchtr/debussy-omr-fullpage-lvl

Debussy OMR – Full-Page Level Dataset

A dataset of 101 paired samples of full-page images from handwritten music scores written by Claude Debussy (French Composer, 1862 - 1918) and their corresponding MusicXML transcriptions.

The manuscript images originate from the collections of the Bibliothèque nationale de France (BnF) and were accessed through the IIIF protocol.

Dataset Description

Property	Value
Samples	101
Documents	37
Image format	JPEG
Transcription format	MusicXML 4.0
Image width	2 872–6 650 px
Image height	2 874–7 346 px
Total size	~246 MB

Each sample represents one full page of a handwritten manuscript. The MusicXML transcription contains all the musical content on that page (all systems).

The dataset can be used to evaluate optical music recognition (OMR) models.

Transcription guidelines

Our primary objective was to produce an Optical Music Recognition (OMR) dataset designed to evaluate SOTA models on historical handwritten sources. To ensure the highest level of fidelity to the original manuscripts, we utilized MusicXML as our primary format, implementing specific encoding conventions to bridge the gap between physical notation and digital representation.

Encoding Conventions & Metadata Mapping

TODO: add html colors code.

To accurately represent the nuances of the handwritten sources within the MusicXML schema, we applied the following mapping logic:

• Implicit Symbols: Musical elements necessary for digital parsing but absent from the physical manuscript are flagged using MuseScore’s invisible function or the MusicXML attribute print-object="no".
• Symbol Mapping via Color Attributes: We repurposed color attributes to categorize specific handwritten phenomena and non-standard notation:
  • Beat Repeats (Magenta - #FF40FF): Rests encoded in this color represent beat repeat symbols.
  • Non-Orthodox Repeats (Yellow): Used to identify non-standard or irregular repeat signs.
  • Specific Trill Types: * Cyan: Indicates flat trills.
    • Brown: Indicates sharp trills.
  • Editorial & Physical Features (Green): Rests in green denote specific manuscript characteristics, such as erasures, editorial notes, or other physical artifacts.

Layout description

System breaks are indicated using the MusicXML attribute <print new-system="yes"/> to provide a basic representation of the original manuscript layout.

Vocabulary description

TODO.

Data Fields

Field	Type	Description
image	Image	Full-page image (JPEG)
musicxml	string	Full MusicXML transcription of the page
document_id	string	BnF ARK identifier (e.g., 12148_btv1b10073993b)
folio	int32	Folio/page number within the document
filename	string	Original filename stem (without extension)

Splits

Split	Samples
test	101

Usage

from datasets import load_dataset

ds = load_dataset("HugoSchtr/debussy-omr-fullpage-lvl", split="test")

sample = ds[0]
sample["image"].show()          # PIL Image
print(sample["musicxml"][:200]) # MusicXML string
print(sample["document_id"])    # e.g. "12148_btv1b10073993b"

Related Datasets

• HugoSchtr/debussy-omr-system-lvl — Same manuscripts cropped at the system (staff-system) level (491 samples).

Source

The manuscript images come from the Gallica digital library of the BnF. The MusicXML transcriptions were produced using MuseScore Studio 4.

Citation

If you use this dataset, please cite:

@misc{debussy_omr_system_lvl,
  title={Debussy OMR System-Level Dataset},
  author={Hugo Scheithauer},
  year={2026},
  url={https://huggingface.co/datasets/HugoSchtr/debussy-omr-system-lvl}
}