ENC-PSL/BSICLE-medieval-illumination-folio-bin-class-dataset

BSICLE Medieval Folio Illumination Dataset

This dataset contains 1,484 medieval and early modern folio images, dating approximately from the 7th to the mid-17th century. annotated for binary image classification: whether a folio contains illumination or not. The dataset is intended to train and evaluate lightweight computer vision models for the fast detection of illuminated folios in medieval manuscript corpora, especially in IIIF-based heritage and research workflows.

Check models: BSICLE — Binary System for Illuminated Folio Classification with Lightweight Engines for models trained on this dataset.

Labels

The dataset uses two labels:

Label ID	Label name
0	non_illuminated
1	illuminated

Dataset structure

The dataset is distributed as a single Parquet file.

Column name	Description
image	Image data, stored as a Hugging Face Image feature.
label	Binary class label: 0 for non_illuminated, 1 for illuminated.
label_name	Human-readable class label: non_illuminated or illuminated.
filename	Normalized filename in the form prefix_sha1.ext.
sha1	SHA-1 checksum of the original image file.
source_split	Original local split before merging: train, val, or test.
mime_type	Detected MIME type of the image file.

Data provenance

This dataset is a curated mixture of existing medieval manuscript image datasets and manually corrected samples. It was assembled to support the training and evaluation of BSICLE models for binary folio-level classification: non_illuminated vs. illuminated. The dataset covers medieval and early Renaissance manuscript and early printed material, with an approximate chronological range from the 7th to the 17th century. This range reflects the documented scope of the source datasets and selected IIIF manifests, but individual images may require consultation of the original repository or holding institution for precise dating. The source datasets were aggregated, harmonised, and partially relabelled. Some samples were corrected manually, while others were labelled or filtered through semi-automatic procedures based on available metadata, folder structures, file naming conventions, or source-specific encoding practices used to describe folio layout, for example SegmOnto.

Because the original datasets were not always designed for binary illumination detection, the labels should be interpreted as research-oriented annotations rather than definitive art-historical judgements.

This dataset may contain residual errors. Although manual correction was performed, several steps of the collection, filtering, and labelling workflow were automated. In particular, some source datasets were recovered or labelled using metadata, directory structure, filename conventions, or encoding information. These procedures may introduce mistakes, especially for ambiguous folios, decorated initials, marginal ornamentation, diagrams, damaged images, cropped details, or partially illuminated pages.

Source datasets

Dataset	Date range	URL
Pinche, A., Gabay, S., Leroy, N., & Christensen, K. Données HTR incunables du 15e siècle [Computer software].	1400–1500	https://github.com/Gallicorpora/HTR-incunable-15e-siecle
Pinche, A., Gabay, S., Leroy, N., & Christensen, K. Données HTR manuscrits du 15e siècle [Computer software].	1400–1500	https://github.com/Gallicorpora/HTR-MSS-15e-Siecle
Clérice, T., Chagué, A., & Vlachou-Efstathiou, M. CREMMA Medii Aevi [Data set].	1100-1599	https://github.com/HTR-United/CREMMA-Medieval-LAT
Claustre, J., Smith, D., Torres Aguilar, S., Bretthauer, I., Brochard, P., Canteaut, O., Cottereau, E., Delivré, F., Denglos, M., Jolivet, V., Julerot, V., Kouamé, T., Lusset, E., Massoni, A., Nadiras, S., Perreaux, N., Regazzi, H., & Treglia, M. (2023). The e-NDP project: collaborative digital edition of the Chapter registers of Notre-Dame of Paris (1326–1504). Ground-truth for handwriting text recognition (HTR) on late medieval manuscripts [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7575693	1326-1504	https://zenodo.org/records/7575693
GAHOM dataset, EHESS.	600-1650	-

Additional IIIF image sources

A selection of images was also collected from individual IIIF manifests, including sources made available by Gallica / Bibliothèque nationale de France, IRHT-CNRS, and the National Library of Wales:

• https://gallica.bnf.fr/iiif/ark:/12148/btv1b55005654j/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b6000621x/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/f6kfe6w3blt9/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b10551125c/manifest.json
• https://damsssl.llgc.org.uk/iiif/2.0/5152608/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/fyi48cz77z7r/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/fv7ewbbnyhpg/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/fgm70hx12f90/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/fomv4uf186n6/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/ff4xkbh232z8/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/fxf00f1mod63/manifest.json
• https://api.irht.cnrs.fr/ark:/63955/f0hkibfh7rl4/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b10560249b/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b72000256/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b8470423c/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b105420825/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b10547049c/manifest.json
• https://gallica.bnf.fr/iiif/ark:/12148/btv1b105450180/manifest.json

License

The dataset metadata and annotations are released under CC BY 4.0. Source images may be subject to the rights and licences of their holding institutions or original datasets.

Citation

If you use this dataset in your research, please cite the following paper:

@dataset{bsicle_medieval_images_bin_class_dataset,
  title        = {BSICLE Medieval Folio Illumination Dataset},
  author       = {Terriel, Lucas and Jolivet, Vincent},
  institution  = {École nationale des chartes -- PSL},
  year         = {2026},
  url          = {https://huggingface.co/datasets/ENC-PSL/BSICLE-medieval-illumination-folio-bin-class-dataset}
}

Funding

This dataset was developed at École nationale des chartes – PSL in the context of the O.D.I.L. project .