MAS: A Millennium of Arabic Manuscripts in Three Styles

MAS (Medieval Arabic Script) is a line-level OCR benchmark of 11,841 expert-annotated lines from nine authentic Arabic manuscript works. It covers Naskh, Taliq, and Nastaliq. Copies date from the 12th to the 20th century; compositions from the 10th to the early 20th.

This dataset accompanies the ICDAR 2026 paper A Millennium of Arabic Manuscripts in Three Styles: A Line-Level OCR Benchmark for Naskh, Taliq, and Nastaliq.

Training configurations, the lmms-eval task, and evaluation notes are in the MAS GitHub repository.

Calligraphic styles in MAS, left to right: Naskh, Taliq, Nastaliq.

Splits

The dataset exposes one configuration, mas:

Split	Lines	Share
train	9,472	80%
test	2,369	20%

The split is 80/20 at image level with no page overlap. All nine works occur in both splits, so this is not a manuscript-held-out evaluation. The published test split is the held-out evaluation set used in the paper.

from datasets import load_dataset

dataset = load_dataset("maximazzik/MAS", "mas")
print(dataset)
print(dataset["train"][0].keys())

Coverage

The paper groups the collection into five domains: astronomy, history, mathematics, religion, and Sufi literature.

The domain field uses six machine-readable labels. siyasat (advice to rulers) and prayers together correspond to the paper's Religion group:

Paper domain	domain	domain_label
Astronomy	astronomy	Astronomy
History	history	History
Mathematics	mathematics	Mathematics
Religion	siyasat	Advice to rulers (siyasat-nama)
Religion	prayers	Prayer collection
Sufi literature	sufi_lit	Sufi literature

Record schema

Field	Meaning
image	Embedded line image
text	Diplomatic UTF-8 transcription
prompt	OCR instruction used to construct the sample
messages	ShareGPT-style user/assistant messages
script	Naskh, Taliq, or Nastaliq
domain, domain_label	Machine-readable and display domain names
work, author, author_dates	Bibliographic metadata
composition_date, copy_date	Work and manuscript chronology
manuscript_id	Stable identifier for one of the nine works
source_pack, line_id	Source grouping and line-level traceability
split	train or test

Annotation

Transcriptions are diplomatic: original spelling, punctuation, dots, diacritics, and historical forms are preserved without modernization or correction. Each line was independently transcribed by two trained researchers, with disagreements resolved through cross-validation and discussion.

Evaluation

Open-source LVLMs were evaluated with lmms-eval. Specialized OCR systems and closed-source models were scored from saved predictions using the same CER/WER definitions. See docs/reproduction.md and docs/results.md.

Citation

If you use the MAS dataset, please cite:

@inproceedings{novopoltsev2027mas,
  title={A Millennium of Arabic Manuscripts in Three Styles: A Line-Level OCR Benchmark for Naskh, Taliq, and Nastaliq},
  author={Novopoltsev, Maxim and Murtazin, Ruslan and Sakhovskiy, Andrey and Bojarskaja, Emilia and Kokh, Vladimir and Ulitin, Ivan and Abdullayev, Botirjon and Aminov, Khamidulla and Ismoilov, Masudkhon and Budennyy, Semen},
  booktitle={Document Analysis and Recognition -- ICDAR 2026},
  pages={643--659},
  year={2027},
  publisher={Springer Nature Switzerland},
  doi={10.1007/978-3-032-36033-5_38}
}