ISKANDAR — Line-Level HTR Ground-Truth for Arabic historical manuscripts (Oriental)

Dataset Description

ISKANDAR is a specialized dataset for Handwritten Text Recognition (HTR) focusing on Arabic historical manuscripts. It contains a curated selection of pages from the al-Iskandar (Romance of Alexander), in Middle Arabic and oriental script.

This HuggingFace dataset provides cropped line-level images paired with their transcriptions and rich metadata for 5 Arabic manuscripts (13 volumes). It is designed as a ready-to-use resource for Arabic historical HTR.

Manuscripts	5 (13 volumes)
Pages	251
Lines	5,221
Words	44,947
Characters	230,450

The full page-level dataset (PageXML + full-page images) is available on GitLab.

Source Documents

Five manuscripts dated to the 18th–19th centuries, held across three European libraries:

• BnF — Bibliothèque nationale de France (Paris): 3 manuscripts (825 folios total)
  • btv1b10031486f, btv1b11003097s, btv1b110032696
• Gotha Research Library: 1 nine-volume set (858 leaves)
  • Ms_orient_A_02385 through Ms_orient_A_02393
• University of Leeds: 1 manuscript (450 folios)
  • leeds-iskandar-ms153

Usage

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("calfa-ai/iskandar")

# Access a sample
sample = ds["train"][0]
sample["image"].show()
print(sample["transcription"])

# Filter by library
leeds = ds["train"].filter(lambda x: x["manuscript_name"].startswith("leeds-iskandar"))
gotha = ds["train"].filter(lambda x: x["manuscript_name"].startswith("Ms_orient"))

Annotation Process

Annotations were made on the Calfa Vision platform during the Alexander Hackathon (November 2022 – May 2023), mobilizing 12 students from French universities. The project aimed to train participants in HTR techniques while advancing the study of popular Levantine Arabic literature. Recognition accuracy reached 90.38–96.25% across manuscripts, with the Gotha collection presenting specific challenges due to its distinct spelling conventions and layout.

Transcription Guidelines (Summary)

The transcription guidelines follow the RASAM specification: transcriptions preserve the text as found in the image, including variant spellings, while excluding vocalization marks unless structurally significant. See RASAM 1 and original paper for details.

Citation

@inproceedings{2024rasam-dataset,
  title     = {{Enhancing Arabic Maghribi Handwritten Text Recognition with RASAM 2: A Comprehensive Dataset and Benchmarking}},
  author    = {Vidal-Gorène, Chahan and Salah, Clément and Lucas, Noëmie and Decours-Perez, Aliénor and Perrier, Antoine},
  url       = {https://enc.hal.science/hal-04722622},
  booktitle = {{Computational Humanities Research (CHR)}},
  address   = {Aarhus, Denmark},
  volume    = {3834},
  pages     = {200--216},
  year      = {2024},
  month     = dec,
}

License

This dataset is released under the Etalab Open Licence 2.0.

Acknowledgements

The Alexander Hackathon has been carried out thanks to the support of the DISTAM consortium and the UMR 8167 Orient & Méditerranée as part of the priority project "Horizons de Renouvellement des Études Arabes" (HoRÉA, "Horizons for the Renewal of Arab Studies") conducted by the Research Consortium Middle-East and Muslim Worlds (GIS MOMM). The Hackathon was hosted at the media library of the Mediterranean House for Human Sciences (MMSH) in Aix-en-Provence. The training session and the production of the HTR model were carried out by Calfa.