calfa-ai/RASAM-1

RASAM 1 — Line-Level HTR Ground-Truth for Arabic historical manuscripts (Maghrebi)

Dataset Description

RASAM 1 is a specialized dataset for Handwritten Text Recognition (HTR) focusing on Arabic historical manuscripts in Maghrebi script.

This HuggingFace dataset provides cropped line-level images paired with their transcriptions and rich metadata for 3 Arabic Maghrebi manuscripts from the BULAC Library. It is designed as a ready-to-use resource for benchmarking and training HTR models on under-resourced historical scripts.

Manuscripts	3
Pages	299
Lines	7,331
Words	97,311
Characters	492,221

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

Included in: this dataset is part of AMIDDA, an aggregated corpus of 66,779 lines for training generic Arabic ATR models.

Source Documents

The corpus consists of three manuscripts from the BULAC Library (Paris), dating from the 17th to the 19th centuries:

• BULAC_MS_ARA_417 (Historical)
• BULAC_MS_ARA_609 (Inheritance law)
• BULAC_MS_ARA_1977 (Historical)

Usage

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("calfa-ai/rasam-1")

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

# Filter by manuscript
bulac_609 = ds["train"].filter(lambda x: x["manuscript_name"].startswith("BULAC_MS_ARA_609"))

Transcription Guidelines (Summary)

The RASAM transcription protocol prioritizes fidelity to the scribe's practice over modern standard orthography to ensure robust HTR training. Key principles include:

• Orthographic Preservation: Scribes' specific spellings (including archaisms and variant usages of characters like dād/ẓāʾ or ṣād/ṭāʾ) are maintained exactly as they appear in the manuscript.
• Normalization Exceptions: To facilitate computational consistency, specific glyphic forms for relative pronouns (e.g., allaḏī/allatī) and the preposition fī are standardized.
• Diacritics & Hamza: Vocalization and diacritics are generally excluded unless semantically essential. Hamzas are only transcribed if explicitly present in the source.
• Handling Uncertainty: Illegible or damaged characters are marked with #. Gaps in the source text are not reconstructed.
• Formatting: Spaces are restored in the transcription regardless of their visual presence in the manuscript layout.

Source: For a detailed breakdown of specifications and paleographical considerations, please refer to the full paper (see below)

Citation

@InProceedings{2021rasam-dataset,
  author    = {Vidal-Gorène, Chahan and Lucas, Noëmie and Salah, Clément and Decours-Perez, Aliénor and Dupin, Boris},
  editor    = {Barney Smith, Elisa H. and Pal, Umapada},
  title     = {RASAM -- A Dataset for the Recognition and Analysis of Scripts in Arabic Maghrebi},
  booktitle = {Document Analysis and Recognition -- ICDAR 2021 Workshops},
  year      = {2021},
  publisher = {Springer International Publishing},
  address   = {Cham},
  pages     = {265--281},
  isbn      = {978-3-030-86198-8}
}

License

This dataset is released under the Apache License 2.0.

Acknowledgements

The RASAM dataset was developed collaboratively between 2021–2023 by the Research Consortium Middle-East and Muslim Worlds (GIS MOMM), DISTAM, Calfa, and the BULAC Library. The project was funded by the French Ministry of Higher Education, Research and Innovation.