Bayanat Labs Arabic Document Atlas

Arabic OCR real scanned pages document understanding word geometry source provenance historical print training data

Key numbers

Counted by script from the released files. Counts are recomputed from the exact tagged release-summary and coverage JSON files. The shard manifest SHA-256 binds the published data files. Image totals are a decoded-pixel union; community text, word records and source canvases are not additional images.

Downloads and identifiers

Get started

Stream the published page configuration

from datasets import load_dataset

pages = load_dataset(
    'Bayanat-Labs/arabic-document-atlas',
    'pages',
    split='train',
    revision='v0.3.0',
    streaming=True,
)
page = next(iter(pages))

# Separate text resource; contains no page images.
community_text = load_dataset(
    'Bayanat-Labs/arabic-document-atlas',
    'community_text',
    split='train',
    revision='v0.3.0',
    streaming=True,
)

# Separate source-associated references; no OCR or word boxes.
reference_pairs = load_dataset(
    'Bayanat-Labs/arabic-document-atlas',
    'reference_pairs',
    split='train',
    revision='v0.3.0',
    streaming=True,
)

# Source lineage for excluded exact-pixel repeats; no new images.
duplicate_occurrences = load_dataset(
    'Bayanat-Labs/arabic-document-atlas',
    'duplicate_occurrences',
    split='train',
    revision='v0.3.0',
    streaming=True,
)

What is released

Attribution: Bayanat Labs, Arabic Document Atlas. Source institutions and source-page attribution are retained per record. https://www.bayanatlabs.com/open/arabic-document-atlas

Intended uses

• Train OCR and document-processing systems using real page images and disclosed automated annotations.
• Keep page citations, source identifiers and word coordinates attached when building document viewers or retrieval chunks.
• Audit recognition noise and apply stricter filters for your target domain.
• Use community text as a separate text resource under its attribution and share-alike conditions.

Out of scope

• Treating machine OCR as verified evaluation transcription.
• Claiming an edition-heldout test set from this training release.
• Applying an annotation license to all underlying source images or community text.
• Assuming that historical library scans represent modern invoices, forms or all Arabic document domains.

Rights at a glance

Training, display, redistribution and synthetic-voice use are separate permissions.

Limitations

• Apple Vision supplies machine OCR, approximate word guides and a reading-order hypothesis. Word scores repeat the parent line’s uncalibrated confidence, not word correctness probabilities. No OCR character-error rate, semantic reading-order accuracy or human box accuracy is measured. Reproducing the native OCR requires compatible macOS/Apple Vision; its system OCR model is proprietary.
• Titles and edition years are source-catalog metadata. Bound volumes can contain other works; the pre1931 catalog filter is rights triage, not independent date verification of every page.
• Books may contain prefaces, stamps, marginalia and other languages; a book-level Arabic label is source metadata.
• Community text retains Wikisource quality labels; Bayanat Labs has not revalidated every character. It is rendered page-body text: header/footer sections are excluded, and linked scan-file versions can postdate the text revision.
• Counts describe this frozen version. Source-inventory canvases, published images and mechanically screened rows have separate definitions.
• Reference pairs preserve source proofreading labels, body scope and the frozen original file version. PDF renders may be born digital; source association is not independent image/text alignment or error-free gold transcription.
• Coverage records 768 enrolled ACO book records: 389 complete, 287 partial and 92 error. Error status can coexist with published pages; 679 books contribute images and 89 have no successful source occurrence. Failure attempts, distinct source pages and never-attempted pages have separate meanings. Exact decoded-image uniqueness does not establish globally distinct works or editions.

Sources and licences

The acquisition archive retains source responses, hashes and exact canvas identifiers. Published pages are bound to the source PDF hash and zero-based PDF page index. Canvas associations follow publisher order with equal page counts; individual canvas/PDF pairs have not all been independently verified. The separate community text preserves its own source and rights lineage. Titles and dates are preserved catalog assertions; OCR text is not used to silently relabel an edition. The reference-pair image cohort requires explicit file licenses and preserved VRT copyright-holder permission evidence; it withholds PDF files with known conflicting rights notices and source versions that postdate their text revision. Source details

How to cite

Cite as:

Bayanat Labs (2026). Arabic Document Atlas, v0.3.0. Bayanat Labs. https://www.bayanatlabs.com/open/arabic-document-atlas

• CITATION.cff

No DOI. Cite the URL in the citation above and the date you read it.

@misc{bayanatlabs2026arabicdocumentatlas,
  author       = {{Bayanat Labs}},
  title        = {Arabic Document Atlas, v0.3.0},
  year         = {2026},
  month        = oct,
  howpublished = {\url{https://www.bayanatlabs.com/open/arabic-document-atlas}}
}

How we handle citations

Changelog

Corrections and removal requests

If something in this release should not be there, or is wrong, email hello@bayanatlabs.com. We handle removal requests within 30 days, and platform requests within 24 hours. How removal requests work

Related product: Bayanat Documents