Bayanat-Labs/arabic-document-atlas

Bayanat Arabic Document Atlas

200,286 unique document images · 60,699,167 machine word records · 26,629 attributed community-text records.

A visual Arabic document resource with JPEG page bytes, machine OCR, approximate word localization and line order, exact source PDF/manifest hashes, edition metadata, duplicate lineage and mechanical quality flags. The main scan tier comes from Arabic Collections Online and its supplying institutions. A separate, smaller PDF image/reference tier preserves Wikisource transcriptions and file-specific permission evidence. Bayanat builds the normalized visual representation and auditable processing pipeline.

Explore actual page examples · Source and pipeline · Research and release record · Usage · Rights and attribution

Snapshot v0.3.0

Component	Records	What a row means
pages	199,997	One retained unique decoded raster, with image bytes and machine OCR
community_text	26,629	A separate attributed Wikisource revision/transclusion body; no image bytes
reference_pairs	289	A source-associated PDF page render plus community body transcription; no machine OCR or word boxes
Machine word annotations	60,699,167	Approximate word guides inside the page rows, not extra pages

The main image shards occupy 159.92 GiB across 679 source books. Reference-pair Parquet occupies 78.42 MiB across 2 original PDFs; permitted originals and credit evidence are included separately. Exact decoded-pixel duplicate occurrences removed within the main tier: 3. Cross-component identical rasters: 0; these are counted once in the 200,286-image union. Mechanical screen passed: 188,473 main-tier pages. That screen uses text amount/script and encoding checks; it is not measured OCR accuracy. Covers and sparse pages may remain in the full image tier with flags. The snapshot has no synthetic images and no human-verified bounding boxes.

Counts describe the pinned snapshot, not the full upstream catalog. Source canvases, variants, community-text records and word records are not added to the unique-image total. See release summary, coverage and failure analysis, integrity report, shard manifest and SHA256SUMS. Pin revision="v0.3.0" for reproducibility; development on main may continue.

Load the image tier

from datasets import load_dataset

pages = load_dataset("Bayanat-Labs/arabic-document-atlas", "pages",
                     split="train", streaming=True, revision="v0.3.0")
page = next(iter(pages))
image = page["image"]
print(page["page_text"])
print(page["source_pdf_sha256"], page["source_pdf_page_index"])

English is the default for documentation and interfaces. Source Arabic is preserved. Streaming avoids downloading the complete image collection.

Useful work this supports

Authentic Arabic document-image pretraining, noisy-supervision OCR research, historical print analysis, visual document retrieval and citation-preserving region-highlighting applications. Related open ACO OCR collections already exist; this release contributes page-image bytes, a spatial representation and quality/provenance controls rather than claiming the first ACO transcription.

Annotation limits

Apple Vision produces the OCR. Its word boxes are approximate guides, not pixel-accurate crop ground truth. Word scores repeat the parent line's uncalibrated confidence and are not word correctness probabilities. Reading order is a machine hypothesis and can fail on columns, poetry, footnotes and tables. No manual table-cell or semantic-layout labels are supplied. OCR errors remain. Native OCR reproduction requires a compatible macOS/Apple Vision system; the pipeline source is open, the system OCR model is proprietary.

bbox=[left,top,right,bottom] uses top-left-origin published-image pixels with half-open bounds. JPEG long edge is at most 2400 pixels; this does not certify scanner DPI or native source resolution. The frozen PDF hash plus zero-based page index identify the processed input. Manifest-to-PDF correspondence follows publisher order with equal page counts; individual canvas/PDF pairs have not all been independently verified.

The separate Wikisource text tier preserves source proofread/validated labels, exact revision and history links, raw body wikitext, rendered text and file-rights metadata. Bayanat did not independently proofread it. The body normally excludes non-transcluded headers/footers, and some current scan versions postdate the text revision. It is not error-free whole-image evaluation gold.

reference_pairs adds original-PDF image bytes only for a narrower cohort with source quality 3/4, explicit compatible file licensing, copyright-holder permission evidence and a binary version that does not postdate the text revision. Those PDF renders may be scanned or born digital. Page association follows the source Page number; it is not independently visually certified. Transcriptions cover the transcluded body and omit some headers/footers. Source proofreading status is preserved, not upgraded to Bayanat gold. These pairs overlap text already in community_text; do not count them as extra transcriptions. See reference provenance and usage.

train is a loading split. There is no claimed held-out benchmark. Related editions and cross-source mirrored scans may overlap; use document/work-level leakage audits for your own evaluation.

Source selection and rights

The main image tier uses ACO's institutional public-domain assertion and a conservative pre-1931 edition filter with no contradictory retrieved rights evidence. This is not independent worldwide legal clearance. Source institutions, exact statements and source links remain in the rows. Underlying scans/works retain their existing rights. Reference images and included originals retain their individual CC BY-SA 4.0 file terms and credits. Bayanat's original annotation representation/docs are CC BY 4.0; adapted Wikisource community text is CC BY-SA 4.0, with attribution/share-alike requirements. Code is Apache-2.0. There is no blanket CC0 license. See component-specific terms and attribution.

Citation

@dataset{bayanat_arabic_document_atlas_2026,
  title = {Bayanat Arabic Document Atlas},
  author = {{Bayanat Labs}},
  year = {2026},
  version = {v0.3.0},
  url = {https://huggingface.co/datasets/Bayanat-Labs/arabic-document-atlas}
}

Also credit the original institution and, for community text, the row's Wikisource revision/history. Corrections and rights concerns: hello@bayanatlabs.com, including page ID and source URL.