Datasets:
The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
Clouda OCR β Real Arabic Printed Book Pages (qualified subset)
This dataset contains only real scanned pages of Arabic printed books with
their paired ground-truth text, published continuously in batches. Every
published batch passed the acceptance pipeline described below. Pages that did
not (yet) qualify are not included here; a separate permanent exclusion
registry (EXCLUSION_REGISTRY.jsonl, aggregated stats only) records each
excluded/deferred page with its precise reason.
Status: acceptance pipeline active. Batches are published only after all acceptance conditions pass. Nothing is published on pairing success or QA checks alone.
Sources
| Source | Publisher / URL | License | Rights gate |
|---|---|---|---|
| freococo/ocr_arabic_books | https://huggingface.co/datasets/freococo/ocr_arabic_books | CC-BY-NC-ND-4.0 (declared by publisher) | RIGHTS_PENDING β redistribution permission under user review |
Per-page metadata keeps: source id, book config, volume/shard, original file name, page index, publisher license string, batch id, SHA-256 of image and label, and the QA/audit trail.
Acceptance conditions (all must hold for publication)
- Real scanned printed-book page (no synthetic renderings; manuscripts and line-crops are excluded by source selection and per-page type).
- Valid image (decodable, sane dimensions) with a correctly linked text (id-level pairing verified; file hashes bound in the tracking DB).
- No major alignment errors: publisher ground truth is a structured page transcription (header / body / footnotes kept verbatim, non-normalized).
- No unacceptable duplicates: exact-hash and perceptual (dhash + text Jaccard β₯0.9) duplicate screening; duplicates excluded, first instance kept.
- Documented text accuracy β₯85% (priority β₯90%): measured by a stratified
visual audit (character-level CER against an independent transcription of
the sampled page). Pairing success, QA checks, or OCR similarity are not
accepted as evidence. Pages without proven accuracy are
UNVERIFIEDand are not published. - No protected-benchmark overlap: pages carry
benchmark_overlap=unknownuntil the protected MISRAJ/KITAB-R hashes are provided; unresolved overlap blocks publication.
Rights and licenses
Downloads being publicly accessible does not imply redistribution rights.
Each batch carries a rights status; batches awaiting the publisher-permission
review are held in RIGHTS_PENDING and are not uploaded. The license tag above
reflects the source license; the final release license is set after the rights
review.
Verification methodology
- Every page image and label is SHA-256 hashed at ingestion and re-verified before packaging (sampling + full-batch integrity checks).
- Duplicates: exact (image/text SHA-256) + perceptual (64-bit dhash, Hamming β€6 and token-Jaccard β₯0.9 text gate).
- Accuracy audit: stratified per book/volume; the auditor transcribes the page
image independently and CER is computed with a documented normalization
recipe (raw code points; no silent NFC normalization). Auditor identity and
limitations are stated in each batch's
AUDIT_REPORT.json. Reviewer: GLM coding-agent visual audit (documented, evidence crops retained) β an AI-agent audit, disclosed as such. - Remote verification: after upload, every file is re-listed from the Hub and its size/hash checked before any local data is cleaned.
Limitations (stated plainly)
- Ground truth of the source dataset derives from publisher-side alignment of scanned pages with e-text; our audit samples measure agreement, not an absolute gold standard.
- Audit coverage is stratified sampling per book/volume, not per-page measurement; per-page accuracy beyond audited pages is not claimed.
- Benchmark-overlap screening is pending the protected hash list.
- Downloads last month
- 65