Bayanat-Labs/arabic-book-quote-retrieval

Arabic Book Quote Retrieval

9,855 exact OCR quotes paired with their source pages across 675 historical books. This dataset tests the narrow task of finding a page containing a supplied excerpt. It is derived from the immutable Arabic Books Page Text Corpus v0.1.0, itself based on Arabic Document Atlas v0.3.0.

Queries are automatically extracted contiguous 12-token substrings of uncorrected machine OCR. They are not natural user questions, paraphrases, human-verified transcriptions or semantic relevance judgments. This is an intentionally easy literal retrieval fixture: exact substring matching is a strong baseline. One source-page association is provided per query; it is not an exhaustive list of relevant pages. Similar passages, editions and unjudged literal occurrences may exist, including outside the split. No negative is labeled as semantically irrelevant. Pages are historical text; OCR errors and their possible effects are retained.

Data and splits

Split	Pages	Queries	Source books
train	7,912	7,912	541
validation	1,012	1,012	71
test	931	931	63
Total	9,855	9,855	675

The pages, queries, and associations configurations each have these splits. A page_id identifies the candidate page. queries contains query_id and query; associations maps query_id to its associated source_page_id and exact start/end Unicode code point offsets in page_text (end exclusive). Keep the associations separate from model inputs. The query and page texts are exactly as produced by the upstream OCR; no text correction or claim of OCR accuracy is implied.

Documents are assigned by a salted SHA-256 hash of source document ID. All pages from one source document stay in one split. Among selected records, exact NFC/whitespace-normalized duplicate pages and queries are discarded globally before release. This is not work- or edition-level deduplication: related or near-duplicate editions may overlap, and public books may already appear in model training data. The full upstream corpus is not an exhaustive held-out candidate set. Each split is meant to be searched against its own published pages.

Each source book contributes at most 15 selected pages. Candidates are chosen by a deterministic hash of page IDs, restricted to pages passing the upstream mechanical screen and at least 300 characters, with usable 12-token contiguous excerpts. The screening is a mechanical filtering rule, not evidence that OCR or language is correct. Original text, title, provider, page index, PDF and manifest identities, rights basis, and Atlas hashes are preserved in the page rows. Query IDs can be joined to these fields through associations.

Load and evaluate

from datasets import load_dataset
repo = "Bayanat-Labs/arabic-book-quote-retrieval"
queries = load_dataset(repo, "queries", split="test", revision="v0.1.0")
pages = load_dataset(repo, "pages", split="test", revision="v0.1.0")
labels = load_dataset(repo, "associations", split="test", revision="v0.1.0")
# Rank page_id values using query and page_text; use labels only for scoring.

A simple untrained literal-substring scan, evaluated on the first 100 released queries in each split, returned the associated page first on 100/100 per split; none of those 300 queries had multiple literal matches in its split. This is a small deterministic sanity check, not a representative difficulty estimate, model baseline, or proof of unique answers throughout the dataset. Report association recall as such, and do not treat unlabeled pages as verified negatives.

Source and rights

Arabic Collections Online (ACO) and contributing institutions state that they believe the displayed materials are public domain; the upstream corpus selected editions with unambiguous publication year at or before 1930 and no contrary retrieved rights evidence. This is not independent worldwide copyright clearance. Underlying works retain their own rights. The exact source of each rights statement is the row's source_manifest_url; source_rights_url links to the ACO information page. Credit ACO, the row's named institution, and Bayanat; retain source links and indicate changes. Bayanat's OCR representation, extraction metadata and documentation are CC BY 4.0; this does not replace underlying work rights. See DATA_LICENSE.md and ATTRIBUTION.md.

The source release's 57 page Parquets were checked against its reviewed SHA-256 manifest before extraction. Page fields were compared back to the source, quote offsets and splits audited, and normalization duplicates removed. See METHODS.md and source/README.md. No page images are included.

@dataset{bayanat_arabic_book_quote_retrieval_2026,
  title = {Arabic Book Quote Retrieval},
  author = {{Bayanat Labs}},
  year = {2026},
  version = {v0.1.0},
  url = {https://huggingface.co/datasets/Bayanat-Labs/arabic-book-quote-retrieval}
}

From Bayanat Labs

This dataset is published by Bayanat Labs (بيانات لابز), a Riyadh-based Arabic AI data company. For licensed or custom Arabic data beyond this open release, see Bayanat Search.