Bayanat-Labs/open-arabic-books

Open Arabic Books

1,545 Arabic books published by the Hindawi Foundation, taken from Andreas Hallberg's The Arabic E-Book Corpus (version 1, CC BY 4.0) and rebuilt into chapters: front matter and licence pages set apart, footnotes split out with their positions, and raw and normalised text. Every row in all three configs has an attribution line and the book's own licence wording; chapters and books also hold the text, and metadata does not.

Config	Rows	What a row is
chapters (default)	27,544	one chapter, preface, appendix or similar section of a book
books	1,545	one whole book: text, footnotes, front matter, licence page
metadata	1,545	one book without its text: credits, dates, categories, licence evidence, source hash

In total: 54,204,272 words of body text (runs of letters and digits after normalisation; footnotes not counted), 27,544 chapters and 102,709 footnotes.

from datasets import load_dataset

chapters = load_dataset("Bayanat-Labs/open-arabic-books", "chapters", split="train")
books = load_dataset("Bayanat-Labs/open-arabic-books", "books", split="train")
metadata = load_dataset("Bayanat-Labs/open-arabic-books", "metadata", split="train")

Arabic Wikisource is published separately as Arabic Wikisource Public-Domain Works.

Licence, book by book

The dataset is released under CC BY 4.0. Each book's colophon was read from the corpus file, on 2026-11-04, and a book is included only when its own statement supports that:

Colophon of the book	Books	original_text_status
Original Arabic text in the public domain; book design CC BY 4.0	1,212	public-domain
Arabic translation and book design CC BY 4.0; original text in the public domain	332	public-domain
Arabic translation and book design CC BY 4.0; original text CC BY 3.0	1	CC-BY-3.0

The full wording is in license_quote, the rest of the colophon in license_page, and the decision in license_decision.

The corpus holds 1,745 books; 200 are not in this release. For 187 translations the colophon does not state the licence of the Arabic translation, for 12 it states a different licence or reserved rights, and for 1 the author's own licence contradicts its public-domain statement. They are listed with their reason and wording in qa/dropped_books.json.

A book counts as a translation when the corpus metadata says so, the title page credits a translator, the colophon says "هذه الترجمة" or "الكتاب الأصلي", or its preface or front matter says it was translated. Those passages were read; qa/rights_readings.json records where the wording turned out not to mean a separate translation.

Where a colophon says the original text is in the public domain and the corpus year is after 1955, the book was read with the author's death year where known: 169 such books are kept, also listed in qa/rights_readings.json.

Public-domain status. "Public domain" above is the publisher's statement on the book. Copyright terms differ by country, and many of these authors died within the last 70 years: Wikidata death years in credits show 356 books whose matched author died after 1955, and many authors have no match. Some works are also recent enough to remain protected where terms run from publication. Check the law of your own jurisdiction before relying on it.

Attribution

Every row has an attribution line naming the title, the authors, translators, editors, compilers and foreword writers credited on the title page, the book's address, its rights and the chain below. Use that line, or this chain, when you reuse the text:

1. Authors, translators, editors, compilers and foreword writers, as each book credits them (credits).
2. Hindawi Foundation, publisher of each edition (book_url).
3. Andreas Hallberg, The Arabic E-Book Corpus, version 1, University of Gothenburg, 2024, https://doi.org/10.5878/7rbh-gy93, CC BY 4.0.
4. Bayanat Labs, who changed the material: split it into chapters, separated front matter, licence pages and footnotes, normalised the text and added identifiers.

This dataset is not affiliated with or endorsed by the Hindawi Foundation, Andreas Hallberg or the University of Gothenburg. Book book_url values are the publisher's addresses as the corpus documents them; on 2026-11-04 the one we requested redirected to safahat.org.

How the books were processed

• Sections. The corpus HTML keeps the file boundaries of the original EPUB. Each file is one row in chapters, with section_type (chapter, preface, dedication, part, volume, appendix, acknowledgements, glossary, references, caption). Part and volume headings become part_title and volume_title on the sections after them; a part or volume page of under 15 words with no notes, tables or pictures gets no row of its own, and its words join that title. A caption row has no heading, under 25 words and an omitted picture or the wording of a picture's caption. Headings the source marks, and a leading line starting الفصل, الباب or القسم, go to chapter_label and chapter_title.
• Front matter and licence page. Cover, title pages and colophon are not in the body. front_matter holds the title-page text; license_page holds the publisher line, ISBN, edition dates and licence wording. Publisher contact lines and cover-designer and reviser credits are left out.
• Footnotes. footnotes holds marker, text, text_norm and offset, the character position of the marker in text_raw (markers are removed from the body). 56 of 102,709 notes have a null offset because their marker could not be placed.
• Text. text_raw keeps the published letters and diacritics. text_norm applies policy bayanat-norm-v1 of arabic-normalize. Poetry lines and table rows are on their own lines; table cells are separated by tabs.
• Not carried. Images and SVG figures, including formulas drawn as SVG, are not in the text. n_images_omitted and n_svg_omitted count them (8,847 and 1,419).
• Names. Authors and translators stay as credited, with titles such as "الدكتور" where the book prints them. credits holds authors, translators, editors, compilers and foreword writers named on the title page. Authors with no title-page credit take the corpus metadata value; when that only repeats the book title, no author is set.
• Wikidata. qid is set only when one person has that exact Arabic label or alias on Wikidata and the dates fit the work, or the English label equals the original author's name, and a match found wrong on reading is removed: 832 books have an author QID and 0 books have a translator QID. Everything else is null, so a null says nothing about the person.
• Contact strings. E-mail addresses and international phone numbers in book text would be replaced with [contact removed]; no released book has any.
• Corpus values. original_author, original_year, original_title and hindawi_edition_date are the corpus's own values, unchanged, typos and odd years included. translation_year is read from the colophon.

Duplicates

Among 27,544 chapters, 7 groups (19 chapters) have identical normalised text and 2 groups (4 chapters) are near-duplicates; among books, 0 exact and 0 near-duplicate groups. Nothing is removed: exact_duplicate_group and near_duplicate_group mark them. Rule: MinHash on word 5-grams of normalised text, confirmed by Jaccard at least 0.8; records with fewer than 50 5-grams are compared only by exact hash.

Checks

• Text. Each released book compared with the corpus plain-text edition: median word recall 0.9995 (lowest 0.9730; the plain text also holds title pages), median precision 1.0000. See qa/text_comparison.json.
• Licence. A separate check re-read all 1,545 source hashes and licence quotations: 0 and 0 mismatches. See qa/licence_recheck.json.
• Reading. Samples from 100 random books were read (credits, chapter titles, the start and end of one chapter, one footnote): chapter boundaries correct in 100. Two defects found on the way were fixed before release. This is a sample check, not a proofread. See qa/reading_check.json.

Columns

Column	Configs	Meaning
book_id, title, authors, translators, is_translation, category_main, n_chapters, n_words	books, metadata (book_id, title also in chapters)	the book; n_words counts body text after normalisation and leaves out footnotes
attribution, license_quote	all	credit line; the colophon's licence wording
original_text_status, front_matter, license_page, near_duplicate_group	books, metadata (the last also in chapters)	status of the original text; title-page text; publisher, ISBN, date and licence lines; duplicate group
chapter_id, chapter_index, source_section, section_type, volume_title, part_title, chapter_label, chapter_title	chapters	the section and its place in the book
text_raw, text_norm, footnotes	chapters, books	text as published; normalised text; notes with positions (chapter_index is added in books)
n_words, n_footnotes, n_images_omitted, n_svg_omitted, n_tables, exact_duplicate_group	chapters	counts for the section; exact duplicate group
hindawi_id, credits, author_qids, translator_qids, original_title, original_author, original_language, categories, hindawi_edition_date, original_year, translation_year, isbn, book_url, n_footnotes	metadata	identifiers, credits with role, role_ar, name, qid, death_year, dates and categories
design_license, translation_covered, license_decision, license_checked_at, rights_statement	metadata	the colophon's reading and the decision
upstream_dataset, source_file, source_file_sha256	metadata	source record and file hash
hallberg_word_count	metadata	the corpus's own count, which also includes footnotes and front matter, so it runs above n_words

Top categories

Categories are the publisher's, as in the corpus; stats/by_category.csv has all of them.

category_main	Books	Words
history	258	15,607,220
literature	209	8,878,901
children.stories	162	1,468,778
detective.fiction	156	3,121,336
biographies	114	4,899,855
novels	114	4,343,446

Names and removal

Book credits name people, some of them living. They appear only as the book credits them. To ask for a name or record to be corrected or removed, write to hello@bayanatlabs.com.

Citation

Please cite the corpus and this dataset:

Hallberg, A. (2024). The Arabic E-Book Corpus (Version 1) [Data set]. University of Gothenburg. https://doi.org/10.5878/7rbh-gy93

Bayanat Labs (2026). Open Arabic Books, version 0.1.0. https://huggingface.co/datasets/Bayanat-Labs/open-arabic-books

Builder code: open-arabic-books, Apache-2.0.

بالعربية

1,545 كتابًا عربيًّا من منشورات مؤسسة هنداوي، مأخوذة من «مدونة الكتب الإلكترونية العربية» لأندرياس هالبرغ (CC BY 4.0)، ومقسمة إلى فصول، مع فصل الصفحات التمهيدية وصفحة الترخيص والحواشي، ونص خام وآخر مطبَّع. وفي كل صف من الإعدادات الثلاثة سطر نسبة ونص الترخيص. قُرئ بيان الترخيص في كل كتاب على حدة، وأُدرج الكتاب فقط إذا كان بيانه يسمح بذلك. للتصحيح أو الحذف: hello@bayanatlabs.com

The Arabic text has not yet been reviewed by a native Arabic writer; the English governs.