Datasets:

islamlab
/

islamic-sciences

license: cc-by-nc-sa-4.0
pretty_name: islamlab — The Islamic Sciences Corpus
language:
  - ar
  - en
task_categories:
  - text-generation
  - text-retrieval
  - question-answering
size_categories:
  - 1M<n<10M
tags:
  - islam
  - arabic
  - classical-arabic
  - fiqh
  - sirah
  - hadith-sciences
  - tasawwuf
  - islamic-history
  - arabic-linguistics
  - retrieval
configs:
  - config_name: works
    default: true
    data_files:
      - split: train
        path: works/train-*.parquet
  - config_name: sira
    data_files:
      - split: train
        path: sira/train-*.parquet
  - config_name: fiqh
    data_files:
      - split: train
        path: fiqh/train-*.parquet
  - config_name: aqidah
    data_files:
      - split: train
        path: aqidah/train-*.parquet
  - config_name: rijal
    data_files:
      - split: train
        path: rijal/train-*.parquet
  - config_name: tasawwuf
    data_files:
      - split: train
        path: tasawwuf/train-*.parquet
  - config_name: tarikh
    data_files:
      - split: train
        path: tarikh/train-*.parquet
  - config_name: lugha
    data_files:
      - split: train
        path: lugha/train-*.parquet
  - config_name: adab
    data_files:
      - split: train
        path: adab/train-*.parquet
  - config_name: mutafarriqat
    data_files:
      - split: train
        path: mutafarriqat/train-*.parquet
  - config_name: passages
    data_files:
      - split: train
        path: passages/train-*.parquet
  - config_name: parallel_en
    data_files:
      - split: train
        path: parallel_en/train-*.parquet

islamlab — The Islamic Sciences Corpus

The Islamic sciences other than Qur'an and hadith, as their authors wrote them: 4,022 works by scholars who died between the 0st and the 14th Hijri century, cut along their own chapter and biographical-entry boundaries into 1,864,389 units (3.41 billion characters of Arabic), each carrying the volume and page it sits on so a quotation can be cited rather than merely produced.

Scope is Ahl al-Sunnah wa'l-Jamāʿah, and the gate is the author, not the title. A work enters only if its author died by 1350 AH, so every text here is in the public domain and is carried verbatim.

islamlab is a volunteer effort. Our aim is to present these works carefully and at high quality, and to help them travel faithfully — in the spirit in which they were written — not to claim them as ours. Who to thank, and what we consulted without ever quoting, is set out in SOURCES.md.

Subsets

config	science	works	units	chars
`sira`	Sīrah	110	37,634	107M
`fiqh`	Fiqh	974	386,382	1457M
`aqidah`	ʿAqīdah	414	23,745	116M
`rijal`	Rijāl	844	824,734	741M
`tasawwuf`	Taṣawwuf	298	25,662	105M
`tarikh`	Tārīkh	299	239,785	360M
`lugha`	Lughah	378	235,028	292M
`adab`	Adab	512	73,342	196M
`mutafarriqat`	Mutafarriqāt	193	18,077	33M
`works`	bibliographic spine	4,022	—	—
`passages`	retrieval chunks	—	3,575,392	—
`parallel_en`	Arabic–English alignments	17	45,972	—

The shelves

sira — Sīrah. The life of the Prophet, his described character, and the Companions around him. Among the largest: سبل الهدى والرشاد في سيرة خير العباد — الصالحي الشامي (d. 942); إمتاع الأسماع بما للنبي من الأحوال والأموال والحفدة والمتاع — المقريزي (d. 845); دلائل النبوة - البيهقي — أبو بكر البيهقي (d. 458); السيرة النبوية من البداية والنهاية - ت عبد الواحد — ابن كثير (d. 774).

fiqh — Fiqh. The law and its theory — the four schools, legal maxims, fatwa collections, judgeship and governance. Among the largest: المبسوط للسرخسي — شمس الأئمة السرخسي (d. 483); حاشية ابن عابدين = رد المحتار - ط الحلبي — ابن عابدين (d. 1252); البحر الرائق شرح كنز الدقائق ومنحة الخالق وتكملة الطوري — زين الدين ابن نجيم (d. 970); حاشية الجمل على شرح المنهج = فتوحات الوهاب بتوضيح شرح منهج الطلاب — الجمل (d. 1204).

aqidah — ʿAqīdah. Creed, the schools and sects, and the refutations written between them. Among the largest: منهاج السنة النبوية — ابن تيمية (d. 728); درء تعارض العقل والنقل — ابن تيمية (d. 728); العواصم والقواصم في الذب عن سنة أبي القاسم — ابن الوزير (d. 840); درء التعارض — ابن تيمية (d. 728).

rijal — Rijāl. The sciences about hadith — narrator criticism, hidden defects, terminology, takhrīj, and the biographical dictionaries they rest on. Among the largest: تاريخ دمشق لابن عساكر — أبو القاسم ابن عساكر (d. 571); سير أعلام النبلاء - ط الرسالة — شمس الدين الذهبي (d. 748); تهذيب الكمال في أسماء الرجال — المزي، جمال الدين (d. 742); الوافي بالوفيات — الصفدي (d. 764).

tasawwuf — Taṣawwuf. Sober spiritual discipline, ethics, heart-softening literature and litanies. Among the largest: اتحاف السادة المتقين بشرح إحياء علوم الدين — الزبيدي، مرتضى (d. 1205); إحياء علوم الدين — الغزالي (d. 505); حسن التنبه لما ورد في التشبه — نجم الدين محمد الغزي (d. 1061); الفتوحات الربانية على الأذكار النواوية — ابن علان (d. 1057).

tarikh — Tārīkh. Chronicles, city histories, geography and travel. Among the largest: تاريخ الإسلام - ت تدمري — شمس الدين الذهبي (d. 748); البداية والنهاية - ت التركي — ابن كثير (d. 774); تاريخ الإسلام - ت بشار — شمس الدين الذهبي (d. 748); تاريخ الإسلام - ط التوفيقية — شمس الدين الذهبي (d. 748).

lugha — Lughah. The Arabic language sciences — grammar, morphology, rhetoric, and the great lexicons. Among the largest: تاج العروس من جواهر القاموس — مرتضى الزبيدي (d. 1205); لسان العرب — ابن منظور (d. 711); تهذيب اللغة — أبو منصور الأزهري (d. 370); تمهيد القواعد بشرح تسهيل الفوائد — المعروف بناظر الجيش (d. 778).

adab — Adab. Letters and poetry, from the pre-Islamic dīwāns to the chancery anthologies. Among the largest: نهاية الأرب في فنون الأدب — النويري، شهاب الدين (d. 733); صبح الأعشى في صناعة الانشاء — القلقشندي، الفزاريو الشافعي (d. 821); خزانة الأدب ولب لباب لسان العرب للبغدادي — عبد القادر البغدادي (d. 1093); نفح الطيب من غصن الاندلس الرطيب ت إحسان عباس — المقري التلمساني (d. 1041).

mutafarriqat — Mutafarriqāt. Works whose author came through the gate but whose science the sources do not settle — shelved as miscellanea rather than guessed at or dropped. Among the largest: kitāb al-Khabar ʿan al-Bashar Fī ’ansāb al-‘Arab wa Nasab Sayyed al-Bashar — المقريزي (d. 845); مجموع رسائل الحافظ ابن رجب الحنبلي — زين الدين أبو الفرج عبد الرحمان بن أحمد بن رجب عبد الرحمان بن الحسن بن محمد بن أبي البركات مسعود البغدادي السلامي الدمشقي (d. 795); تعطير الأنام في تعبير المنام — عبد الغني النابلسي (d. 1143); مجموعة رسائل الإمام الغزالي — الغزالي (d. 505).

How the rows are shaped

works is the spine: one row per work, carrying work_id, our category and subcategory, madhhab where the work declares one, the Arabic and transliterated title and author, the author's death year in both calendars, and the size of the text. source says which library our copy came through and text_quality warns where the digitisation was never proof-read. Everything else joins to it on work_id.
The science configs hold the text. A row is one heading-delimited unit: heading and heading_path place it in the book, kind says whether it is a section or a biographical entry, and page_start / page_end anchor it in the printed edition. markers carries the structural flags the source records about that heading — auto and check say how the heading was found, bio_man / bio_wom / dic_top / dic_lex / dic_nis / dic_bib say what kind of entry it opens, and paratext / editor warn that the block is the modern edition's apparatus rather than the author's text. Filter on it or ignore it; it is never mixed into the Arabic.
passages is the same text re-cut to ~1,200 characters on sentence boundaries for retrieval, each chunk carrying a ready Arabic citation and the unit_id it came from.
parallel_en is the corpus's one multilingual seam: sentence-aligned Arabic–English for seventeen works that have a published translation — among them al-Shāfiʿī's Risāla, Ibn Khaldūn's Muqaddima, Ibn Khallikān's Wafayāt, Ibn Baṭṭūṭa's Riḥla and Ibn Ḥazm's Ṭawq al-Ḥamāma. It is a separate table because it is a different kind of thing: a translation is someone's reading, not the author's words.

The Arabic itself is untouched: continuation lines are rejoined, page stamps and editorial flags are lifted into their own columns, whitespace is collapsed, and no letter or vowel is altered. The one substitution is notational — where the source brackets a Qur'anic quotation with its own ASCII markers, the text carries ﴿ ﴾ instead, which is how Arabic prints it. There is no search-normalised copy; strip the diacritics yourself if your retriever wants them gone.

from datasets import load_dataset

fiqh  = load_dataset("islamlab/islamic-sciences", "fiqh",  split="train")
works = load_dataset("islamlab/islamic-sciences", "works", split="train")

Where this sits

Three organisations, three gates; a work belongs to exactly one of them.

Dataset family	What it holds
`quranlab/*`	Qur'an, translations, tafsīr, qirā'āt, hadith text and its commentary, recitation
`risaleinur/*`	The Risale-i Nur corpus and its derivatives
*`islamlab/`**	Every other Islamic science — this dataset

Four datasets are derived from this one, each re-cutting it for a job the corpus itself does not do well:

Dataset	What it re-cuts it for
`islamlab/arabic-lexicons`	The dictionaries, entry by entry, with the headword and root split out — for looking a word up
`islamlab/islamic-biographies`	Every biographical notice in the ṭabaqāt literature, one row per notice — for prosopography
`islamlab/islamic-scholars`	One row per author instead of per book — the authority file the rest joins to
`islamlab/islamic-sciences-training`	Domain-adaptation text, retrieval pairs with hard negatives, and citation questions with a work-level held-out split

Alongside it, islamlab/persian-islamicate carries the Persian side of the same tradition — chiefly historiography — as a separate dataset, because OpenITI's Persian metadata does not support filing a work under a science the way the Arabic side does.

The line that matters most in practice: hadith text belongs to quranlab/hadith; the sciences built about that text — narrator criticism, ʿilal, muṣṭalaḥ, takhrīj — are the rijal config here.

Notes

Keys are stable: work_id, unit_id ({work_id}:{seq}), passage_id.
subcategory and madhhab are empty where the work does not declare one. An empty cell means we do not know, never that the answer is "none".
Corrections, additions and removal requests are welcome on the Community tab, and are acted on promptly.

License

The works are public domain; the machine-readable transcriptions they reached us through are licensed CC BY-NC-SA 4.0, so this dataset carries the same terms — attribute, keep it non-commercial, share alike. The details are in LICENSES.md and the credits in SOURCES.md.

Citation

@misc{islamlab_islamic_sciences,
  title  = {islamlab — The Islamic Sciences Corpus},
  author = {islamlab contributors},
  year   = {2026},
  howpublished = {Hugging Face Datasets}
}