quranlab/quran-ijaz

QuranLab — Iʿjāz, Letter Structure and Abjad Reckoning

QuranLab 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.

The text these numbers are counted from reaches you through the work of the Tanzil Project. Who to thank, and what we consulted without ever quoting, is set out in SOURCES.md.

What this is

The Qurʾānic sciences have always counted. Scribes counted āyāt to settle where a juzʾ begins, reciters counted letters to fix a qirāʾa, and scholars from al-Bāqillānī onward argued about what in the text cannot be imitated. This dataset carries the counting part of that inheritance, done carefully and in the open, over the same 6,236-āya backbone as every other QuranLab corpus.

It publishes arithmetic, not conclusions. Every row here was computed from the published text and can be recomputed by you. What a number means — whether it is remarkable, whether it points to anything — is a separate question this dataset does not answer on your behalf.

The one thing worth knowing before you use it

There is no such thing as "the number of letters in the Qurʾān."

Not because the question is improper, but because it is underspecified. Three choices have to be fixed before any count exists, and each is a real scholarly tradition rather than a matter of taste:

Choice	The options	Why it moves the number
Orthography	Uthmānī rasm · imlāʾī	The rasm writes ٱ and the superscript ٰ where modern spelling writes a plain ا
Ordering	mashriqī (eastern) · maghribī (western)	ṣād is 90 in one and 60 in the other; shīn is 300 against 1000
Convention	five documented ones	Whether ة counts as hāʾ (5) or tāʾ (400), whether ى is yāʾ (10) or alif (1)

Counted honestly, the whole Qurʾān has between 325,386 and 337,326 letters in the Uthmānī rasm and exactly 330,709 in imlāʾī spelling — the rasm figure moving with the convention you apply, the imlāʾī one steady because modern orthography has no superscript letters to make a ruling about. Every one of those totals is correct. None of them is the total.

The conventions config puts all ten side by side in ten rows. If you have ever met a claim of the form "the Qurʾān contains exactly N of letter X," that table is the first thing to check it against — and it is why we built this before building anything that interprets.

Configurations

Config	Rows	What it answers
verse-abjad	6,236	Abjad totals and letter counts per āya, in both orderings and both scripts.
verse-letters-uthmani	6,236	Per-āya count of each of the 28 letters in the Uthmānī rasm.
verse-letters-imlaei	6,236	Per-āya count of each of the 28 letters in the imlāʾī orthography.
lafz-al-jalala	6,236	Per-āya count of ٱللَّه under each counting rule, and what a text search misses.
surah-structure	114	Per-sūra āya/word/letter/abjad totals, revelation order, page and juzʾ span.
page-structure	604	Per-page totals across the 604-page muṣḥaf, including the lafẓ al-jalāla.
divine-names	9	Corpus-wide count of each divine name, by QAC lemma and by written form.
muqattaat	156	Each disconnected opening letter against its frequency in its own sūra.
letter-frequency	56	Corpus-wide frequency and rank of each letter, in both scripts.
conventions	10	What the whole-Qurʾān totals become under each documented convention.
verifiable-qa	5,176	Questions whose answers can be rechecked against the rows they came from.
chapters	1,765	A table of contents across all nine works, with the subject each title announces.
verse-coverage	6,236	Per āya: how many of the classical works quote it, and which.
citations	19,275	Every Qurʾānic quotation in the classical works, linked to its āyāt.
work-baqillani-ijaz-ar	2,332	Iʿjāz al-Qurʾān — Abū Bakr al-Bāqillānī (d. 403 AH), by paragraph.
work-khattabi-bayan-ar	236	Bayān iʿjāz al-Qurʾān — Abū Sulaymān al-Khaṭṭābī (d. 388 AH), by paragraph.
work-jurjani-dalail-ar	1,924	Dalāʾil al-iʿjāz — ʿAbd al-Qāhir al-Jurjānī (d. 471 AH), by paragraph.
work-jurjani-asrar-ar	1,673	Asrār al-balāgha — ʿAbd al-Qāhir al-Jurjānī (d. 471 AH), by paragraph.
work-zarkashi-burhan-ar	10,395	al-Burhān fī ʿulūm al-Qurʾān — Badr al-Dīn al-Zarkashī (d. 794 AH), by paragraph.
work-suyuti-itqan-ar	10,097	al-Itqān fī ʿulūm al-Qurʾān — Jalāl al-Dīn al-Suyūṭī (d. 911 AH), by paragraph.
work-ibntaymiyya-muqaddima-ar	174	Muqaddima fī uṣūl al-tafsīr — Ibn Taymiyya (d. 728 AH), by paragraph.
work-ibnkhaldun-muqaddima-ar	686	al-Muqaddima — Ibn Khaldūn (d. 808 AH), by paragraph.
work-baqillani-tamhid-ar	2,384	Tamhīd al-awāʾil wa-talkhīṣ al-dalāʾil — Abū Bakr al-Bāqillānī (d. 403 AH), by paragraph.

Reading the muqaṭṭaʿāt table

Twenty-nine sūras open with disconnected letters — الم, يس, ق, ن and the rest. A classical observation, recorded by al-Zarkashī among others, is that these letters run unusually frequent inside the very sūra they open.

The muqattaat config lets you check that rather than repeat it. Each row gives one opening letter's share of its own sūra against its share of the whole Qurʾān, as frequency_ratio. Above 1.0 means over-represented. Some openings bear the observation out strikingly; others do not. Both kinds of row are published, because a table that only showed the confirming cases would not be a table, it would be an argument.

Counting ٱللَّه

The same lesson arrives a second time, and more sharply, in the one quantity the numerical literature leans on hardest.

Counted by lemma — the way the Quranic Arabic Corpus segments the text — the lafẓ al-jalāla occurs 2,699 times. That is the figure the tradition carries. Counted by searching the text for الله, it occurs 2,555 times. The difference is not rounding:

• 151 occurrences are missed. In 149 of them the alif elides after a prefixed lām and the lafẓ is written له — لِلَّهِ. The remaining two are at 10:59 and 27:59, where an interrogative hamza makes it آلله.
• 7 are added that are not the name at all. Five are ٱللَّهُمَّ, the vocative. The other two are ٱللَّهْو — the diversion, at 62:11 — and ٱللَّهَب — the flame, at 77:31. Ordinary nouns that happen to begin with the same four letters.

The two errors are counted in separate columns, never netted. At 10:10 a real lafẓ written له sits beside ٱللَّهُمَّ: one miss and one false positive, in the same āya, cancelling to a difference of zero. Reporting only the difference would show that verse as correctly counted when both of its figures are wrong.

Every divine name in the divine-names config behaves this way, and the table gives both figures with the gap in a column of its own. رب is the extreme case: 977 by lemma against 1,241 by search, because رب is a substring of much else.

This is why the counting method is a published column here and not an implementation detail. A numerical argument assembled with a text search is wrong in both directions before it begins.

Tevafuk, and what this dataset can and cannot say about it

Tevafuk — the observed correspondence in how occurrences of ٱللَّه fall on the page — is associated above all with Said Nursi, who wrote about it at length and had a muṣḥaf typeset to display it. Naming a claim and its author is what this dataset does; it is not the same as reproducing his text, which is a separate question governed by its own permissions.

What we publish is the substrate, computed and checkable: the count of the lafẓ per āya (lafz-al-jalala) and per page (page-structure), across all 604 pages.

What we do not publish is a verdict on the alignment, and the reason is concrete rather than diplomatic. Tevafuk is a property of a particular typesetting — the 15-line tevafuklu muṣḥaf. Our backbone paginates by the 604-page Madinah printing. Counts transfer between the two; positions on the page do not. Testing the alignment needs the line-level layout of the muṣḥaf the claim is actually made about, and until that is sourced, anything we said about it would be about the wrong book.

So: the counts are here and they are ours to stand behind. The alignment is named, attributed, and left open.

Why letter_delta is on every āya

verse-abjad carries the per-āya difference between the imlāʾī and Uthmānī letter counts. For most āyāt it is small. Where it is large, the āya is one whose rasm spelling diverges sharply from modern orthography — and it is exactly those āyāt that make or break a numerical claim, depending on which text its author happened to use. Publishing the gap per row means nobody has to take our word for how big it is.

For agents and pipelines

Deterministic facts an automated consumer needs, so nothing has to be inferred from a sample.

• Config naming: <grain>-<subject>; verse-grain configs carry the 6,236-āya backbone
• Join key: verse_key (surah:ayah), identical to every other QuranLab corpus
• Default config: verse-abjad
• Format: Parquet, one directory per config, split train.
• Terms and credit are per config in metadata/, never per row — join on the config name when you need them.
• • Every row in this repo is claim_status = computed: arithmetic over the published text, not a report of what anyone said about it.
• • A number is only defined once you fix an orthography and an ordering and a convention. conventions is the table that shows why.

Inspect it without downloading the data:

# every config and split
curl https://datasets-server.huggingface.co/splits?dataset=quranlab/quran-ijaz

# a first page of rows
curl "https://datasets-server.huggingface.co/first-rows?dataset=quranlab/quran-ijaz&config=verse-abjad&split=train"

# machine-readable schema and provenance (Croissant JSON-LD)
curl https://huggingface.co/api/datasets/quranlab/quran-ijaz/croissant

Query it in place with SQL — no full download:

hf datasets sql quranlab/quran-ijaz "SELECT * FROM 'verse-abjad' LIMIT 5"

import duckdb
duckdb.sql("SELECT count(*) FROM 'hf://datasets/quranlab/quran-ijaz/verse-abjad/*.parquet'")

Quality, and how to check it yourself

Nothing is released until the integrity gate passes. Every claim below is something you can re-run rather than take on trust — here on the published data, not in our build tree:

What is guaranteed	Evidence
6,236 āyas, 114 sūras, on the shared verse_key backbone	verse-abjad row count and a join against quranlab/quran
Abjad tables reproduce the classical anchors	الله = 66, محمد = 92, the basmala = 786 (mashriqī, standard)
Both orderings cover the same 28 letters	metadata/abjad_values.csv; the six that diverge are marked
No letter count is presented as the letter count	conventions publishes all ten totals side by side
The lafẓ al-jalāla reproduces the classical 2,699	lafz-al-jalala, counted by QAC lemma; the text-search figure is published beside it, not instead of it

import pandas as pd
c = pd.read_parquet("hf://datasets/quranlab/quran-ijaz/conventions/train-00000-of-00001.parquet")
print(c[["convention", "script", "n_letters", "abjad_mashriqi"]])   # ten different right answers

Where a source is genuinely missing, the row is published empty and labelled rather than filled by guessing from a neighbouring entry. Coverage numbers in this card count only real content.

The works themselves

The counting above is ours. This part is not: it is the Sunni literature on what is inimitable about the Qurʾān, carried paragraph by paragraph with each author's name on it.

Work	Author	Died	Passages
Iʿjāz al-Qurʾān	Abū Bakr al-Bāqillānī	403 AH	2,332
Bayān iʿjāz al-Qurʾān	Abū Sulaymān al-Khaṭṭābī	388 AH	236
Dalāʾil al-iʿjāz	ʿAbd al-Qāhir al-Jurjānī	471 AH	1,924
Asrār al-balāgha	ʿAbd al-Qāhir al-Jurjānī	471 AH	1,673
al-Burhān fī ʿulūm al-Qurʾān	Badr al-Dīn al-Zarkashī	794 AH	10,395
al-Itqān fī ʿulūm al-Qurʾān	Jalāl al-Dīn al-Suyūṭī	911 AH	10,097
Muqaddima fī uṣūl al-tafsīr	Ibn Taymiyya	728 AH	174
al-Muqaddima	Ibn Khaldūn	808 AH	686
Tamhīd al-awāʾil wa-talkhīṣ al-dalāʾil	Abū Bakr al-Bāqillānī	403 AH	2,384

Every one passes the same theology gate as the rest of QuranLab, and the classification that let it in is recorded per work in metadata/works.csv rather than assumed. They are also in conversation with each other: al-Suyūṭī's chapter on iʿjāz opens by naming who wrote monographs on it — al-Khaṭṭābī, al-Rummānī, al-Rāzī, and al-Qāḍī Abū Bakr al-Bāqillānī — and three of those four are in the table above.

The fourth, and why he is not here

al-Rummānī's al-Nukat fī iʿjāz al-Qurʾān is one of the three founding treatises of this literature and has been quoted with approval by Sunni authors ever since. It is available, it is short, and it is not in this dataset: Abū al-Ḥasan al-Rummānī was a Muʿtazilī theologian of the Baghdad school, and the corpus admits Ahl al-Sunnah works without exception.

Applying the rule to a book we wanted is the only thing that makes it a rule. The exclusion is recorded, with its reason, in metadata/excluded_works.csv rather than left as a silent absence.

Markers kept inside the text

The digitisations carry a few marks inside a paragraph that look like noise and are not. They are kept, and metadata/markup.csv says what each one means:

• % separates the two hemistichs of a line of verse. al-Bāqillānī quotes a great deal of poetry, because setting the finest Arabic verse beside the Qurʾān is his argument — collapsing the hemistichs would flatten it.
• / < … > / marks a variant Qurʾānic reading, where a work is discussing the seven aḥruf rather than quoting the Ḥafṣ text.
• & … & encloses a heading or technical term set off inside the text, and ( … ) the printed edition's own brackets.

What was stripped is genuine markup: page and milestone markers, and — caught by the integrity gate rather than by eye — eleven al-Zarkashī passages carrying an unclosed @ 263 page remnant that would have published as though he had written it.

Finding your way in

Nine works come to thirty thousand paragraphs, and without an index the abjad discussion inside al-Zarkashī is reachable only by grep. chapters is a table of contents across all of them — 1,765 chapters, each with the title its own author gave it, its size, and its page in the printed edition.

117 of those titles announce a subject, and those carry topic keys: ijaz, balagha, qiraat, muqattaat, abjad-huruf, naskh and the rest. The classification is ours rather than the authors', so it is published with its rules — metadata/topics.csv lists the exact Arabic each key matched on, and you can disagree with it. Matching is against titles only, never body text: a title is the author saying what the chapter is about, while a word in a paragraph means only that the word occurs there.

The other chapters are titled faṣl and nothing more, and they get no topic rather than a guessed one.

verse-coverage runs the other way, from the muṣḥaf inward: 3,848 of the 6,236 āyāt are quoted by at least one of these works. The 2,388 that are not are published as rows of zeroes, because an absence is a fact about the literature and dropping the row would hide it.

Linking them to the muṣḥaf

OpenITI tags Qurʾānic quotations in these texts explicitly, which is far better than guessing at where a scholar stopped quoting and resumed arguing. There are 19,275 of them, and the citations config carries every one, matched against our own text and linked to the āyāt it covers.

• 14,764 resolve to a single place — 77% — and 890 of those span more than one āya, because scholars quote passages rather than verses. حم تنزيل الكتاب is 40:1–2 and was a failure until the matching worked in ranges.
• 1,943 occur in several places and are published with their match count and no verse key.
• 88 were searched and not found, and 2,480 are too short to search. They stay unresolved. A quotation from memory, an elision, or a reading other than Ḥafṣ is a real thing to record; attaching it to the nearest plausible āya would be inventing an attribution.

match_status says which of those a row is, so nothing rests on a link that was never made. As a sanity check on the whole exercise, the most-cited āya across all four works is 3:7 — the verse on the clear and the ambiguous, which is where any discussion of interpreting the Qurʾān begins.

Questions you can mark

Qurʾānic numerology is the subject a language model is most likely to be confidently wrong about. It is also, unusually, a subject where being right is decidable — so verifiable-qa carries 5,176 questions whose answers are checkable by recomputation rather than by judgement.

Every answer names the row it came from, in source_config and source_key, so it can be rechecked against this repository without trusting whatever produced it. The build gate re-derives all 5,171 answerable ones from their own sources and refuses to publish on a single mismatch.

The remaining handful are the reason the rest of the dataset exists. Asked how many letters are in the Qurʾān, the right answer is not a number — it is "which orthography, which convention". Those items carry answer_type = clarification-required. A question set made only of the answerable ones would train exactly the overconfidence this repository was built to correct.

Scope, and what is not here

Two layers are complete: the computed one, arithmetic anyone can re-run, and the attributed one, somebody else's sentence with their name on it. Three things are deliberately absent, and it is more useful to say why than to leave them looking forthcoming.

• Tevafuk's alignment. Checked and blocked. It is a property of the 15-line tevafuklu muṣḥaf; the twelve layouts published by QUL include several 15-line typesettings, none of them that book. A different 15-line muṣḥaf would answer a question about the wrong object. The counts are here; the alignment waits for the layout.
• The abjad tradition in practice — chronograms, inscriptions, manuscript dating. Its criticism is here and now findable: chapters indexes Ibn Khaldūn's fifty-third chapter, on apocalyptic prediction and on what the word jafr actually refers to, under abjad-huruf. He traces letter-divination to its origin and takes it apart — the classical objection arriving from inside the tradition rather than from outside it. The practice would need a corpus of dated chronograms we do not have on terms that allow republication.
• Scientific iʿjāz. Not excluded on principle — every claim in it would need a named proponent and a citable text, and no corpus meeting that standard is available to us. Publishing the claims without the names is the one thing this repository does not do.

Nothing here will ever assert a number, or what a number means, without saying whose assertion it is.

Terms

Terms differ by config and each one states its own in metadata/configs.csv. In short: our computation is CC BY 4.0 over Tanzil's text (CC BY 3.0); the classical works come from the OpenITI release under CC BY-NC-SA 4.0, so reuse of those configs must stay non-commercial and keep the same licence. Full detail, and how to ask for a correction or a removal, is in LICENSES.md and SOURCES.md.

Citation

@misc{quranlab_quran_ijaz_2026,
  title        = {QuranLab — Qur'anic Iʿjāz, Letter Structure and Abjad Reckoning},
  author       = {QuranLab},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/quranlab/quran-ijaz}}
}

---

QuranLab is volunteer work, and help is genuinely welcome — an error found, a convention we have documented poorly, a classical source we should be counting against. The Community tab is the place.