Bayanat-Labs/arabic-data-index

Arabic Data Index

Dated statistics on Arabic AI datasets on Hugging Face and in the Masader catalogue: how many there are once copies are counted once, what licences they declare, and which modalities, Arabic varieties, tasks and creation months they cover. By Bayanat Labs. Snapshot 2026-11, read on 1 November 2026 (UTC).

As of 1 November 2026, 3,564 de-duplicated Arabic AI datasets are listed on Hugging Face or in Masader. 46.9% declare a licence in the open-commercial class and 12.7% declare no licence. Open-noncommercial: 9.3%; restricted: 7.4%; other, custom or unknown terms: 23.7%. The classes describe declared licences only and are never added together.

See the charts in the browser: Bayanat-Labs/arabic-data-index (English and Arabic, works offline). Code: github.com/Bayanat-Labs/arabic-data-index.

Snapshot 2026-11 in numbers

Each figure below is a row of snapshots/2026-11/arabic-data-index_2026-11.csv, with its metric id, numerator, denominator and coverage.

Figure	Value	Metric
De-duplicated Arabic datasets (headline datasets)	3,564	ADI-CNT-01
Hugging Face dataset repositories with an Arabic language tag (copies and datasets with 10 or more languages included)	4,333	ADI-CNT-02
Datasets that declare 10 or more languages, counted apart	1,324 (27.1% of 4,888)	ADI-CNT-03
New on Hugging Face in October 2026 (still on the Hub)	404	ADI-GRW-02
Created on Hugging Face in 2026, up to the snapshot	1,626	ADI-GRW-03
Listed on Hugging Face only / in both / in Masader only	2,589 / 213 / 762	ADI-CNT-04
Declare a variety or dialect beyond "Arabic"	30.7%	ADI-DIA-02
Declare at least one regional dialect	10.1% (361)	ADI-DIA-03
Speech and audio / text / image datasets	17.2% / 70.2% / 6.7%	ADI-MOD-01
Speech and audio datasets with an open-commercial declared licence	25.6%	ADI-LIC-04
Declare a software licence (MIT, Apache, BSD, GPL and similar) for their data	25.1%	ADI-LIC-03
Hugging Face datasets without a licence tag on the Hub	16.9% of 2,802	ADI-LIC-05
Hugging Face datasets that are gated	9.1%	ADI-HLT-01
Rows in Hugging Face datasets with a complete viewer size	15.23 billion (coverage 71.7%)	ADI-SIZE-01
Open, commercially usable Arabic speech, released audio hours (quoted, as of 30 October 2026)	36,221	ADI-SIZE-04
Share of Hugging Face datasets held by the 10 accounts with the most	14.0%	ADI-CON-01
Share of 30-day downloads that goes to the 20 most downloaded	61.4%	ADI-CON-02

Declared licence class of the 3,564 headline datasets (ADI-LIC-01):

Class	Datasets	Share
Open, commercial use allowed (permissive, share-alike or software licences)	1,673	46.9%
Open, non-commercial only (CC BY-NC variants)	330	9.3%
Restricted (no derivatives, research or agreement-based)	264	7.4%
Other, custom or unknown terms	844	23.7%
No licence declared	453	12.7%

Licence classes describe the licence each dataset declares (the Hub licence tag, or Masader's licence when the Hub has none). They were not checked against the data or its sources and are not legal advice.

What is counted, and what is not

• Counted: every dataset repository on Hugging Face with an Arabic language tag (32 language codes: ar, ara, the ISO 639-3 Arabic varieties that carry a Hub tag, and Hassaniya), and every entry of Masader at commit ce2c342 (1,163 entries).
• Counted once: repositories in one copy group of the Arabic Dataset Rights Index, and same-size copies on Hugging Face checked by Bayanat Labs (the same complete dataset-viewer size and matching repository names, or a group of 3 or more repositories with one size); 224 repositories are folded into another. Masader entries are joined to a Hugging Face repository through their Hub link.
• Counted apart: datasets that declare 10 or more languages (1,324). Headline figures cover datasets in Arabic only (45.2% of all de-duplicated datasets), Arabic and one other language, or 3 to 9 languages. Masader entries marked multilingual (181) are counted as multilingual because Masader gives no number of languages; entries known to cover 10 or more languages (17) are counted apart.
• Not covered in this snapshot: Kaggle, and datasets published elsewhere that are neither on Hugging Face with an Arabic tag nor in Masader.
• Declared only: licence, modality, dialect and task come from what each repository or Masader entry declares. There is no verified licence figure in this snapshot.

How to use

from datasets import load_dataset
agg = load_dataset("Bayanat-Labs/arabic-data-index", "aggregates", split="train")
ent = load_dataset("Bayanat-Labs/arabic-data-index", "entities", split="train")

Or read the CSV directly: snapshots/2026-11/arabic-data-index_2026-11.csv (one row per metric and dimension value) and snapshots/2026-11/entities_2026-11.csv.gz (one row per dataset). The headline figures are also in latest.json.

To cite one number, give its metric id and the snapshot, for example: "3,564 de-duplicated Arabic datasets (Bayanat Labs Arabic Data Index, snapshot 2026-11, ADI-CNT-01)".

Files

Path	Licence	What
snapshots/2026-11/arabic-data-index_2026-11.csv	CC BY 4.0	249 rows, 32 metrics: snapshot_id, metric_id, metric_en, metric_ar, dimension, dimension_value, licence_basis, numerator, denominator, value, unit, coverage_pct, mode, method, input_revisions, second_check, notes
snapshots/2026-11/entities_2026-11.csv.gz	CC0 1.0	4,888 rows, one per de-duplicated dataset: Hub id or Masader id, copies, creation date, declared languages and Arabic share, modality, tasks, dialect labels with the source's own labels, declared licence and class, gated, viewer health and size, size category, 30-day downloads, likes, and Masader's factual fields (masader_*)
snapshots/2026-11/*.parquet	as the CSV	Parquet copies
snapshots/2026-11/FROZEN.json	CC BY 4.0	input revisions, the SHA-256 of every Hub page read (35 pages), read times, configuration digests and the quoted speech-hours figure
snapshots/2026-11/MODE.json	CC BY 4.0	input mode and source choices
data/aggregates/, data/entities/, data/history/, data/definitions.parquet	as above	the configs above (history: the creation-month series and Masader years)
latest.json	CC BY 4.0	headline figures of the snapshot
qa/	CC BY 4.0	second computation, re-run check, merge checks
PROVENANCE.jsonl	CC BY 4.0	one row per request (8,887 rows): URL, status, UTC time, SHA-256, purpose
SHA256SUMS, MANIFEST.json		checksums

Metric dictionary

The dictionary in English and Arabic, with notes, is data/definitions.parquet.

Id	Metric
ADI-CNT-01	De-duplicated Arabic datasets (headline datasets; massive multilingual datasets excluded)
ADI-CNT-02	Hugging Face dataset repositories tagged with an Arabic language code (all, copies and massive multilingual datasets included)
ADI-CNT-03	De-duplicated datasets by Arabic share (only, bilingual, multilingual, massive)
ADI-CNT-04	Headline datasets by where they are listed (Hugging Face only, both, Masader only)
ADI-CNT-05	Hugging Face repositories counted once with another repository as copies of the same dataset
ADI-CNT-06	Headline datasets listed in Masader, by their Hugging Face status
ADI-CNT-07	Masader entries read
ADI-CON-01	Share of headline Hugging Face datasets held by the 10 accounts with the most
ADI-CON-02	Share of 30-day downloads of headline Hugging Face datasets that goes to the 20 most downloaded
ADI-CON-03	The 20 most downloaded headline Hugging Face datasets (30-day downloads)
ADI-DIA-01	Headline datasets by declared Arabic variety or dialect
ADI-DIA-02	Headline datasets that declare an Arabic variety or dialect beyond Arabic
ADI-DIA-03	Headline datasets that declare at least one regional dialect
ADI-GRW-01	Headline Hugging Face datasets still on the Hub, by creation month
ADI-GRW-02	Headline Hugging Face datasets created in the last full month before the snapshot
ADI-GRW-03	Headline Hugging Face datasets created in the snapshot year, up to the snapshot
ADI-GRW-04	Headline datasets listed in Masader, by the year Masader gives
ADI-HLT-01	Headline Hugging Face datasets that are gated
ADI-HLT-02	Headline Hugging Face datasets whose dataset viewer works
ADI-HLT-03	Headline Hugging Face datasets by dataset-viewer size status
ADI-LIC-01	Headline datasets by declared licence class
ADI-LIC-02	Headline datasets by declared fine licence class
ADI-LIC-03	Headline datasets that declare a software licence (MIT, Apache, BSD, GPL and similar) for their data
ADI-LIC-04	Share of headline datasets with a declared open-commercial licence, by modality
ADI-LIC-05	Headline Hugging Face datasets without a licence tag on the Hub
ADI-MOD-01	Headline datasets by modality
ADI-SIZE-01	Rows in headline Hugging Face datasets with a complete size from the dataset viewer
ADI-SIZE-02	Parquet bytes of headline Hugging Face datasets with a complete size from the dataset viewer
ADI-SIZE-03	Headline Hugging Face datasets by Hub size category
ADI-SIZE-04	Open, commercially usable Arabic speech, released audio hours, overlap-adjusted (quoted)
ADI-TSK-01	Headline datasets by task family
ADI-TSK-02	Headline datasets with no task declared

How it was built

• Hugging Face. The datasets API with filter=language:<code> for each code, 1,000 rows per page, following Link: rel="next", read between 2026-11-01T06:05:32.537Z and 2026-11-01T06:06:07.211Z. The Hub's ?language=ar parameter returned unfiltered results and is not used. The dataset viewer's /size and /is-valid were asked for every repository that is not gated, between 2026-11-01T06:07:08.203Z and 2026-11-01T08:35:24.845Z; no data file was downloaded. The 609 viewer requests that failed (HTTP 500, 501 or 502) were requested once more, about 2.5 hours later (1.1 to 3.4 hours after the original request): 1 then succeeded and is used, and 608 failed again and are counted with the status of their original reply (viewer error, not supported, own loading code or empty; ADI-HLT-03).
• Masader (ARBML). git clone at commit ce2c342efcba2dd2333be62421b6305089a55b42 (2026-10-05); the Hub copy arbml/masader was last modified on 2024-12-21, so the git repository is used. Only factual fields are kept: name, links, licence, year, language, dialect and dialect subsets, form, volume and unit, host, access and tasks. No description or other prose is copied. Source: Masader (ARBML).
• Requests. User-Agent BayanatLabs-OpenData/1.0 (+https://bayanatlabs.com/open; hello@bayanatlabs.com), no token, at least one second between requests to a host, the Hub's rate-limit header read after every reply, and a stop on 429, 403, a challenge or a login redirect (none occurred).
• One dataset, one row. Copy groups of the Arabic Dataset Rights Index (revision a422db2) count once. Review candidates are name matches (token-set ratio of at least 92, year within one), Masader entries that share a link, same-size copies on Hugging Face (repositories with an identical complete dataset-viewer size whose names match, or a group of 3 or more such repositories with at least 10 MB of Parquet files), and every Masader link to a Hugging Face repository that receives two or more Masader entries. Bayanat Labs checked 406 pairs: all 296 candidates (61 name matches, 30 shared links, 188 same-size copies and 17 Masader links to a shared repository), a seeded random sample of 112 rule-based joins (80 Masader Hub links and 32 copy-group members; 3 of them are also candidates), and 1 pair found while entity rows were reviewed. 201 candidate pairs were merged (188 of them same-size copies), 111 sampled joins were confirmed, and 3 Masader Hub links that name a different dataset were undone. The checks are in qa/dedup_check_2026-11.csv.
• Classes. Arabic share, modality, task families, dialect labels and licence classes follow the configuration files in the code repository (src/arabic_data_index/config/). The licence tag table and the software and non-commercial tests are those of the Arabic Dataset Rights Index, so both releases read a tag the same way. Dialect labels come from language codes and Masader's dialect fields only; a plain ar tag is "not specified".
• Two computations. The aggregates are computed in SQL (DuckDB) and again in pandas; the two agree on all 249 of 249 rows (numerator, denominator, value and coverage). Recomputing the aggregates from the frozen entities table and FROZEN.json gives the same CSV byte for byte.
• Speech hours are quoted from the Open Arabic Speech Hours Counter (revision 696cc8e, as of 30 October 2026), not recomputed.

Considerations

• The creation-month series counts datasets that still exist on the Hub, so earlier months are undercounted. 32 headline repositories carry the Hub's migration date of 2 March 2022 (they were created earlier) and are counted in their own bucket.
• 19.3% of headline datasets declare no task, and dialect is declared beyond "Arabic" for 30.7%; labels are not inferred where nothing is declared.
• Rows mean different things in text, speech and image datasets. The row and byte totals cover 71.7% of headline Hugging Face datasets (those with a complete viewer size) and are counts, not a measure of data. One dataset holds 29.6% of these rows, more than any other.
• Same-size copies are found from complete viewer sizes; copies whose viewer size is partial, failed or lacks the original file size can still be counted more than once.
• The speech-hours figure is quoted from the Open Arabic Speech Hours Counter as published, with its as-of date; that release lists the corpora it covers and how they are classed.
• ADI-CON-01 does not name accounts; ADI-CON-03 lists the 20 most downloaded datasets by repository id as published on the Hub, and ids include account names.

Licensing

The aggregates, definitions, charts and history are released under CC BY 4.0. The per-dataset entities table (entities_*) holds factual metadata only and is released under CC0 1.0. Masader-derived columns are credited "Source: Masader (ARBML)" (https://github.com/ARBML/masader); see LICENSE-sources.md. Each indexed dataset keeps its own licence; nothing from the datasets themselves is included.

Citation

Bayanat Labs Arabic Data Index, snapshot 2026-11, CC BY 4.0. https://huggingface.co/datasets/Bayanat-Labs/arabic-data-index

@misc{bayanatlabs_arabic_data_index_2026_11,
  title  = {Arabic Data Index, snapshot 2026-11},
  author = {{Bayanat Labs}},
  year   = {2026},
  note   = {CC BY 4.0 (aggregates), CC0 1.0 (entities table)},
  url    = {https://huggingface.co/datasets/Bayanat-Labs/arabic-data-index}
}

Corrections

Corrections and removal requests: hello@bayanatlabs.com.

بالعربية

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

مؤشر البيانات العربية من بيانات لابز: إحصاءات مؤرّخة عن مجموعات بيانات الذكاء الاصطناعي العربية في Hugging Face وفي فهرس Masader. اللقطة 2026-11، قُرئت في 1 نوفمبر 2026.

• مجموعات البيانات العربية بعد إزالة التكرار: 3,564 (ADI-CNT-01)، دون 1,324 مجموعة تعلن 10 لغات أو أكثر.
• مستودعات Hugging Face التي تحمل وسم لغة عربية: 4,333 (ADI-CNT-02).
• تعلن 46.9% منها ترخيصًا مفتوحًا يسمح بالاستخدام التجاري، و12.7% لا تعلن أي ترخيص (ADI-LIC-01). الفئات تصف الترخيص المعلن فقط، ولا تُجمع، وليست استشارة قانونية.
• تُعدّ مدخلات Masader الموصوفة بأنها متعددة اللغات (181) ضمن متعددة اللغات لأن Masader لا يذكر عدد اللغات؛ أما المدخلات المعروف أنها تغطي 10 لغات أو أكثر (17) فتُعدّ وحدها.
• طُلبت طلبات عارض البيانات الفاشلة (609) مرة أخرى بعد نحو 2.5 ساعة؛ نجح منها 1 وفشل 608 مرة أخرى، وتُعدّ بحالة ردها الأصلي.
• لا يذكر ADI-CON-01 أسماء الحسابات؛ ويسرد ADI-CON-03 أكثر 20 مجموعة تنزيلًا بمعرّف المستودع كما يُنشر في المنصة، والمعرّفات تتضمن أسماء الحسابات.
• الجداول المجمّعة بترخيص CC BY 4.0، وجدول الكيانات بترخيص CC0 1.0. للتصحيح: hello@bayanatlabs.com