Can I use this Arabic dataset?

Can a commercial model be trained on this Arabic dataset? Can copies be shared? The search page of the Arabic Dataset Rights Index by Bayanat Labs answers for 120 datasets on Hugging Face: the 100 most-downloaded datasets tagged Arabic and 20 more repositories in sets of known copies. For each dataset it shows what the declared licence and the dataset card say about commercial model training, redistributing copies and share-alike, which flags the card raises, and which copies of the same data declare a different licence. Each answer opens to its evidence: the quoted line, the commit read and the retrieval date. Assessed from the declared licence and card text on 19 October 2026; not legal advice; check the source terms.

Open the search page · Data on Hugging Face · Code on GitHub · Release record

What you can do with it

• Search by repository name, set name or licence identifier, for example MASC or cc-by-nc.
• Filter by the commercial-training answer, the redistribution answer, a flag (such as "licence not declared", "card restricts use or says the underlying data keep their own licences", "gated: access terms apply" or "a linked copy of the same dataset declares a different licence") or the selection.
• Open an entry to read the sentences behind the answers, the linked copies and their licences, and each quote with a link to its line.
• Share a search: the filters stay in the address after the # sign.

What the declared licences and cards say

Among the 100 most-downloaded datasets tagged Arabic on 19 October 2026:

• Commercial model training: permitted for 61, not permitted for 5, and unclear for 34.
• Redistributing copies: permitted for 6, permitted with obligations such as attribution or share-alike for 56, and unclear for 38.
• Use restrictions in the card: the cards of 14 restrict use (to research, non-commercial or personal use) or say the underlying data keep their own licences. Such a card makes an answer that would permit unclear: for 6 datasets the declared licence alone would permit commercial training.
• Software licences on data: 33 declare a software licence (MIT, Apache, GPL or similar) for their data, and 28 of the 61 commercial-training permits rest on one.
• No licence: 11 declare none, so their answers are unclear.
• Personal data: the cards of 17 mention personal data or privacy.

Copies that declare different licences

The index found 45 groups of linked copies, with 118 repositories. In 4 groups the copies declare different licences:

Which licence applies is a question for the rights holder; the index shows the disagreement and the evidence.

From Python and the command line

from datasets import load_dataset
verdicts = load_dataset("Bayanat-Labs/arabic-dataset-rights-index", "verdicts", split="train")
row = verdicts.filter(lambda r: r["repo_id"] == "pain/MASC")[0]
print(row["commercial_training"], "|", row["commercial_training_text"])

pip install "git+https://github.com/Bayanat-Labs/arabic-dataset-rights-index@v0.1.0"
hf download Bayanat-Labs/arabic-dataset-rights-index --repo-type dataset --local-dir rights-index
bl-rights lookup rights-index pain/MASC

Python 3.10 or later. The package is not on PyPI yet; install it from GitHub as above. The same tool rebuilds the index from a clone of the repository, which holds the search configuration: bl-rights harvest reads public Hub metadata, cards and licence files with at least one second between requests, and bl-rights build writes the tables.

git clone --branch v0.1.0 https://github.com/Bayanat-Labs/arabic-dataset-rights-index && cd arabic-dataset-rights-index && pip install .
bl-rights harvest --config config/search.yaml --raw raw/
bl-rights build --config config/search.yaml --raw raw/ --out build/

How the answers are computed

The selection is the 100 datasets tagged Arabic with the most downloads on the Hub on 19 October 2026, plus 20 repositories in sets of known copies. For each dataset that is not gated, the card and any licence file are read at the commit the Hub lists. The declared licence is taken from the card metadata, a licence file and the card text, and checked with the dataset licence checker (license-check 0.1.0, rules version 2026.10.19) for commercial model training and for redistributing copies. When declared licences disagree, or one is not covered by the rules, the answer is unclear. When the card limits use to research, non-commercial or personal use, or says the underlying data keep their own licences, an answer that would permit is unclear. Card statements are found by fixed patterns, and each quote is an exact span of the file. The build made 637 requests to the Hub and recorded 432 evidence rows. The full method is in docs/METHOD.md.

Limits

• The answers concern the declared licence and the card; the data may include material under other terms, and a dataset built from sources with their own terms can carry a licence its sources do not allow.
• Verdicts are computed by fixed rules and are not individually reviewed. A wrong or missing licence tag on the Hub gives a wrong answer here, and a use restriction the patterns do not find leaves an answer as it is.
• Gated datasets are not opened, and copies with changed files and no shared link are not found.

Cite as

Bayanat Labs (2026). Arabic Dataset Rights Index, version 0.1.0, assessed 2026-10-19. https://huggingface.co/datasets/Bayanat-Labs/arabic-dataset-rights-index. CC0 1.0 (Hub metadata facts), CC BY 4.0 (verdicts, flags and notes).

The release record, with licences and changelog, is the Arabic Dataset Rights Index on Bayanat Labs Open.

Licence and contact

Hub metadata facts: CC0 1.0. Verdicts, flags, notes and documentation: CC BY 4.0. Code and search page: Apache-2.0. Quoted card text remains under its owners' terms. Corrections and removal requests: hello@bayanatlabs.com.

Need Arabic training data with clear licensing?

The index is free. For licensed Arabic text collections with their own counts and licensing details, see Bayanat Knowledge.