fatmajlali/tunisian-nlp-resources

Tunisian Arabic NLP Resources

An open, maintained inventory of the NLP resources that exist for Tunisian Arabic (Derja / Tounsi / تونسي, ISO 639-3 aeb).

👉 Full map, and how to add to it: https://github.com/jjlalli/Tunisian-Derja-NLP-Resources

Tunisian NLP has a findability problem more than a quantity problem. People build good things — a dataset, a fine-tune, a benchmark — and they end up on a lab page or a personal account, so the next person spends weeks rebuilding what already existed. This is the place those things get recorded.

What's in it

	Entries
Text datasets, benchmarks, lexicons & papers	71
Speech corpora (ASR, SLU, translation, TTS)	28
Pretrained models (LLMs, encoders, ASR, TTS, OCR)	21
Researchers, labs & companies	30
Total	150

This table export carries 120 rows — the resource entries, regenerated from the markdown inventory by build-dataset-csv.py; the researcher/lab entries live in the repository only.

Every entry carries an honest access tag — of 151 tags applied: 113 open, 13 paywalled, 12 paper-only, 8 on request, 3 gated, 2 commercial. At entry level, which is what the repository's open badge tracks, the export resolves to 90 open of 120 rows; grouped headings that carry several different tags are marked "varies" rather than counted as open. A resource that exists but can't be downloaded stays on the list with the truth attached, rather than being quietly dropped or oversold.

Multi-dialect and pan-Arabic resources are included, with the Tunisian portion described rather than the whole thing counted as Tunisian. Derivatives and repackaged slices of existing corpora are cross-referenced under their parents so nothing is counted twice.

Load it

from datasets import load_dataset

ds = load_dataset("fatmajlali/tunisian-nlp-resources")

Columns

• id — row number
• name — resource name
• type — text / speech / model
• category — section of the inventory (sentiment, dialect ID, ASR…)
• access — open / on request / paywalled / paper only / gated / varies (grouped entry)
• description — one-line summary
• url — canonical link
• file — which inventory file it came from

Scope and limits

This indexes resources; it does not redistribute them. Access status was verified by hand, and links rot. Uncertain Tunisian coverage is flagged rather than dropped.

Add your work

If you have built something for Tunisian Arabic, or you know something missing:

• Add a resource — paste a link; formatting and verification are the maintainer's job
• Report something wrong — a wrong entry is worse than a missing one
• Pull requests — the first external one was merged in September 2026; contributed entries are credited in the repository's "Recently added" table

Citation

@misc{jlali2026tunisiannlp,
  author       = {Jlali, Fatma},
  title        = {Tunisian Arabic NLP Resources},
  year         = {2026},
  howpublished = {\url{https://github.com/jjlalli/Tunisian-Derja-NLP-Resources}},
  doi          = {10.5281/zenodo.21779464},
  note         = {Open inventory of Tunisian Arabic (aeb) NLP resources}
}

Maintained by Fatma Jlali. Licensed CC BY 4.0.