Ilo-v1: Arabic ↔ Mauritanian languages parallel corpus

Ilo-v1 links Modern Standard Arabic to the four national languages of Mauritania: Hassaniya, Pulaar, Wolof and Soninke. It has 2,000 pairs per language, 8,000 pairs in total. It is the dataset behind MauriCorpus, presented at Deep Learning Indaba 2026 (poster IX-40).

Mauritania is one of the few countries where Arabic and Sub-Saharan African languages are spoken side by side in the same communities. Yet public multilingual resources contain almost no sentence pairs linking Arabic to Pulaar, Soninke or Wolof, and Soninke is not covered by any major multilingual model (NLLB-200, MADLAD-400).

Subsets

Config	Pairs	Languages	Content
arabic_pulaar (default)	2,000	Arabic, Pulaar	Culture and history articles
arabic_hassaniya	2,000	Arabic, Hassaniya (+ French, English)	Everyday conversational sentences
arabic_wolof	2,000	Arabic, Wolof	Same Arabic sentences as arabic_pulaar
arabic_soninke	2,000	Arabic, Soninke (+ French)	Vocabulary, example sentences and short texts
Total	8,000

from datasets import load_dataset

pulaar    = load_dataset("mwiml/Ilo-v1", "arabic_pulaar", split="train")
hassaniya = load_dataset("mwiml/Ilo-v1", "arabic_hassaniya", split="train")
wolof     = load_dataset("mwiml/Ilo-v1", "arabic_wolof", split="train")
soninke   = load_dataset("mwiml/Ilo-v1", "arabic_soninke", split="train")

Each subset has a single train split, so make your own train/test split.

How each subset was made

The subsets were built in different ways. Please read this before you train or evaluate on them.

Subset	Arabic side	Other side
arabic_pulaar	Human-written	Machine translation (Google Cloud Translation, Fula ff), then all 2,000 pairs reviewed by a native Pulaar speaker
arabic_hassaniya	Human-written	Human translation
arabic_wolof	Human-written	Machine translation (NLLB-200-distilled-600M, wol_Latn). Not reviewed by native speakers
arabic_soninke	Machine translation from the French (Google Cloud Translation)	Soninke–French collected by engineer Maimouna Waiga from published Soninke sources

What this means in practice:

• Hassaniya is human-written on both sides.
• Pulaar started as Google Translate output and was reviewed in full by a native speaker. If you use it for evaluation, systems that translate the way Google Translate does may score a little higher.
• Wolof is unreviewed machine output. Treat it as silver data, useful for augmentation or pre-training, and don't use it as gold test references. It was produced with NLLB-200, so NLLB-based models will score unrealistically high against it.
• Soninke: the Soninke and French are human-written, collected by engineer Maimouna Waiga. The Arabic was machine-translated from the French and may contain errors. About 60% of the entries are single words ("level": "word").
• The subsets are kept exactly as used in the baseline experiments, so they contain a few duplicate pairs (1 in Hassaniya, 27 in Soninke). Remove them if you need unique pairs.

Fields

arabic_pulaar and arabic_wolof

Both subsets share the same 2,000 Arabic sentences in the same order. The id values match, so they can be joined into Arabic–Pulaar–Wolof triples.

{
  "id": 1,
  "arabic": "افتتحت رابطة أدباء آدرار ليل السبت موسمها الثقافي ...",
  "pulaar": "Fedde winndooɓe Adrar udditi jamma aset ...",
  "category": "History",
  "source_url": "https://www.culture.gov.mr/index.php/ar/node/112"
}

Field	Description
pulaar / wolof	The translation
category	Topic of the source article: History (909), General (360), Education (262), Environment (167), Health & Wellness (100), Art & Culture (78), and others
source_url	Article the Arabic sentence came from (Mauritanian Ministry of Culture website). null for 93 sentences whose source could not be matched

When you split for training, keep sentences from the same source_url on the same side of the split.

arabic_hassaniya

{
  "id": 1,
  "arabic": "يمكنني أن آخذها.",
  "hassaniya": "أنگد نگبظها.",
  "french": "Je peux le prendre.",
  "english": "I can take it."
}

french and english are reference translations of the same sentence.

arabic_soninke

{
  "id": 1,
  "level": "sentence",
  "arabic": "أنا، سوكونا تونكارا، المقيم حاليًا في بالي، سأروي هذه القصة لأبنائي وأحفادي.",
  "soninke": "In ke Soxona Tunkara, in wo taaxunu Balle.",
  "french": "Moi, Sokona Tounkara, actuellement installé à Ballé, vais narrer ce conte pour mes enfants et mes petits-enfants.",
  "category": null,
  "source": "https://www.calameo.com/books/0009039477a9d2d6c6af5"
}

Field	Description
level	word (1,210), sentence (780) or text (10)
category	Thematic category for vocabulary items, in French (Animaux, Santé, Verbes, Famille, …). null if none
source	Publication the Soninke–French entry was collected from

The Soninke–French entries were collected by engineer Maimouna Waiga from these sources:

Source	Pairs
Parlons soninké, Christian Girier, L'Harmattan, 1996	988
Lexique Soninké (Mali)–Français–Anglais, SIL	384
Calaméo publication 0009039477a9d2d6c6af5	241
Soninkara.org: Histoire des forgerons de Barâgo, Doudou Sakho	70
SOONINKARA magazine, A.P.S., no. 1, 1988	29
Dangamaanu folk tales, SIL Mali	10
Source not recorded	278

Baseline results

We fine-tuned NLLB-200-distilled-600M on all four subsets and trained the eight directions jointly. NLLB-200 has no tokens for Soninke or Hassaniya, so we added two new language tokens: snk_Latn (initialised from Bambara) and mey_Arab (initialised from Arabic). The model was trained for 10 epochs.

Direction	n	BLEU	chrF++
Arabic → Pulaar	400	17.9	42.9
Pulaar → Arabic	400	7.2	32.0
Arabic → Hassaniya	398	5.4	30.5
Hassaniya → Arabic	398	12.8	35.3
Arabic → Wolof	400	14.2	28.8
Wolof → Arabic	400	4.4	22.2
Arabic → Soninke	403	0.6	10.1
Soninke → Arabic	403	0.7	10.1

Some test references started as machine translation (see How each subset was made), so read these scores as a starting point, not as absolute translation quality. Arabic has rich morphology, which lowers BLEU for translation into Arabic, so chrF++ is the more informative metric for those directions.

Limitations

• The subsets cover different domains: Hassaniya is conversational, Pulaar and Wolof lean towards culture and history articles, and Soninke is mostly vocabulary. None is a balanced sample of its language.
• Pulaar, Soninke and Wolof are written in Latin script. Arabic-script (Ajami) writing is not covered yet.
• Spelling of the African languages varies across sources and has not been standardised.

License

Ilo-v1 is released under CC BY-NC-SA 4.0. You may use and share it for non-commercial purposes, with attribution, under the same license. For commercial use, contact the authors.

Some source text belongs to third parties and remains under its original terms: Parlons soninké (Christian Girier, L'Harmattan, 1996), SIL Mali (Soninke lexicon and folk tales), Soninkara.org, the SOONINKARA magazine of the A.P.S., and the Mauritanian Ministry of Culture website (culture.gov.mr). The Wolof side was generated with NLLB-200 (CC BY-NC 4.0).

Citation

@misc{boughale2026mauricorpus,
  title        = {MauriCorpus: Bridging the Gap Between Arabic and Mauritanian African Languages},
  author       = {Boughale, Rahma and Aw, Mamadou Abdoulaye and Maimouna, Oiga and Ba, Ahmed},
  year         = {2026},
  howpublished = {Poster, Deep Learning Indaba 2026}
}

Acknowledgements

Soninke data collected by engineer Maimouna Waiga. Published by MWiML (Mauritanian Women in Machine Learning), with support from the Mauritanian Machine Intelligence Hub (2MIH) and native-speaker communities in Mauritania.

Contact

Rahma Boughale: rahma.boughale@gmail.com