This dataset contains French-Adja parallel text created in collaboration with the Adja-speaking community in Benin. Access requires agreement to the terms below. Please fill out all fields.

French-Adja Parallel Corpus

The first publicly available parallel text corpus for Adja machine translation, targeting an under-resourced Gbe language spoken by approximately 1,000,000 people in Benin and Togo.

Dataset Description

This dataset has three configs.

default: 10,000 French-Adja sentence pairs from Tatoeba French, translated with native Adja speakers in the Couffo region of Benin.
structured: 4,251 translated pairs from a grammar-guided curriculum (5 modules). The French design set had 4,266 prompts. 15 prompts do not have Adja translations yet, so the released count is 4,251.
paper_r10k_s4k: the mixed training regime used for the public NLLB models named r10ks4k (Tatoeba 10k + structured curriculum, repaired May 2026 text).

All Adja text is Unicode NFC.

About Adja

Adja (ISO 639-3: ajg) is a Gbe language of the Niger-Congo family, closely related to Fon, Ewe, and Gen. It is spoken by approximately 1 million people, primarily in the Couffo and Mono departments of southern Benin and in southeastern Togo. Despite its significant speaker population, Adja had no publicly available text-based NLP resources prior to this work: no parallel corpora, no machine translation systems, and no labeled computational datasets. Concurrent work by Justin et al. (2025) introduced Eyaa-Tom, a multi-language dataset for Togolese languages that includes Adja among its targets, but the publicly released Adja data consists only of a small amount of audio. No French-Adja parallel text for MT had been made available before this corpus.

Languages

	Language	ISO 639	Script
`fr`	French	fra	Latin
`adj`	Adja	ajg	Latin (with diacritics: ɔ, ɛ, ɖ, ŋ, tonal marks)

Configs

`default` (Tatoeba 10k)

Split	Count
train	8,000
validation	1,000
test	1,000

Columns: fr, adj.

Created with native speakers through oral translation of Tatoeba French (seed 42 random splits). This config is unchanged from the original public release.

`structured` (grammar curriculum)

Split	Count
train	3,462
validation	334
test	455
total	4,251

Columns: fr, adj, sentence_id, base_sentence_id, module.

Splits are group-aware by base_sentence_id (seed 42), so M1-M5 variants of the same base stay in one split.

`paper_r10k_s4k` (thesis / NLLB mix)

Split	Count
train	10,316
validation	1,146
test	1,455

Columns: fr, adj, module, sentence_id, dataset_source.

This is the repaired local R10K+S4K condition behind the public models: JosueG/adja-nmt-nllb-600m-forward-r10ks4k-seed42 and JosueG/adja-nmt-nllb-600m-reverse-r10ks4k-seed42. The shared test set mixes 455 structured pairs and 1,000 Tatoeba pairs.

Dataset Creation

Translation Process (`default`)

Source sentences: 10,000 French sentences selected through uniform random sampling from Tatoeba
Translation team: 5 native Adja speakers from the Couffo region of Benin
Process: French sentences were read aloud, discussed, then translated orally into Adja and transcribed
Duration: 6 months of collaborative work
Quality control: Unicode normalization, spacing cleanup, punctuation checks

Structured curriculum

French prompts were generated module by module (present, negation, past, future, questions). Native speakers translated them into Adja. A May 2026 source/target repair pass cleaned French contamination in the structured text before this public release.

Why Oral Translation?

Adja is primarily a spoken language. Oral translation before transcription helps keep the Adja natural.

Usage

from datasets import load_dataset

# Tatoeba 10k (original public release)
ds = load_dataset("JosueG/french-adja-parallel-corpus")

# Structured grammar curriculum
structured = load_dataset("JosueG/french-adja-parallel-corpus", "structured")

# Mix used by the public NLLB r10ks4k models
paper = load_dataset("JosueG/french-adja-parallel-corpus", "paper_r10k_s4k")

print(ds["train"][0])
print(structured["train"][0])

Ethical Considerations

This corpus was created in collaboration with native Adja speakers in Benin. The translation team was compensated for their work.
The dataset is released under a non-commercial license (CC BY-NC-SA 4.0).
Adja is primarily spoken. This written corpus does not capture full tonal and dialectal variation.
The Tatoeba source sentences reflect global French usage and may not match everyday French in Benin and Togo.

Citation

If you use this dataset, please cite:

@inproceedings{godeme2026french-adja,
    title     = {A 10,000-Sentence French-Adja Parallel Corpus for Machine Translation},
    author    = {Godeme, Josue and Coto-Solano, Rolando},
    booktitle = {Proceedings of the 2026 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2026)},
    year      = {2026},
    note      = {To appear}
}

Acknowledgements

We are deeply grateful to the Adja-speaking community members in the Couffo region of Benin who dedicated their time and expertise to translating and transcribing this corpus. Their commitment to documenting and advancing their language made this work possible.