Translation pairs
Native translations that preserve meaning and local phrasing, with train/held-out split at collection time.
- English ↔ Hausa
- English ↔ Yoruba
- English ↔ Nigerian Pidgin
Hausa · Yoruba · Pidgin datasets
Datasets for the Nigerian languages that break generic models.
Commission clean, consented Hausa, Yoruba, and Nigerian Pidgin data for model training and evaluation: translation, speech, preference, safety, and cultural-quality tasks.
Human contact: Chat on WhatsApp · [email protected]
Pidgin
Naturalness and culture judgement
Yoruba
Fresh train/held-out translation
Hausa
Regional coverage and speech
Speech datasets
ASR and TTS collection for Nigerian language speech, with recording protocol and consent.
- Prompted voice reads
- Transcript review
- Speaker consistency options for TTS
RLHF preferences
A/B judgement tasks where Nigerian speakers choose the more natural, helpful response.
- NGPT-compatible export
- Majority voting
- Gold-check quality scoring
Cultural quality ratings
Human scores for whether text sounds like a Nigerian wrote it, not just whether BLEU or chrF looks good.
- Pidgin naturalness
- Politeness and formality
- Local context fit
Commercial-safe collection
Consent and provenance are built into the export path instead of added later.
- AI-data reuse consent
- Anonymised contributor ids
- Gold probes excluded from corpus
Delivery formats
Ship data in the format your training pipeline expects: JSONL, CSV, TMX-like translation tables, or custom manifests.
- Dataset card
- Language and split fields
- Provenance metadata
Own the Nigerian-language data loop.
Start with one language and one task type, then expand once quality and cost are proven.