Inkuba-Instruct
Lelapa AIInkuba-Instruct is an instruction-tuning dataset for Hausa, Yoruba, Swahili, isiZulu and isiXhosa, built by recasting existing task datasets such as MAFAND-MT, MasakhaNER, MasakhaPOS, AfriQA, SIB-200, MasakhaNEWS and AfriSenti into instruction, input and output records. It spans translation, named-entity recognition, part-of-speech tagging, question answering, topic classification and sentiment. Lelapa AI released it with the InkubaLM paper.
Only the train and dev splits are public; the card says the test split will follow later.
Openness
2 medium confidence- license
- cc-by-nc-4.0(card body
- access
- auto(automatic Hugging Face gate
- dataset_card
- present(source datasets per task, sample counts and record structure)
The card restricts it to non-commercial use behind an automatic click-through, and its test split is not yet released. The source datasets keep their own licenses, which the card does not reconcile with its own.
- https://huggingface.co/api/datasets/lelapa/Inkuba-instruct recorded 2026-09-24
"gated": "auto"; no license in cardData.
- https://huggingface.co/datasets/lelapa/Inkuba-instruct recorded 2026-09-24
"License: CC BY-NC 4.0"; "Only the train and dev set are currently open sourced. The test set will be made open-source at a later date."
Adoption
1 high confidenceHugging Face downloads of the single repository, behind an automatic gate. Downloads count file fetches, not models tuned on it.
- https://huggingface.co/api/datasets/lelapa/Inkuba-instruct recorded 2026-09-24
24 downloads in the trailing 30 days for lelapa/Inkuba-instruct
Capability
4 medium confidenceInkuba-Instruct turns well-known African task datasets into instruction data for five languages, lists every source on its card, and the Pula-8B model is tuned on it. Like WangchanThaiInstruct it is documented and has a named model tuned on it, but its content recasts existing human-annotated sets rather than new writing by people.
- https://huggingface.co/datasets/lelapa/Inkuba-instruct recorded 2026-09-24
Task table naming source datasets (Mafand-MT, MasakhaNER2, MasakhaPOS, afriqa, SIB-200, MasakhaNEWS, AfriSenti); models trained on it include OxxoCodes/Pula-8B-v0.1.
Verified 2026-09-24