Datasets

AfriQA

Datasets

Cross-lingual open-retrieval question-answering dataset with human-translated QA pairs for 10 African languages (incl. Hausa, Igbo, Yoruba), totaling 12,159 examples across train/validation/test splits. From the Masakhane initiative.

Hausa Visual Genome (HausaVG)

Datasets

Multimodal Hausa-English dataset of 32,923 images with paired English/Hausa region descriptions (train/dev/test/challenge splits), post-edited by HausaNLP and Bayero University Kano translators for English-to-Hausa machine translation and image description.

MasakhaNER 2.0

Datasets

Largest high-quality named-entity-recognition corpus for 20 African languages (incl. Nigerian Pidgin, Hausa, Igbo, Yoruba) with PER/ORG/LOC/DATE tags over news-domain text, totaling ~152,786 rows. Built by the Masakhane community.

MasakhaNEWS

Datasets

News-topic-classification dataset for 16 widely spoken African languages (incl. Hausa, Igbo, Yoruba, Nigerian Pidgin), ~31,088 rows in CSV/Parquet with train/val/test splits across seven topic categories. Built by the Masakhane community.

MasakhaPOS

Datasets

Part-of-speech tagging dataset for 20 African languages (incl. Nigerian Pidgin, Hausa, Igbo, Yoruba) using Universal Dependencies tags, with per-language train/validation/test splits. Built by the Masakhane community.

AfriHate Hate Speech Datasets

Datasets

Multilingual collection of hate speech and abusive language datasets covering 15 African languages, built from tweets annotated by native speakers. Each instance carries labels from 3 to 4 annotators with anonymous annotator IDs, downloadable on HuggingFace. Published at NAACL 2025.

AfriMMLU (IrokoBench)

Datasets

Human-translated multiple-choice question-answering evaluation benchmark covering 16 to 17 African languages plus English and French, derived from a subset of MMLU across subjects like maths, geography and law. Distributed as CSV and parquet on HuggingFace and forms part of the IrokoBench suite (MMLU, MGSM, XNLI). Licensed Apache 2.0.

Kencorpus Kenyan Language Corpus

Datasets

Text and speech corpus for three Kenyan languages, Swahili, Dholuo and Luhya, containing 4,442 texts (5.6 million words) and 1,152 speech files (177 hours). It also ships derived NLP sets: POS-tagged Dholuo/Luhya, 7,537 Swahili question-answer pairs and 13,400 translated sentences. Downloadable from Harvard Dataverse; released 2022.

MAFAND-MT (masakhane/mafand)

Datasets

Largest news-domain machine translation benchmark for African languages, covering 21 languages with English or French as source. It contains 142,909 parallel sentences in parquet with train, dev and test splits, hosted on HuggingFace. Licensed CC BY-NC 4.0.