6 July 2026 · Wycord
African training data: verified datasets organized by use case
Africa is underrepresented in most public ML datasets. This guide maps the verified datasets in the Wycord registry by task: language, geospatial, economic, health, and conflict data.
Standard ML training corpora skew toward English, European languages, and Western markets. For developers building products for Africa, the gap surfaces quickly: an NER model that misses Yoruba names, a geocoder that fails outside a capital city, an economic model with no Nigeria-specific signals.
The datasets below are all in the Wycord datasets registry, reviewed for accessibility, format, and active availability. This is a map organized by task, so you can find the right source without a literature search.
Language and NLP
The Masakhane initiative has produced the most comprehensive open African-language NLP datasets available. Several are on HuggingFace and widely cited in academic NLP.
Named entity recognition: MasakhaNER 2.0 covers 20 African languages including Hausa, Igbo, Yoruba, and Nigerian Pidgin, with 152,786 annotated rows using PER, ORG, LOC, and DATE tags. It is the standard NER benchmark for African NLP and the starting point for any language model you are adapting to the region.
Text classification: MasakhaNEWS provides news topic classification for 16 African languages, roughly 31,000 rows across seven topic categories. MasakhaPOS provides part-of-speech tagging using Universal Dependencies tags for 20 languages including Nigerian Pidgin.
Machine translation: MAFAND-MT is the largest news-domain machine translation benchmark for African languages: 21 languages, 142,909 parallel sentences in Parquet, with English or French as source. AfriQA provides cross-lingual open-retrieval question answering for 10 African languages, 12,159 examples with train/val/test splits.
Hate speech detection: AfriHate covers 15 African languages across tweets annotated by native speakers, with multi-annotator labels. Published at NAACL 2025.
Sentiment: NaijaSenti / AfriSenti is the largest Nigerian sentiment corpus: roughly 30,000 annotated tweets per language covering Hausa, Igbo, Yoruba, and Pidgin.
Speech: AfriSpeech-200 is a 200-hour Pan-African accented English speech corpus from 2,463 speakers across 120 African accents and 13 countries, built by Intron Health for clinical and general ASR. AfriSpeech-Dialog adds roughly 6 hours of two-speaker conversational audio across 11 African accents, suited to speaker diarization tasks. Nigerian Pidgin ASR provides 4,277 quality-filtered recordings from 10 native Pidgin speakers with sentence-level transcriptions.
Multimodal: Hausa Visual Genome pairs 32,923 images with Hausa and English region descriptions, useful for image captioning and Hausa-English translation.
Licensing note: most Masakhane and Intron Health datasets carry CC BY-NC licenses. Free for research; commercial deployment requires a separate arrangement.
Geospatial
Administrative boundaries:
- LGA Boundaries (GRID3): OSGOF-backed LGA polygons for all 36 Nigerian states plus FCT, GeoJSON and Shapefile, no auth. This is the authoritative source.
- Kenya Admin Boundaries (OCHA COD): counties, sub-counties, and wards as Shapefile via OCHA HDX.
Buildings and settlements:
- Google Open Buildings (Africa): 1.8 billion building detections with confidence scores and Plus Codes. Download as CSV by S2 cell or query via Google Earth Engine.
- GRID3 Nigeria Settlement Extents v4.0: settlement blocks with building counts, Google Open Buildings 2.5D heights, and Sentinel-2 vegetation metrics. Released March 2026, CC BY-SA 4.0.
Population:
- WorldPop Nigeria 2020: gridded population raster at roughly 100m resolution (~59 MB GeoTIFF), UN-adjusted. CC BY 4.0.
- GRID3 Population Estimates v3.0: modelled 100m raster aggregatable to LGA and ward. Nigeria's last census was 2006; treat all population figures as modelled estimates.
Raw geodata:
- OpenStreetMap Nigeria Extract: daily-updated .osm.pbf (~677 MB) and Shapefile (~2 GB) from Geofabrik under ODbL.
Economic and Financial
No authentication required:
- World Bank API for Nigeria:
api.worldbank.org/v2/country/NG/indicator/returns GDP, CPI, FDI, remittances, and employment indicators as JSON with no API key required. The fastest way to pull Nigerian macro time-series into a notebook or pipeline. - NBS CPI and Inflation: monthly consumer price index from Nigeria's National Bureau of Statistics, rebased to COICOP-2018 methodology, downloadable as Excel.
Registration or login required:
- CBN Statistics Database: official Central Bank of Nigeria store for exchange rates, money-market rates, government securities, and balance-of-payments data, exportable to Excel.
- FEWS NET Nigeria Staple Food Prices: weekly market prices for roughly 13 staple commodities across 73 Nigerian markets since 2021. Direct CSV, XLSX, and JSON download from the FEWS NET portal.
Health
- Health Facility Registry (HUMDATA): 39,081 geo-located Nigerian hospitals and clinics with name, type, LGA, state, coordinates, and registration status, in CSV, no login. Use this HUMDATA mirror; the official
hfr.health.gov.nghad broken TLS as of June 2026. - Sub-Saharan Africa Public Health Facilities: 98,745 public-sector facilities assembled from national master facility lists across 50 sub-Saharan countries. Published in Nature Scientific Data, CC BY 4.0.
- GRID3 Health Facilities v2.0: Nigeria-specific GIS-ready health facility data on the GRID3 Data Hub, no auth.
Conflict, Survey, and Satellite
- ACLED Conflict Data: event-level conflict and crime data with deep Nigeria coverage including banditry, kidnapping, farmer-herder incidents, and oil theft. REST API, CSV, and R/Python packages available. Free with registration.
- Afrobarometer: public-attitude surveys across roughly 39 African countries through Round 9 (2021-2023), covering governance, the economy, and society. SPSS, Stata, and CSV; free with registration.
- Digital Earth Africa EO Data: analysis-ready satellite products for all of Africa, including cropland extent, water observations from space, and coastline change. Served via STAC API and open AWS S3. Continuously updated from Landsat and Sentinel.
What to use when
| You're building... | Start here |
|---|---|
| NER for African-language text | MasakhaNER 2.0 |
| African-accented ASR or TTS | AfriSpeech-200 |
| Machine translation into African languages | MAFAND-MT |
| Nigeria geospatial feature engineering | LGA Boundaries (GRID3) |
| Building density or settlement modeling | Google Open Buildings + GRID3 Settlement Extents |
| Nigeria macroeconomic regression | World Bank API (no key needed) |
| Health facility accessibility modeling | HUMDATA facility registry |
| Agricultural yield or food security | FEWS NET food prices + Digital Earth Africa EO |
| Conflict risk features | ACLED |
| Cross-African governance or sentiment analysis | AfriSenti + Afrobarometer |
Browse and filter the full verified dataset catalog at wycord.com/datasets.