AI Evaluation
African AI Evaluation Dataset
Preference pairs for answers about Nigerian farming, markets and public information.
Dataset description
Small set of prompts with a preferred and a weaker answer, written so reviewers can practise AI response evaluation. Not a leaderboard and not customer traffic.
- Creator
- AI HUB Data Team (demo)
- Records
- 20,000
- Format
- JSONL
- Collection period
- Demo catalogue
- Updated
- 2026-11-07
- Downloads of samples
- 0
Quality score 92/100
Research sample
Samples use a 10-minute grant. Full files are not hosted.
Quality breakdown
20% completeness + 18% accuracy + 14% consistency + 12% annotation agreement + 14% documentation + 12% duplicate health + 10% metadata.
Completeness
95%
Accuracy
90%
Consistency
92%
Duplicate rate
0.5%
Annotation agreement
87%
Documentation
93%
Metadata
90%
Data preview
Safe sample only. Restricted fields are not shown.
| prompt | preference | note |
|---|---|---|
| How should a Kano tomato farmer rotate crops? | A | A names a practical rotation. B is generic. |
Data card
- Purpose
- Score whether a model answer is useful, local and safe.
- Collection method
- Written evaluation items.
- Annotation method
- Raters pick the better answer. Gold items are hidden.
- Source
- Demo catalogue.
- Size
- 20,000 catalogue target. Sample only is hosted.
- Fields
- prompt, response_a, response_b, preference, rationale
- Language
- English, Hausa
- Geography
- West Africa
- Limitations
- Narrow topics. Agreement on the full 20,000 is not measured here.
- Potential bias
- English prompts dominate. Hausa items are few.
- Recommended uses
- Teaching evaluation guidelines.
- Prohibited uses
- Publishing a model ranking from this sample.
- Licence
- research
Versions
v1.0 · 2026-03-12
Demo evaluation card.
+20000 / −0 · Initial · prompt, A, B, preference
Report dataset
Use this for privacy, copyright, harmful content or a misleading card. Do not attach more personal data than the report needs.
You can also use the report desk.