AfriStereo

African Stereotype Evaluation Dataset

bias-evaluation stereotypes safety ~5,000 pairs Open Source

Description

First open-source African stereotype dataset grounded in local socio-cultural contexts. Stereotype-antistereotype pairs for evaluating bias in models.

Quick Start

Python
# Dataset coming soon on HuggingFace

Project Ideas

LLM Bias Comparison

Evaluate GPT-4 vs Llama-2 on African gender stereotypes

Intermediate

Reward Model Training

Train reward model to penalize stereotyping

Advanced

Ethical Considerations

  • Contains offensive stereotypes for evaluation purposes