8,011 participants from 75 countries contributed 30,000+ human-LLM conversations. These interactions are paired with 10,000+ preference labels and fine-grained demographic metadata for each contributor.
Use Cases
- Train reward models for RLHF using the preference labels and conversation history
- Evaluate model alignment across different cultures using the country and demographic columns
- Analyze how human values vary by demographic group using the participant metadata
Strengths
- 8,011 unique participants from 75 countries
- 30,000+ multi-turn human-LLM conversations
- 10,000+ pairwise preference labels
- Demographic metadata including age, gender, and education