Dataset · allenai
Human preference dataset by AllenAI (MultiPref)
allenai/multipref
10,000 English examples with cross-evaluations from crowdworkers and experts for LLM alignment.
What's inside
The collection includes 10,000 examples in English, each evaluated four times by two regular annotators and two domain experts, totaling approximately 40,000 annotations. In addition to overall response quality ratings, the markup includes several additional preference aspects. The dataset was released by the AllenAI institute to study hybrid human-AI feedback.
What you can do with it
- Training reward models and aligning language models using RLHF or DPO methods
- Studying discrepancies between evaluations by regular users and narrow specialists
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 20.5K
- Likes
- 24
How to load it
from datasets import load_dataset
ds = load_dataset("allenai/multipref")