Dataset · allenai
Advanced benchmark for evaluating reward models
allenai/reward-bench-2
Updated dataset based on new human evaluations for testing reward models (RLHF).
What's inside
The second, significantly more challenging version of AllenAI's RewardBench benchmark, compiled from previously unused English human data. Includes 1k to 10k examples in tabular and text formats to test models across new categories, including factual accuracy (hallucination detection) and precise instruction following.
What you can do with it
- Evaluating reward model quality during LLM tuning via RLHF
- Testing model ability to detect hallucinations and response errors
- Benchmarking adherence to complex user instructions
- Tasks
- Question answering
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 11.3K
- Likes
- 37
How to load it
from datasets import load_dataset
ds = load_dataset("allenai/reward-bench-2")