Skip to content

Dataset · allenai

RewardBench benchmark for evaluating reward models

allenai/reward-bench

Dataset for testing reward models on chats, complex dialogues, and preferences; English text only.

What's inside

RewardBench is an evaluation dataset for assessing reward model capabilities. The card lists Chat categories with simple chat subsets (alpacaeval-easy, alpacaeval-length, alpacaeval-hard, mt-bench-easy, mt-bench-medium) and Chat Hard with complex ones (mt-bench-hard, llmbar-natural, llmbar-adver-neighbor, llmbar-adver-GPTInst, llmbar-adver-GPTOut). The language is English; the format is not detailed in the card.

What you can do with it

  • Evaluating and comparing reward models for RLHF
  • Choosing a reward model prior to PPO/DPO training
  • Testing model robustness on complex and adversarial dialogues
Tasks
Question answering
Languages
English
Size
1K–10K rows
Downloads this month
19.8K
Likes
113

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("allenai/reward-bench")