What capabilities does RewardBench test?

RewardBench evaluates llm judging, safety.

How can a model improve its RewardBench score?

Tools linked to RewardBench on Sophon include Reward Bench RL Env (Prime Intellect), HelpSteer2 - RL environments, datasets, and scaffolds that target this eval.

What license is RewardBench under?

RewardBench is available under ODC-BY-1.0.

RewardBench

Active

2,985 prompt-chosen-rejected triples across chat, reasoning, safety, and code - a benchmark for evaluating reward models and LLM judges.

Open

Publisher: Allen Institute for AI (Ai2)
Capabilities: LLM Judging Safety
Format: HF Dataset
Size: 2985 tasks
License: ODC-BY-1.0
Published: Dec 2023
Notable for: Benchmark for evaluating llm judging and safety.
Canonical: github.com/allenai/reward-bench
Also on: huggingface.co/spaces/allenai/reward-bench

Cite

Notes

Only stored in your browser.

Related tools

View all

Implementations, trainers, datasets and scaffolds linked to this eval.

Reward Bench RL Env (Prime Intellect)

Prime Intellect

Evaluates pair-wise answers from RewardBench datasets

Trains towardRL EnvSafetyPair WiseMulti Lingual

HelpSteer2

NVIDIA

NVIDIA's permissively-licensed human-annotated preference dataset with 5-axis Likert ratings - engineered to train high-quality reward models.

Training dataPreferenceHallucinationSafetyInstruction Following

Contributors

NNathan Lambert

FAQ

What is RewardBench?: 2,985 prompt-chosen-rejected triples across chat, reasoning, safety, and code - a benchmark for evaluating reward models and LLM judges.
What capabilities does RewardBench test?: RewardBench evaluates llm judging, safety.
How can a model improve its RewardBench score?: Tools linked to RewardBench on Sophon include Reward Bench RL Env (Prime Intellect), HelpSteer2 - RL environments, datasets, and scaffolds that target this eval.
What license is RewardBench under?: RewardBench is available under ODC-BY-1.0.