Dataset Summary

RaR-Science is a dataset curated for training and evaluating language models on science domain using structured rubric-based supervision. Each example includes a science related question, a reference answer, and checklist-style rubric annotations generated via OpenAI's o3-mini. This dataset is introduced in Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains.

Dataset Structure

Data Fields

Each example contains:

• question: the open-ended medical question
• reference_answer: high-quality expert reference response
• question_source: source of the original question
• rubric_list: list of rubric criteria used to evaluate the model response
• rubric: dictionary mapping each rubric criterion to a score
• rubric_count: number of rubric criteria used

Usage

from datasets import load_dataset
ds = load_dataset("ScaleAI/RaR-Science")

Citation

If you use this dataset, please cite:

@article{gunjal2025rubrics,
  title={Rubrics as rewards: Reinforcement learning beyond verifiable domains},
  author={Gunjal, Anisha and Wang, Anthony and Lau, Elaine and Nath, Vaskar and He, Yunzhong and Liu, Bing and Hendryx, Sean},
  journal={arXiv preprint arXiv:2507.17746},
  year={2025}
}