RaR-Science-20k-o3-mini

Anisha Gunjal, Anthony Wang, Elaine Lau, Vaskar Nath, Yunzhong He, Bing Liu, Sean Hendryx

Type: dataset
Venue: HuggingFace (anisha2102)
Year: 2026
Source: huggingface
Access: free
Language: English
Added: 2026-07-17T20:18:03.704464+00:00
Verified: 2026-07-17T20:18:03.704464+00:00

Summary

RaR-Science-20k-o3-mini is a dataset of ~22.9k science questions (18.3k train / 2.29k val / 2.29k test) each paired with a reference answer and checklist-style rubric criteria generated by OpenAI's o3-mini model. It is released as part of the 'Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains' paper (arXiv:2507.17746) to enable structured, multi-criteria reward supervision for RL training of language models on science reasoning tasks.

Keywords

rubrics reinforcement-learning science reasoning rewards o3-mini rlvr

Topics

Science / Reasoning

Research notes

Part of the Rubrics as Rewards (RaR) framework. Source questions appear to come from science QA datasets (e.g. SciBench, GPQA-style problems). Rubrics generated by o3-mini.