YuChuang1205/ScienceQA-RAD-Version

ScienceQA-RAD is the rationale-augmented version of ScienceQA, introduced in MIND (ICML 2026). It is constructed using the proposed Rationale Augmentation and Discrimination (RAD) paradigm to enhance multimodal Chain-of-Thought (MCoT) reasoning, logical discrimination, and self-correction.

Dataset Overview

The original ScienceQA dataset contains 21,208 multimodal multiple-choice questions covering three major scientific domains:

• Natural Science: Biology, physics, chemistry, and other natural sciences.
• Social Science: Geography, history, and related social sciences.
• Language Science: Grammar, vocabulary, reading comprehension, and language-related reasoning.

The dataset includes text, image, and no-context questions, together with ground-truth answers, lectures, and explanations.

RAD Augmentation

Following the RAD paradigm proposed in MIND, ScienceQA-RAD extends the original dataset with:

• Positive Rationales (R⁺): Diverse, semantically consistent reasoning chains that preserve correct reasoning logic and conclusions.
• Hard Negative Rationales (R⁻): Plausible but logically incorrect reasoning chains designed to expose challenging reasoning errors.

Through large-model-based rationale generation, splitting, and cleaning, the dataset achieves approximately 1000× rationale augmentation compared with the original ScienceQA.

These augmented rationales provide rich supervision for multi-rationale learning, reasoning discrimination, and logical correction, supporting the Understand → Rethink → Correct paradigm introduced in MIND.

Related Resources

• MIND Paper (ICML 2026): Multi-rationale INtegrated Discriminative Reasoning Framework for Multi-modal Large Models
• Code: MIND GitHub Repository
• Original Dataset: ScienceQA
• Related Dataset: M3CoT-RAD-Version
• Related Dataset: VersaCoT-10K-RAD-Version
• Related Dataset: A-OKVQA-RAD-Version

Citation

If you find this dataset useful in your research, please cite both MIND and the original ScienceQA paper.

MIND (ICML 2026) — Paper | Google Scholar

@inproceedings{pmlr-v306-yu26bv,
  title={{MIND}: Multi-rationale {IN}tegrated Discriminative Reasoning Framework for Multi-modal Large Models},
  author={Yu, Chuang and Zhao, Jinmiao and Zhao, Mingxuan and Liu, Yunpeng and Shu, Xiujun and Feng, Yuanhao and Wang, Bo and Yue, Xiangyu},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  pages={151737--151756},
  year={2026},
  volume={306},
  series={Proceedings of Machine Learning Research},
  publisher={PMLR},
  url={https://proceedings.mlr.press/v306/yu26bv.html}
}

ScienceQA (NeurIPS 2022) — Paper | Google Scholar

@inproceedings{lu2022learn,
  title={Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering},
  author={Lu, Pan and Mishra, Swaroop and Xia, Tony and Qiu, Liang and Chang, Kai-Wei and Zhu, Song-Chun and Tafjord, Oyvind and Clark, Peter and Kalyan, Ashwin},
  booktitle={Advances in Neural Information Processing Systems},
  volume={35},
  pages={2507--2521},
  year={2022}
}