VisCoR-55K

# VisCoR-55K Dataset VisCoR-55K is a high-quality dataset for visual reasoning, spanning five categories: **General, Reasoning, Math, Graph/Chart, and OCR**. This release contains three components: - **VQA Samples**: Original visual question-answer pairs. - **Contrastive Counterparts**: Matched contrastive VQA pairs constructed to encourage faithful reasoning. - **Generated Rationales**: High-quality rationales synthesized by our **VC-STaR** framework. (**This dataset**) We hope this resource will facilitate future research on faithful and robust visual reasoning in vision-language models. - **GitHub**: https://github.com/zhiyupan42/VC-STaR - **Paper**: https://arxiv.org/abs/2603.02556 📝 Citation If you find our work useful for your research, please consider citing our paper: ``` @inproceedings{pan2026through, title={Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs}, author={Pan, Zhiyu and Wu, Yizheng and Hua, Jiasheng and Feng, Junyi and Yan, Shaotian and Deng, Bing and Cao, Zhiguo and Ye, Jieping}, booktitle={The Fourteenth International Conference on Learning Representations}, year={2026} } ```