This is the VQA-Verify dataset, introduced in the paper SATORI-R1: Incentivizing Multimodal Reasoning with Spatial Grounding and Verifiable Rewards.
Arxiv Here | Github
VQA-Verify is a 12k dataset annotated with answer-aligned captions and bounding boxes. It's designed to facilitate training models for Visual Question Answering (VQA) tasks, particularly those employing free-form reasoning. The dataset addresses limitations in existing VQA datasets by providing verifiable intermediate steps and… See the full description on the dataset page:
https://huggingface.co/datasets/justairr/VQA-Verify.
