Datasets:

glowol
/

RealXBench

license: apache-2.0
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - vision
  - question-answering
  - multimodal
size_categories:
  - 1K<n<10K

RealXBench

RealXBench is a comprehensive visual question answering benchmark dataset. The full dataset contains 300 high-quality image-question-answer triplets. Due to internal regulations, only a subset of 194 samples is released in this open-source version.

Dataset Structure

Each example contains:

query: The question about the image (in English)
answer: The ground truth answer(s), with multiple answers separated by "or"
perception: Difficulty level for perception task (1 if required, 0 otherwise)
search: Difficulty level for search task (1 if required, 0 otherwise)
reason: Difficulty level for reasoning task (1 if required, 0 otherwise)
image: The corresponding image file

Usage

from datasets import load_dataset

dataset = load_dataset("glowol/RealXBench")

Citation

If you use this dataset, please cite:

@article{deepEyesV2,
  title={DeepEyesV2: Toward Agentic Multimodal Model},
  author={Jack Hong and Chenxiao Zhao and ChengLin Zhu and Weiheng Lu and Guohai Xu and Xing Yu},
  journal={arXiv preprint arXiv:2511.05271},
  year={2025},
  url={https://arxiv.org/abs/2511.05271}
}