metadata

dataset_info:
  features:
    - name: id
      dtype: int64
    - name: multimodal_question
      dtype: string
    - name: answer
      dtype: string
    - name: rationale
      dtype: string
    - name: text_only_question
      dtype: string
    - name: image_source
      dtype: string
    - name: evidence
      dtype: string
    - name: resolution
      dtype: string
    - name: proportion_of_roi
      dtype: string
    - name: category
      dtype: string
    - name: text_in_image
      dtype: string
    - name: rationale_granularity
      dtype: string
    - name: image
      dtype: image
    - name: cropped_image
      dtype: image
  splits:
    - name: train
      num_bytes: 157153160
      num_examples: 129
  download_size: 157133331
  dataset_size: 157153160
configs:
  - config_name: default
    data_files:
      - split: train
        path: hard_data/train-*

VisualSimpleQA

Introduction

VisualSimpleQA is a multimodal fact-seeking benchmark with two key features. First, it enables streamlined and decoupled evaluation of LVLMs in visual and linguistic modalities. Second, it incorporates well-defined difficulty criteria to guide human annotation and facilitates the extraction of a challenging subset, VisualSimpleQA-hard. Experiments on 15 LVLMs show that even state-of-the-art models such as GPT-4o achieve merely 60%+ correctness in multimodal fact-seeking QA on VisualSimpleQA and 30%+ on VisualSimpleQA-hard. Furthermore, the decoupled evaluation based on this benchmark across different models highlights substantial opportunities for improvement in both visual and linguistic modules. The dataset reviewer above illustrates 129 samples from VisualSimpleQA-hard.

arXiv: https://arxiv.org/pdf/2503.06492

Data Example:

  {'id': 369,
   'multimodal_question': 'Which institution did the creator of this cartoon duck donate her natural science-related paintings to?',
   'answer': 'The Armitt Museum, Gallery, Library', 'rationale': 'Jemima Puddle-Duck',
   'text_only_question': 'Which institution did the creator of Jemima Puddle-Duck donate her natural science-related paintings to?',
   'image_source': 'https://www.gutenberg.org/files/14814/14814-h/images/15-tb.jpg',
   'evidence': 'https://www.armitt.com/beatrix-potter-exhibition/\nhttps://en.wikipedia.org/wiki/Beatrix_Potter',
   'resolution': '400x360',
   'proportion_of_roi': '0.2232',
   'category': 'research and education',
   'text_in_image': 'absence',
   'rationale_granularity': 'fine-grained',
   'image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=400x360 at 0x7FE82C270D70>,
   'cropped_image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=164x196 at 0x7FE82C329550>}

File Structure

data/ This directory contains all 500 samples of VisualSimpleQA stored in parquet files.

hard_data/ This directory contains 129 VisualSimpleQA-hard samples stored in a parquet file. These samples are selected based on well-defined criteria to ensure they represent more challenging cases from VisualSimpleQA.

Disclaimer

This dataset contains images collected from various sources. The authors do NOT claim ownership or copyright over the images. The images may be subject to third-party rights, and users are solely responsible for verifying the legal status of any content before use.

Intended Use: The images are provided for non-commercial research purposes only.
Redistribution Prohibition: You may NOT redistribute or modify the images without permission from original rights holders.
Reporting Violations: If you encounter any sample potentially breaching copyright or licensing rules, contact us at [email protected]. Verified violations will be removed promptly.

The authors disclaim all liability for copyright infringement or misuse arising from the use of this dataset. Users assume full legal responsibility for their actions.

License

Text Data: Licensed under CC-BY-4.0
Images: Subject to custom terms (see Disclaimer above).

Citation

BibTeX:

@article{wang2025visualsimpleqa,
  title={VisualSimpleQA: A Benchmark for Decoupled Evaluation of Large Vision-Language Models in Fact-Seeking Question Answering},
  author={Yanling Wang and Yihan Zhao and Xiaodong Chen and Shasha Guo and Lixin Liu and Haoyang Li and Yong Xiao and Jing Zhang and Qi Li and Ke Xu},
  journal={arXiv preprint arXiv: 2503.06492},
  year={2025}
}