MIRACL-VISION is a multilingual visual retrieval dataset released by NVIDIA in 2025, with the related research paper titled "MIRACL-VISION: A Large, multilingual, visual document retrieval benchmark", designed to evaluate multilingual, multimodal retrieval pipelines.

The dataset contains 7,898 user queries and 338,734 Wikipedia article images with corresponding annotations, supporting 18 languages including Arabic, Bengali, English, Spanish, Persian, Finnish, French, Hindi, Indonesian, Japanese, Korean, Russian, Swahili, Telugu, Thai, Chinese, and Yoruba.

Dataset Composition

• Language Subsets: The dataset covers 18 languages, with each language corresponding to an independent configuration (config) and data files.

• Queries: Stores user query text, in Parquet format.

• Corpus: Stores Wikipedia articles and their associated image data, in Parquet format.

• Qrels: Stores relevance information between queries and documents, in Parquet format.

• Images: Stores image data associated with documents, in Parquet format.

Citation

@misc{osmulsk2025miraclvisionlargemultilingualvisual,
      title={MIRACL-VISION: A Large, multilingual, visual document retrieval benchmark}, 
      author={Radek Osmulsk and Gabriel de Souza P. Moreira and Ronay Ak and Mengyao Xu and Benedikt Schifferer and Even Oldridge},
      year={2025},
      eprint={2505.11651},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2505.11651}, 
}