jaagli/ravenea

Dataset Card for Dataset Name

RAVENEA: A Benchmark for Multimodal Retrieval-Augmented Visual Culture Understanding

Dataset Description

RAVENEA (Retrieval-Augmented Visual culturE uNdErstAnding) is a benchmark designed to advance visual culture understanding through retrieval-augmented methods. It focuses on two tasks: culture-focused visual question answering (cVQA) and culture-informed image captioning (cIC). The dataset integrates over 10,000 Wikipedia documents curated and ranked by human annotators, aiming to enhance the cultural comprehension capabilities of vision-language models (VLMs).

Dataset Sources [optional]

• Repository: RAVENEA Code
• Paper: RAVENEA: A Benchmark for Multimodal Retrieval-Augmented Visual Culture Understanding
• Project Page: RAVENEA

Uses

RAVENEA is intended for evaluating and enhancing the performance of vision-language models in tasks requiring cultural understanding, specifically through culture-focused visual question answering and culture-informed image captioning

Direct Use

from datasets import load_dataset

data = load_dataset("jaagli/ravenea", split="combination")

Dataset Structure

Overview of the dataset.

Dataset({
    features: ['query_id', 'img_id', 'image', 'country', 'task_type', 'doc_ids', 'doc_urls', 'culture_relevance_scores', 'questions', 'options', 'answers', 'human_captions', 'category'],
    num_rows: 1868
})

An example from cIC task:

{'query_id': 'ccub_0',
'img_id': 'China_10',
'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=899x600 at 0x7F8B09469550>,
'country': 'China',
'task_type': 'cIC',
'doc_ids': ['38201816', '1681878', '1590884', '19422108', '10979580', '1364408', '30087328', '21678136', '27343349', '8197312'],
'doc_urls': ['https://en.wikipedia.org/wiki/Wangu%20Pavilion', 'https://en.wikipedia.org/wiki/Marble%20Boat', 'https://en.wikipedia.org/wiki/Chinese%20garden', 'https://en.wikipedia.org/wiki/Dunedin%20Chinese%20Garden', 'https://en.wikipedia.org/wiki/Chinese%20Garden%2C%20Singapore', 'https://en.wikipedia.org/wiki/Paifang', 'https://en.wikipedia.org/wiki/China%20pavilion%20at%20Expo%202010', 'https://en.wikipedia.org/wiki/Dule%20Temple', 'https://en.wikipedia.org/wiki/Expo%202010%20pavilions', 'https://en.wikipedia.org/wiki/Chinese%20Pavilion%20at%20Drottningholm'], 'culture_relevance_scores': [1, 1, 0, -1, -1, -1, -1, -1, -1, -3],
'questions': [],
'options': [],
'answers': [],
'human_captions': 'the Chinese pavilion by the riverside',
'category': 'architecture'}

Source Data

Wikipedia, CVQA, CCUB

Licensing Information

We adhere to the licensing terms derived from the source data.

BibTeX Citation

@misc{li2025raveneabenchmarkmultimodalretrievalaugmented,
      title={RAVENEA: A Benchmark for Multimodal Retrieval-Augmented Visual Culture Understanding}, 
      author={Jiaang Li and Yifei Yuan and Wenyan Li and Mohammad Aliannejadi and Daniel Hershcovich and Anders Søgaard and Ivan Vulić and Wenxuan Zhang and Paul Pu Liang and Yang Deng and Serge Belongie},
      year={2025},
      eprint={2505.14462},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.14462}, 
}