lbourdois/VQA-neulab-CulturalGround-clean

Description

French part of the neulab/CulturalGround dataset that we processed to display images directly as a PIL object, and questions & answers as individual columns.

The dataset contains images from 42 countries from Wikidata.
For each country, two types of questions (generated via Qwen/Qwen2.5-VL-72B-Instruct according this this file) are possible:

• Open-Ended VQA (OE splits), i.e. the model answers directly from the image
• Multiple-Choices VQA (MCQs splits), i.e. the model answers MCQ or True/False questions on facts related to the image.

For these two types of questions, the original dataset has two categories: raw questions/answers, and filtered questions/answers (i.e. high quality data keep by the authors).
In the present dataset, we have kept only the filtered data hence the names OE_Filtered and MCQs_Filtered.

Citation

@misc{nyandwi2025groundingmultilingualmultimodalllms,
      title={Grounding Multilingual Multimodal LLMs With Cultural Knowledge}, 
      author={Jean de Dieu Nyandwi and Yueqi Song and Simran Khanuja and Graham Neubig},
      year={2025},
      eprint={2508.07414},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.07414}, 
}