M4-RAG: A Massive-Scale Multilingual Multi-Cultural Multimodal RAG

Paper | Code

M4-RAG is a massive-scale benchmark spanning 42 languages, 56 regional dialects and registers, and 189 countries, comprising over 80,000 culturally diverse image-question pairs for evaluating retrieval-augmented Visual Question Answering (VQA) across languages and modalities.

This repository specifically contains the Wikipedia Retrieval Corpus, a controlled retrieval environment containing millions of carefully curated multilingual documents relevant to the query domains.

Dataset Structure

The dataset consists of two configurations:

• cvqa: Wikipedia articles relevant to the Culturally-Aware Visual Question Answering domain.
• worldcuisines: Wikipedia articles relevant to the food-related visual question answering domain.

Sample Usage

You can load the retrieval corpus using the Hugging Face datasets library:

from datasets import load_dataset

# Load the CVQA Wikipedia retrieval corpus
cvqa_corpus = load_dataset("davidanugraha/M4-RAG", "cvqa", split="train")

# Load the WorldCuisines Wikipedia retrieval corpus
worldcuisines_corpus = load_dataset("davidanugraha/M4-RAG", "worldcuisines", split="train")

Related Datasets

• CVQA Images: Available at davidanugraha/cvqa
• WorldCuisines Images: Available at worldcuisines/vqa-v1.1

Citation

If you use M4-RAG in your research, please cite:

@article{anugraha2025m4rag,
  title={M4-RAG: A Massive-Scale Multilingual Multi-Cultural Multimodal RAG},
  author={Anugraha, David and Irawan, Patrick Amadeus and Singh, Anshul and Lee, En-Shiun Annie and Winata, Genta Indra},
  journal={arXiv preprint arXiv:2512.05959},
  year={2025},
  url={https://arxiv.org/abs/2512.05959}
}