Dataset · afaji
Multilingual culturally-aware VQA benchmark
afaji/cvqa
Over 10,000 multilingual image-question pairs with cultural context for evaluating multimodal models.
What's inside
The dataset contains over 10,000 image-question pairs (from 10K to 100K entries), covering 39 language-country pairs and 10 thematic categories. Questions are written in native languages (including Indonesian, Japanese, Sundanese, Javanese, Minangkabau, Breton) as well as English. The set was created by MBZUAI researchers and is intended for use as a test split.
What you can do with it
- Evaluating multimodal models (VLM) on cultural context understanding
- Checking visual question answering (VQA) quality in low-resource languages
- Tasks
- Question answering
- Languages
- Indonesian, Sundanese, Japanese, Javanese, Minangkabau, Breton, Irish, Spanish and 4 more
- Size
- 10K–100K rows
- Downloads this month
- 2K
- Likes
- 40
How to load it
from datasets import load_dataset
ds = load_dataset("afaji/cvqa")