Dataset · ByteDance
Multilingual Text-Image VQA (9 Languages)
ByteDance/MTVQA
Image-text question-answer pairs in nine languages, including Russian, for evaluating and training multimodal models.
What's inside
Inside are images with text and questions about them, annotated by native speakers according to strict rules. The dataset covers 9 languages (including Russian, Korean, Japanese, German, Arabic, and others) and contains between 1 and 10 thousand examples. Collected by the ByteDance team specifically for visual-question-answering and image-to-text tasks.
What you can do with it
- Evaluating multimodal LLM quality in text recognition and understanding on images (OCR + VQA)
- Fine-tuning vision-language models on multilingual text scenes, including Russian
- Tasks
- Visual question answering, Image → text
- Languages
- Multilingual, Arabic, German, Vietnamese, Japanese, Korean, French, Russian and 2 more
- Size
- 1K–10K rows
- Downloads this month
- 2.2K
- Likes
- 42
How to load it
from datasets import load_dataset
ds = load_dataset("ByteDance/MTVQA")