Skip to content

Dataset · neulab

Multilingual multimodal instruction dataset

neulab/PangeaInstruct

1-10 million question-answer and VQA pairs in 39 languages for training multimodal LLMs.

What's inside

PangeaInstruct is a multilingual multimodal dataset for visual-question-answering and question-answering. The metadata indicates a size of 1 to 10 million examples and 39 languages. The card describes it as part of the open Pangea project featuring the Pangea-7B model and PangeaBench benchmark; the format is not specified in the visible part.

What you can do with it

  • Training the Pangea-7B multimodal LLM
  • LoRA for VQA in low-resource languages
  • Evaluating multilingual VQA
Tasks
Visual question answering, Question answering
Languages
Amharic, Arabic, Bulgarian, Bangla, Czech, German, Greek, English and 4 more
Size
1M–10M rows
Downloads this month
1.6K
Likes
88

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("neulab/PangeaInstruct")