Datasheet for COCO-ARVQA
Motivation
The dataset was created to support Arabic Visual Question Answering and multilingual multimodal research.
Composition
- Total examples: 104,006
- Train examples: 93,645
- Validation examples: 10,361
- Total unique referenced images: 55,438
- Image source: COCO 2017 train2017
- Images included in this repository: No
Collection and Processing
The dataset links Arabic question-answer annotations to COCO image file names. Users must download COCO 2017 train2017 images separately.
Recommended Uses
Arabic VQA, multimodal learning, Arabic vision-language modeling, and parameter-efficient adaptation studies.
Licensing
Arabic annotations: CC BY 4.0. COCO images: not redistributed; users must comply with original COCO/Flickr image terms.