README.md

metadata

language:
  - ar
pretty_name: 'COCO-ARVQA: Arabic Visual Question Answering over COCO 2017'
license: cc-by-4.0
task_categories:
  - visual-question-answering
  - image-to-text
tags:
  - arabic
  - vqa
  - visual-question-answering
  - coco
  - coco2017
  - multimodal
  - vision-language
  - question-answering
  - low-resource-language
  - arabic-nlp
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/validation.jsonl

COCO-ARVQA: Arabic Visual Question Answering over COCO 2017

Dataset Summary

COCO-ARVQA is an Arabic Visual Question Answering dataset built over images from MS COCO 2017 train2017.
It provides Arabic questions, Arabic answers, answer lists, question identifiers, image identifiers, and COCO image file names.

This repository does not redistribute COCO images. Both the training and validation splits reference images from the official COCO 2017 train2017.zip archive.

Official COCO image archive used by this dataset:

http://images.cocodataset.org/zips/train2017.zip

Dataset Structure

coco-arvqa/
├── README.md
├── CITATION.cff
├── LICENSE
├── dataset_stats.json
├── data/
│   ├── train.jsonl
│   └── validation.jsonl
└── assets/
    ├── split_examples.png
    ├── unique_images.png
    ├── answer_type_distribution.png
    └── top_question_types_train.png

Splits

Split	Examples	Unique Images	Image Source
Train	93,645	49,894	COCO 2017 train2017
Validation	10,361	5,544	COCO 2017 train2017
Total	104,006	55,438	COCO 2017 train2017

Dataset Visualizations

Examples per Split

Unique Images per Split

Answer Type Distribution

Top Question Types in the Train Split

Answer Type Distribution

Answer Type	Train	Validation	Total
yes/no	36,157	3,925	40,082
number	26,591	2,921	29,512
other	30,897	3,515	34,412

Dataset Examples

The following examples illustrate the structure of COCO-ARVQA. Images are not redistributed in this repository. Each example can be matched with its COCO image using the image_file_name field after downloading the official COCO 2017 train2017.zip archive.

Official COCO image archive:

http://images.cocodataset.org/zips/train2017.zip

Training Examples

Split	Image File	Arabic Question	Arabic Answer	Answer Type
train	000000458752.jpg	ما لون قميص اللاعبين؟	برتقالي	other
train	000000262146.jpg	ما لون الثلج؟	أبيض	other
train	000000262146.jpg	ماذا يفعل الشخص؟	التزلج	other
train	000000262146.jpg	ما لون غطاء رأس الشخص؟	أحمر	other
train	000000524291.jpg	ماذا يحمل الشخص في يده؟	قرص طائر	other
train	000000524291.jpg	هل الكلب ينتظر؟	نعم	yes/no

Validation Examples

Split	Image File	Arabic Question	Arabic Answer	Answer Type
validation	000000393227.jpg	هل لدى الرجل وشم؟	نعم	yes/no
validation	000000393227.jpg	على ماذا يركب هذا الرجل؟	لوح تزلج	other
validation	000000393227.jpg	كم وشمًا يمكن رؤيته على جسم هذا الرجل؟	1	number
validation	000000131127.jpg	هل الرجل غير سعيد؟	لا	yes/no

Example JSON Record

{ "task": "vqa_ar", "question_id": 458752002, "image_id": 458752, "image_file_name": "000000458752.jpg", "coco_split": "train2017", "coco_images_zip": "http://images.cocodataset.org/zips/train2017.zip", "question_ar": "ما لون قميص اللاعبين؟", "prompt": "أجب عن السؤال بالعربية: ما لون قميص اللاعبين؟", "answer_ar": "برتقالي", "multiple_choice_answer_ar": "برتقالي", "answers_ar": [ "برتقالي", "برتقالي", "برتقالي", "برتقالي", "برتقالي", "برتقالي", "برتقالي", "برتقالي", "برتقالي", "برتقالي" ], "question_type_en": "what color is the", "answer_type": "other", "source_dataset": "MS COCO 2017 train2017", "dataset_name": "COCO-ARVQA" }

A small preview file is also provided at:

examples/sample_examples.jsonl

Data Fields

Each JSONL example contains the following fields:

Field	Type	Description
task	string	Task name, usually vqa_ar.
question_id	integer/string	Question identifier.
image_id	integer/string	COCO image identifier.
image_file_name	string	File name of the image inside COCO train2017.
coco_split	string	COCO image split used by the dataset, here train2017.
coco_images_zip	string	Official COCO archive URL for the images.
question_ar	string	Arabic visual question.
prompt	string	Instruction-style Arabic prompt.
answer_ar	string	Main Arabic answer.
multiple_choice_answer_ar	string	Arabic multiple-choice-style answer, when available.
answers_ar	list[string]	List of Arabic answer annotations.
question_type_en	string	Original or inherited English question type category.
answer_type	string	Answer category: yes/no, number, or other.
source_dataset	string	Image source dataset.
dataset_name	string	Dataset name.

Example

{
  "task": "vqa_ar",
  "question_id": 458752002,
  "image_id": 458752,
  "image_file_name": "000000458752.jpg",
  "coco_split": "train2017",
  "coco_images_zip": "http://images.cocodataset.org/zips/train2017.zip",
  "question_ar": "ما لون قميص اللاعبين؟",
  "prompt": "أجب عن السؤال بالعربية: ما لون قميص اللاعبين؟",
  "answer_ar": "برتقالي",
  "multiple_choice_answer_ar": "برتقالي",
  "answers_ar": ["برتقالي", "برتقالي"],
  "question_type_en": "what color is the",
  "answer_type": "other"
}

How to Use

Load the annotations from Hugging Face

from datasets import load_dataset

ds = load_dataset("MouaffakAyoub/coco-arvqa")
print(ds)
print(ds["train"][0])

Download and match COCO images

Download and extract the official COCO 2017 train images:

wget http://images.cocodataset.org/zips/train2017.zip
unzip train2017.zip

Then match each example with its image:

from pathlib import Path

image_root = Path("train2017")
example = ds["train"][0]
image_path = image_root / example["image_file_name"]

print(image_path)

Dataset Creation

The dataset was created for Arabic Visual Question Answering research.
It links Arabic question-answer annotations to COCO 2017 image file names. The training and validation splits are internal splits of COCO-ARVQA; both reference images from COCO 2017 train2017.

Intended Uses

This dataset is intended for:

• Arabic Visual Question Answering.
• Multilingual and Arabic vision-language modeling.
• Parameter-efficient adaptation of multimodal models.
• Low-resource multimodal learning research.
• Evaluation of Arabic VQA systems.

Out-of-Scope Uses

This dataset is not intended for:

• Biometric identification or surveillance.
• Face recognition or person identification.
• Legal, medical, or safety-critical decision-making.
• Commercial use of COCO images without checking the original image licenses.
• Any use that violates COCO, Flickr, or original image-owner terms.

Limitations

• Images are not included in this repository.
• Both train and validation examples reference COCO 2017 train2017.
• Arabic questions and answers may contain translation or generation noise.
• The dataset may inherit visual, cultural, and social biases from COCO images and from the automatic generation/translation process.
• Some answers are short, such as colors, numbers, or yes/no responses.

Licensing and Copyright

Arabic VQA Annotations

The Arabic question-answer annotations in this repository are released under CC BY 4.0, unless otherwise specified by the dataset authors. Users should cite the dataset and preserve attribution.

COCO Images

COCO images are not redistributed in this repository. Users must download the images from the official COCO archive and comply with the original image licenses and Flickr terms.

COCO Annotations

The original COCO annotation set belongs to the COCO Consortium and is commonly distributed under Creative Commons Attribution 4.0. COCO images themselves follow Flickr/image-owner licensing terms.

Citation

If you use this dataset, please cite it as:

@dataset{mouaffak2026cocoarvqa,
  title  = {COCO-ARVQA: Arabic Visual Question Answering over COCO 2017},
  author = {Mouaffak, Ayoub},
  year   = {2026},
  url    = {https://huggingface.co/datasets/MouaffakAyoub/coco-arvqa},
  note   = {Arabic Visual Question Answering annotations over COCO 2017 train2017 images}
}

Acknowledgements

This dataset builds on MS COCO 2017 image identifiers and image file names. Users should also cite the Microsoft COCO paper when using the associated images.

Contact

For questions, corrections, or citation updates, contact the dataset maintainer: ayoubmouaffak@gmail.com.