---
language:
  - ar
  - multilingual
  - arb
  - ary
  - arz
  - apc
  - afb
language_details: "ary_Arab,arz_Arab"
language_bcp47:
  - ar-jo
  - ar-ae
  - ar-eg
  - ar-ma
size_categories:
- 1K<n<10K
license: mit
task_categories:
- text-generation
- image-to-text
- visual-question-answering
dataset_info:
  features:
  - name: id
    dtype: string
  - name: dialect
    dtype: string
  - name: image
    dtype: image
  - name: image_source
    dtype: string
  - name: description
    dtype: string
  - name: description_dialect
    dtype: string
  - name: question1
    dtype: string
  - name: question2
    dtype: string
  - name: question3
    dtype: string
  - name: question4
    dtype: string
  - name: question5
    dtype: string
  - name: answer1
    dtype: string
  - name: answer2
    dtype: string
  - name: answer3
    dtype: string
  - name: answer4
    dtype: string
  - name: answer5
    dtype: string
  - name: no_answer1
    dtype: bool
  - name: no_answer2
    dtype: bool
  - name: no_answer3
    dtype: bool
  - name: no_answer4
    dtype: bool
  - name: no_answer5
    dtype: bool
  splits:
  - name: train
    num_bytes: 4737581
    num_examples: 2178
  download_size: 2309353
  dataset_size: 4737581
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
tags:
- Dialectal Arabic
---


# Dataset Card for JEEM



![MBZUAI](MBZUAI.png)



📊 **Curated by**: [Toloka](http://toloka.ai), [MBZUAI](https://mbzuai.ac.ae/research-department/natural-language-processing-department/)


🌐 Language(s): Modern Standard Arabic, 🇯🇴 Jordanian dialect, 🇪🇬 Egyptian dialect, 🇦🇪 Emirati dialect, 🇲🇦 Moroccan dialect


## 🔍 Dataset Description

JEEM is a benchmark dataset designed to evaluate Vision-Language Models (VLMs) in the context of Arabic dialectal diversity. It includes one representative dialect from each major Arabic dialectal region: 🇯🇴 **J**ordanian (Levantine), 🇪🇬 **E**gyptian, 🇦🇪 **E**mirati (Khaleeji), 🇲🇦 **M**oroccan (Maghrebi).

The dataset comprises two main tasks:

1. **Image Captioning** – Evaluating a model’s ability to generate captions that accurately describe images using Modern Standard Arabic (MSA) and dialectal Arabic.
2. **Visual Question Answering (VQA)** – Assessing how well a model understands and responds to questions about images while incorporating dialectal and culturally specific reasoning.



👾 **Repository**: TBD

🗞️ **Paper**: TBD

🪪 **License**: MIT

## 🖊️ Dataset Creation

![annotation-pipeline-bg.pdf](annotation-pipeline-bg.pdf)

 The JEEM’s creation is carried out in the following steps:

1. Regional-specific images are sourced manually from open-source databases.
2. Annotator A describes the image in both MSA and their dialect (image caption).
3. Annotator B formulates questions in their dialect based only on the image caption.
4. Annotator C reviews the image, caption, and questions, then provides answers in their dialect.

Image captions, questions and answers are written by an in-house team of annotators, who are native speakers of target dialects. 

## 🗒️  Dataset Structure

### Dataset Instances

Each dataset instance looks as follows:

```json
    {
        "id": "66ebab6aeab3ac803338a42d214dd523",
        "dialect": "Moroccan",
        "image": {"bytes": "...", "path": "..."},
        "image_source" : "Wiki",
        "description": "في الصورة يظهر رجل جالس على الأرض وهو يعزف على آلة موسيقية معروفة في المغرب باسم 'الجمبري'. الرجل يرتدي عباءة  حمراء مزخرفة، وعلى رأسه طربوش أحمر مزين كذلك. بجانب الرجل توجد حقيبة بنية قديمة بعض الشيء.",
        "description_dialect": "فالصورة كيبان واحد الراجل جالس فالأرض وكيعزف على آلة موسيقية لي معروفة فالمغرب بسمية 'الجمبري'، الراجل لابس واحد   الغندورة حمرة مزوقة، وفراسو طربوش حمر حتى هو مزوق. حدا الراجل كاين واحد الصاك قهوي قديم شوية.",
        "question1": "واش هذا الراجل كيعزف بوحدو ولا مع شي فرقة؟",
        "question2": "واش كينة شي حاجة فداك الصاك؟",
        "question3": "واش الراجل جالس على شي حاجة ولا ديريكت على  الأرض؟",
        "question4": "شنو اللون ديال الزواق اللي فالغندورة؟",
        "question5": "واش باين شمن مدينة تخادت فيها هاد التصويرة؟",
        "answer1": "لا الراجل كيعزف غير بوحدو",
        "answer2": "ما كيبانش شنو الداخل ديال الصاك ولكن يقدر يكون الغشا  ديال الجمبري.",
        "answer3": "داير بحال شي وسادة بيضة فوق مانطة فالموف.",
        "answer4": "الأغلبية ديال الزواق فالبيض وكاين حتى شوية ديال الصفر.",
        "answer5": "لا ولكن غالبا تكون مراكش حيت الحيط اللي مور الراجل حمر.",
        "no_answer1": False,
        "no_answer2": False,
        "no_answer3": False,
        "no_answer4": False,
        "no_answer5": False
    }
```

### Dataset Fields

| Column Name          | Description |
|----------------------|------------|
| `id`                | A unique identifier for the data instance |
| `dialect`           | The Arabic dialect used in the dataset instance |
| `image_source`      | The source of the image (Wiki / Flickr / Private) |
| `image`             | The image as a PIL Image object |
| `description`       | A description of the image in Modern Standard Arabic (MSA) |
| `description_dialect` | The same description written in a dialectal variant of Arabic |
| `question1` - `question5` | A set of five questions related to the image, phrased in Arabic dialect |
| `answer1` - `answer5` | The corresponding answers to the five questions |
| `no_answer1` - `no_answer5` | Indicators of whether a question is unanswerable |


## ⚖️ Licensing Information

The images are subject to the underlying licensing terms of Wikimedia Commons ([Wikimedia Commons Licensing](https://commons.wikimedia.org/wiki/Commons:Licensing/en)) and Flickr ([Flickr Creative Commons](https://www.flickrhelp.com/hc/en-us/articles/4404070159636-Creative-Commons)). The images from private archives, image captions, questions, and answers are distributed under the MIT license ([MIT License](https://opensource.org/license/mit)).

## 📜 Cite us


```
@misc{kadaoui2025jeemvisionlanguageunderstandingarabic,
      title={JEEM: Vision-Language Understanding in Four Arabic Dialects}, 
      author={Karima Kadaoui and Hanin Atwany and Hamdan Al-Ali and Abdelrahman Mohamed and Ali Mekky and Sergei Tilga and Natalia Fedorova and Ekaterina Artemova and Hanan Aldarmaki and Yova Kementchedjhieva},
      year={2025},
      eprint={2503.21910},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.21910}, 
}
````

## ✉️ Contact us

- Yova Kementchedjhieva ([Yova.Kementchedjhieva@mbzuai.ac.ae](mailto:Yova.Kementchedjhieva@mbzuai.ac.ae))
- Ekaterina Artemova ([Katya-Art@toloka.ai](mailto:katya-art@toloka.ai))