Datasets:

QCRI
/

ImageEval-ArabicNLP26

license: cc-by-nc-4.0
language:
  - ar
  - en
pretty_name: ImageEval-ArabicNLP26
tags:
  - multimodal
  - arabic
  - visual-question-answering
  - hallucination-detection
  - speech
  - culture
  - text-to-image
  - evaluation
configs:
  - config_name: task1a_en
    data_files:
      - split: train
        path: task1a/train_en.jsonl
      - split: dev
        path: task1a/dev_en.jsonl
      - split: devtest
        path: task1a/devtest_en.jsonl
      - split: test
        path: task1a/test_en.jsonl
  - config_name: task1a_msa
    data_files:
      - split: train
        path: task1a/train_msa.jsonl
      - split: dev
        path: task1a/dev_msa.jsonl
      - split: devtest
        path: task1a/devtest_msa.jsonl
      - split: test
        path: task1a/test_msa.jsonl
  - config_name: task1b_en
    data_files:
      - split: train
        path: task1b/train_en.jsonl
      - split: dev
        path: task1b/dev_en.jsonl
      - split: devtest
        path: task1b/devtest_en.jsonl
      - split: test
        path: task1b/test_en.jsonl
  - config_name: task1b_msa
    data_files:
      - split: train
        path: task1b/train_msa.jsonl
      - split: dev
        path: task1b/dev_msa.jsonl
      - split: devtest
        path: task1b/devtest_msa.jsonl
      - split: test
        path: task1b/test_msa.jsonl
  - config_name: task2
    data_files:
      - split: train
        path: task2/train.jsonl
      - split: dev
        path: task2/dev.jsonl
      - split: test
        path: task2/test.jsonl

ImageEval-ArabicNLP26 👁️

ImageEval-ArabicNLP26 is the dataset of the ImageEval 2026 Shared Task at ArabicNLP 2026. It covers both of the shared task's tasks: AynVQA (Task 1), a culturally grounded Arabic multimodal benchmark for spoken visual question answering and hallucination detection, and CRAI-Bench (Task 2), which evaluates the cultural accuracy of Arabic text-to-image generation.

The shared task has concluded. All gold labels are released, including the blind test splits, and the official results are on the leaderboard.

💬 Join our Slack

Join the ImageEval Slack channel for announcements about data releases, deadlines, and updates, and to connect with the organisers and other participants.

🎯 Tasks

Spoken VQA (Task 1a). Given an image and the spoken question and options (audio), choose the correct option.

Prediction: the option index 0, 1 or 2.

Hallucination detection (Task 1b). Given an image and three statements, decide for each statement whether it is True (grounded in the image) or False (a hallucination). Exactly one statement is grounded.

Prediction: a True/False label per statement.

Cultural accuracy evaluation (Task 2). Given a reference image of a Qatari cultural scene, a caption, and an AI-generated image produced from that caption, judge how faithfully the generated image represents the culture.

Prediction: CRAI scores across five dimensions plus a composite.

👁️ Task 1: AynVQA

Ayn (عين, "eye") tests whether a model can read a culturally specific image, both from a spoken Arabic question and by telling grounded descriptions apart from plausible but hallucinated ones. Each subtask is offered as two language tracks, English and Modern Standard Arabic (MSA), scored separately.

Starter kit, format checker and official scorer: ImageEval2026-tasks/task1.

🗂️ Subsets

config	task	language	Codabench
`task1a_en`	Spoken VQA	English	compete
`task1a_msa`	Spoken VQA	MSA	compete
`task1b_en`	Hallucination	English	compete
`task1b_msa`	Hallucination	MSA	compete

The English and MSA tracks of a task are parallel: same images, same answers, and the questions are translations of each other.

🌍 Countries

Task 1 spans 18 Arab countries (Task 2 is grounded specifically in Qatari culture):

Algeria, Bahrain, Egypt, Iraq, Jordan, Kuwait, Lebanon, Libya, Morocco, Oman, Palestine, Qatar, Saudi Arabia, Sudan, Syria, Tunisia, UAE, Yemen.

🔊 Audio

The Task 1a questions in train, dev and devtest are synthetically generated using voice cloning (TTS). The questions in the final blind test set will be human-recorded; expect a speaker/recording-condition shift between the dev-phase audio and the test audio.

📂 Files

images/<id>.jpg            one image per item, shared across tasks and languages
audio/<lang>/<id>.wav      spoken question and options (Task 1a)
task1a/<split>_<lang>.jsonl
task1b/<split>_<lang>.jsonl

Media is referenced by relative path keyed on id, so inputs join to files directly.

🏷️ Fields

Task 1a (task1a/<split>_<lang>.jsonl):

field	type	description
`id`	str	item id
`image`	str	`images/<id>.jpg`
`audio`	str	`audio/<lang>/<id>.wav`, the spoken question and the three options (no text is given; listen and answer)
`label`	int	index (0–2) of the correct option

Task 1b (task1b/<split>_<lang>.jsonl):

field	type	description
`id`	str	item id
`image`	str	`images/<id>.jpg`
`statements`	list[str]	three statements, exactly one grounded
`labels`	list[bool]	truth value of each statement (one `true`)

Every split, including devtest and test, now carries the full schema: labels plus country, category and subcategory. The competition has ended and the previously blind splits are fully labelled.

📊 Splits

split	labels	items	use
`train`	yes	3000	training and fine-tuning
`dev`	yes	500	validation
`devtest`	yes	500	development-phase evaluation (labels now released)
`test`	yes	1000	official competition ranking (labels now released)

📏 Evaluation

Each of the four tracks is scored separately. One ranking metric decides the leaderboard; the remaining columns are reported as diagnostics. A missing or unparseable prediction always counts as wrong.

Spoken VQA

A 3-way multiple choice: predict the option index 0, 1 or 2.

metric	role	meaning
Accuracy	ranking	fraction of items answered correctly
Balanced accuracy	reported	mean per-class recall over the three positions
Macro-F1	reported	macro-averaged F1 over the three positions

Hallucination detection

Three statements per image; predict True/False for each. Exactly one statement is grounded ("Q+"); the other two are hallucinated ("Q−").

metric	role	meaning
Contrastive Instability (CI)	ranking	of the items with at least one of the three statements correct, the fraction that are not fully correct (all three right); lower is better
Combined accuracy	reported	fraction of items where all three labels are correct (grounded → true, both hallucinated → false); higher is better
CFHR	reported	of items where the grounded statement was correctly identified, the fraction that still affirmed a hallucinated one; lower is better
Q+ accuracy	reported	grounded statement correctly marked true
Q− accuracy	reported	hallucinated statements correctly marked false (over all false statements)

True/False is read from the prediction with the shared‑task evaluate_tf parser, which handles English and Arabic verdicts (e.g. true/false, صح/خطأ).

🧪 Baselines & example notebooks

Starter Colab notebooks that run end to end (download the data → run a model → write a Codabench-ready prediction.zip) are here:

Reference notebooks (Google Drive)

Open-model baseline: Task 1a (Qwen2.5-Omni) and Task 1b (Qwen2.5-VL), runnable on a free Colab T4 (4-bit).
Cascaded API example (Task 1a): Fanar Aura-STT (speech → text) → Oryx (image understanding); no GPU required.

These are references only; participants are free to use their own models, prompts, and configurations.

Reference baseline scores (greedy decoding, do_sample=False, so deterministic). Contrastive instability is lower-is-better:

task	track	model	ranking metric	score
1a	English	Qwen2.5-Omni	accuracy	0.6640
1a	MSA	Qwen2.5-Omni	accuracy	0.3980
1b	English	Qwen2.5-VL	contrastive instability	0.3133
1b	MSA	Qwen2.5-VL	contrastive instability	0.4900

For 1b, combined accuracy was 0.6840 (English) and 0.5080 (MSA).

🎨 Task 2: CRAI-Bench

CRAI-Bench evaluates whether AI-generated images faithfully represent Qatari and Arab cultural scenes. Given a reference image of an authentic cultural scene, a caption, and an AI-generated image produced from that caption, systems predict scores on the Cultural Representation Accuracy Index (CRAI), a five-dimensional framework validated against human annotation. Each of the reference images has five caption versions, ranging from fully Qatari-specific (v1) to entirely generic (v5). See the Task 2 page for the full framework and scoring rubric.

Starter kit, format checker, baselines and the official scorer: ImageEval2026-tasks/task2.

📂 Files

One jsonl per split; each row is one instance with its caption, image paths and the human gold scores:

task2/<split>.jsonl                  id, image_id, version, caption, ref_image, gen_image, category, CRAI_* scores
task2/<split>/imgs/ref/              reference images (one per image_id)
task2/<split>/imgs/generated/        generated images (one per id)

📊 Splits

split	reference images	instances (5 caption versions each)	gold
`train`	24	120	yes
`dev`	8	40	yes
`test`	8	40	yes (released after the competition)

📏 Evaluation

metric	role	meaning
Spearman	ranking	correlation between predicted and gold `CRAI_composite`
MAE	tiebreaker	mean absolute error on `CRAI_composite` (lower is better)

🧪 Baselines

From the starter kit, on dev:

system	Spearman	MAE
Mean baseline	N/A	0.3450
GPT-4 LLM-as-a-judge	0.6250	0.2248

📜 License and contact

The dataset is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Website: https://imageeval2026.github.io/
Slack: https://join.slack.com/t/mm-eval/shared_invite/zt-44zii4e53-IiZZ3O73Za9XL4umb9mjUQ
Contact: imageeval2026@gmail.com

Citation

The shared task overview paper covers both tasks. If you use the Task 1 data, please also cite the dataset papers below.

@inproceedings{imageeval-2026,
    title = {{ImageEval 2026}: Culturally Grounded {A}rabic Multimodal Evaluation},
    author = {Abdaljalil, Samir and
              Bhatti, Hunzalah Hassan and
              Bashiti, Ahlam and
              Amir, Farina and
              Hasan, Md Arid and
              Mousi, Basel and
              Durrani, Nadir and
              Dalvi, Fahim and              
              Sheikh Ali, Zien and
              Serpedin, Erchin and
              Kurban, Hasan and
              Jarrar, Mustafa and              
              Chowdhury, Shammur Absar and
              Alam, Firoj},
    booktitle = {Proceedings of the Fourth Arabic Natural Language Processing Conference: Shared Tasks},
    month = oct,
    year = {2026},
    address = {Budapest, Hungary},
    publisher = {Association for Computational Linguistics}
}

@article{alam2025everydaymmqa,
  title = {{OASIS}: A Multilingual and Multimodal Dataset for Culturally Grounded Spoken Visual QA},
  author = {Alam, Firoj and Shahroor, Ali Ezzat and Hasan, Md. Arid and Ali, Zien Sheikh and Bhatti, Hunzalah Hassan and Kmainasi, Mohamed Bayan and Chowdhury, Shammur Absar and Mousi, Basel and Dalvi, Fahim and Durrani, Nadir and Milic-Frayling, Natasa},
  journal = {arXiv preprint arXiv:2510.06371},
  year = {2025},
}

@inproceedings{mousi-etal-2026-correct,
    title = "Once Correct, Still Wrong: Counterfactual Hallucination in Multilingual Vision-Language Models",
    author = "Mousi, Basel  and
      Dalvi, Fahim  and
      Chowdhury, Shammur Absar  and
      Alam, Firoj  and
      Durrani, Nadir",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.234/",
    doi = "10.18653/v1/2026.findings-acl.234",
    pages = "4763--4788",
    ISBN = "979-8-89176-395-1",
}


@inproceedings{mousi2026said,
  title     = {Said Aloud, Read Different: Cross-Modal Instability in Multimodal Models},
  author    = {Mousi, Basel and Dalvi, Fahim and Chowdhury, Shammur and Alam, Firoj and Durrani, Nadir},
  booktitle = {Proceedings of Interspeech 2026},
  year      = {2026},
  address   = {Sydney, Australia},
  note = {accepted}
}