💬 Join the ImageEval 2026 Slack for announcements & questions Join now →
ImageEval 2026

Cultural Grounding in Arabic Multimodal Generation and Understanding

Task 1: AynVQA is a culturally grounded Arabic multimodal benchmark covering 18 Arab countries. It tests how well vision-language models understand Arab cultural imagery, and how reliably they tell what is really in an image apart from what only sounds right.

Overview

Task 1 has two subtasks, each offered in two language tracks (English and Modern Standard Arabic). That makes four separate leaderboards, and you may enter any of them independently.

Spoken VQA (1a) Hallucination Detection (1b) English track MSA track
Subtask What you are given What you predict
1a. Spoken VQA an image and a spoken question with three spoken options the index of the correct option (0, 1, or 2)
1b. Hallucination Detection an image and three statements True or False for each statement (exactly one is true)

Subtask 1a: Spoken VQA

Given an image and a spoken Arabic question with three spoken answer options, predict which option is correct. The question and the options live only in the audio, so a system has to listen and look at the same time.

Input

One JSON record per item. The image and audio paths resolve inside the HuggingFace dataset.

{
  "id": "4a405ef6...d10de6a0",
  "image": "images/4a405ef6...d10de6a0.jpg",
  "audio": "audio/en/4a405ef6...d10de6a0.wav",
  "country": "Algeria",
  "category": "Objects, Materials & Clothing",
  "label": 0
}

The audio says something like “What is shown in this image? Option 0 … Option 1 … Option 2 …“. The correct index label (0, 1, or 2) is provided for train and dev, and withheld for test.

Output

A CSV with one row per item: the predicted option index.

id,prediction
4a405ef6...d10de6a0,0
9b1c2d3e...77aa01ff,2

Evaluation

Metric Role What it measures
accuracy official ranking share of questions answered correctly
balanced accuracy secondary mean per-class recall across the three option positions
macro-F1 secondary macro-averaged F1 across the three option positions

Subtask 1b: Hallucination Detection

Given an image and three statements about it, label each statement True (grounded in the image) or False (hallucinated). In every item, exactly one statement is grounded; the other two are culturally plausible but not supported by the image.

Input

One JSON record per item.

{
  "id": "4a405ef6...d10de6a0",
  "image": "images/4a405ef6...d10de6a0.jpg",
  "statements": [
    "Islamic art is prominently featured in this image. True or False?",
    "Abstract art is prominently featured in this image. True or False?",
    "Renaissance art is prominently featured in this image. True or False?"
  ],
  "country": "Algeria",
  "category": "Objects, Materials & Clothing",
  "labels": [true, false, false]
}

The labels list is provided for train and dev, and withheld for test.

Output

A CSV with three rows per item, one per statement.

id,statement_index,prediction
4a405ef6...d10de6a0,0,true
4a405ef6...d10de6a0,1,false
4a405ef6...d10de6a0,2,false

Evaluation

The grounded statement is called Q+; the two hallucinated ones are Q−.

Metric Role What it measures
Contrastive Instability (CI) official ranking of items with at least one correct label, the share not fully correct, all three right (lower is better)
combined accuracy secondary items where all three labels are correct
CFHR secondary of items whose true statement was found, the share that still affirmed a hallucination (lower is better)
Q+ accuracy secondary the grounded statement correctly marked true
Q− accuracy secondary hallucinated statements correctly marked false

Exact formulas live in the scorer.

Dataset

The data for Task 1 is available on HuggingFace: ImageEval2026 Task1 data.

The images and audio live there; the JSONL records reference them by relative image and audio paths. Labels and metadata are included for train and dev only; devtest and the blind test split are released without labels.

Split Labels Items Use
train yes 3,000 training and fine-tuning
dev yes 500 local validation (run the scorer here)
devtest no 500 development-phase leaderboard
test no 1,000 blind final ranking

The train and dev records also carry country (one of 18 Arab countries) and a cultural category, so you can break results down by region and topic.

For Subtask 1a, the train, dev, and devtest audio is synthetic (voice cloning); the blind test audio is human-recorded, so expect a shift in speaker and recording conditions.

Starter Kit

The starter kit ships three ready-to-run Colab notebooks that download the data, run a model, and write a submission-ready file. Set the LANG flag (en or msa) and run. Each link below opens the notebook straight in Colab, with no need to browse the repository.

The first two are reference baselines with published scores. The third is a no-GPU cascaded example that shows an alternative system design (speech-to-text into an image model).

Subtask Notebook Score (devtest) Colab
1a Qwen2.5-Omni-3B baseline (image + audio) accuracy: EN 0.664, MSA 0.398 Open in Colab
1b Qwen2.5-VL-3B baseline (True/False per statement) CI (lower=better): EN 0.3133, MSA 0.4900 Open in Colab
1a Fanar cascade example (no GPU) not scored Open in Colab

Participation

You may use any model and any pipeline, cascaded or end to end. For Subtask 1a, you might transcribe the audio and pass the text to a vision-language model, or use a single multimodal model that reads the audio and image directly. We encourage creative approaches: data augmentation, model merging, new cascade designs (for example, speech recognition then translation then reasoning), and novel prompting are all welcome.

Open and closed models are both allowed. We especially encourage open models, since they suit low-resource settings and make results easier to reproduce.

Please take part in good faith and run each track end to end in its own language. Because the English and MSA tracks share the same images and answers (the questions are translations of each other), it is technically possible to copy one track’s predictions onto the other. Please do not. Build a system that actually processes each track’s own audio and statements. A pipeline that goes through translation is welcome, as long as it consumes the track’s real inputs.

Every participating team is expected to submit a short system description paper. See the submission guidelines and the FAQ.

Submission

  1. Produce predictions in the CSV format shown above.
  2. Validate them with the format checker, and check your score locally on a labelled split with the scorer.
  3. Zip the CSV as prediction.zip and upload it to the matching Codabench leaderboard.
Subtask Track Leaderboard
Spoken VQA (1a) English Codabench 17049
Spoken VQA (1a) MSA Codabench 17048
Hallucination (1b) English Codabench 17051
Hallucination (1b) MSA Codabench 17050

Questions

See the FAQ or email imageeval2026@gmail.com. The dataset is released for research use under CC BY-NC 4.0.