DisasterVQA 🌍

A Visual Question Answering Benchmark Dataset for Disaster Scenes

Introduced at ICWSM 2026  ·  4,405 QA pairs  ·  1,395 images  ·  7 models evaluated

---

Dataset Overview

Property	Value
Total QA pairs	4,405
Unique images	1,395
Question types	Yes/No (48.9%), Multiple-Choice (38.4%), Open-Ended (12.7%)
Disaster categories	earthquake, flood, hurricane, fire, accident, storm, wildfire, landslide, other
Image sources	MEDIC, CrisisMMD, Incidents1M
Evaluated models (7)	GPT-4o-mini, GPT-4.1-mini, Llama 3.2, Mistral Small, Molmo-7B-D, Pixtral, Qwen2.5-VL

---

Quick Links

Resource	Link
📄 Paper	arXiv:2601.13839
📦 Zenodo	10.5281/zenodo.18365212
🤗 Dataset	HuggingFace
🏛️ Conference	ICWSM 2026

---

Repository Structure

DisasterVQA/
├── dataset/                    # Benchmark dataset
│   ├── disasterVQA_dataset.json
│   ├── disasterVQA_allmodel_judge_outputs.json
│   └── README.md
├── prompts/                    # LLM prompts for question generation and judging
│   ├── question_generation.txt
│   ├── judge_binary.txt
│   ├── judge_mcq.txt
│   └── judge_open_ended.txt
├── inference/                  # Model inference scripts (7 models)
│   ├── gpt4o_mini.py
│   ├── gpt41_mini.py
│   ├── llama32.py
│   ├── mistral_small.py
│   ├── molmo.py
│   ├── pixtral.py
│   ├── qwen25_vl.py
│   └── README.md
├── judge/                      # LLM-as-judge post-processing
│   ├── judge_postprocess.py
│   └── README.md
├── classification/             # Humanitarian category classification
│   ├── classify_humanitarian.py
│   ├── taxonomy.json
│   └── README.md
└── evaluation/                 # Evaluation scripts
    ├── evaluate_by_question_type.py
    ├── evaluate_by_region.py
    ├── evaluate_by_humanitarian_category.py
    └── README.md

Note: Images are not included in this repository. The full dataset including images is available on Zenodo and HuggingFace.

---

Quick Start

Skip to step 4 if you want to use the pre-computed outputs already in dataset/disasterVQA_allmodel_judge_outputs.json.

Pipeline: inference/ → judge/ → classification/ → evaluation/

1. Install dependencies

pip install -r requirements.txt

2. Run inference

Each script in inference/ produces a raw output JSON in the same format. Example using GPT-4o-mini:

python inference/gpt4o_mini.py \
    --input_json  dataset/disasterVQA_dataset.json \
    --output_json outputs/gpt4o_mini_raw.json \
    --deployment  <your-deployment-name> \
    --endpoint    <your-azure-endpoint> \
    --api_key     <your-api-key>

See inference/README.md for instructions for every model.

3. Judge raw outputs

Combine raw outputs from all models into a single JSON (one entry per question, one key per model), then run the judge:

python judge/judge_postprocess.py \
    --input           outputs/all_models_raw.json \
    --output          outputs/all_models_judged.json \
    --prompt-yesno    prompts/judge_binary.txt \
    --prompt-mcq      prompts/judge_mcq.txt \
    --prompt-open     prompts/judge_open_ended.txt \
    --endpoint        <your-azure-endpoint> \
    --api-key         <your-api-key> \
    --deployment-name <your-deployment-name>

See judge/README.md for details on the expected input format.

4. Classify humanitarian categories

Assign each QA entry a humanitarian response category using the taxonomy in classification/taxonomy.json:

python classification/classify_humanitarian.py \
    --input      outputs/all_models_judged.json \
    --output     dataset/disasterVQA_allmodel_judge_outputs.json \
    --deployment <your-deployment-name> \
    --endpoint   <your-azure-endpoint> \
    --api-key    <your-api-key>

See classification/README.md for the full taxonomy and optional flags.

5. Evaluate

# Per-model metrics by question type (accuracy, precision, recall, F1)
python evaluation/evaluate_by_question_type.py \
    --input  dataset/disasterVQA_allmodel_judge_outputs.json \
    --output results/metrics_by_question_type.xlsx

# Per-model metrics by geographic region
python evaluation/evaluate_by_region.py \
    --input  dataset/disasterVQA_allmodel_judge_outputs.json \
    --output results/metrics_by_region.xlsx

# Per-model metrics by humanitarian category
python evaluation/evaluate_by_humanitarian_category.py \
    --input  dataset/disasterVQA_allmodel_judge_outputs.json \
    --excel  results/metrics_by_humanitarian_category.xlsx \
    --csv    results/metrics_by_humanitarian_category.csv

---

Results

Overall per-model performance across all question types.

By humanitarian category

By disaster type

By geographic region

Error analysis

To reproduce these results, see the scripts in evaluation/.

---

License

The benchmark is released under the Creative Commons Attribution Share Alike 4.0 International (CC BY-SA 4.0) license. The underlying images belong to their respective source datasets (MEDIC, CrisisMMD, Incidents1M) — please refer to the original dataset licenses for usage terms.

---

Citation

If you use this dataset in a publication, please also cite the ICWSM 2026 paper:

@inproceedings{disastervqa_icwsm2026,
  author    = {Al-Mohannadi, Aisha and Firoz, Ayisha and Yang, Yin and Imran, Muhammad and Ofli, Ferda},
  title     = {DisasterVQA: A Visual Question Answering Benchmark Dataset for Disaster Scenes},
  booktitle = {Proceedings of the International AAAI Conference on Web and Social Media (ICWSM)},
  year      = {2026},
  address   = {Los Angeles, California, USA},
  url       = {https://arxiv.org/abs/2601.13839}
}

---

Contact

For questions or issues, please open a GitHub Issue or contact the authors via the paper.