---
language:
- en
license: cc-by-4.0
task_categories:
- visual-question-answering
tags:
- multimodal
- hallucination
- benchmark
- vision
- vqa
size_categories:
- 10k<n<100k
pretty_name: PRE-HAL
configs:
  - config_name: default
    data_files:
      - split: test
        path: dataset.csv
---

# PRE-HAL: Multimodal Hallucination Evaluation Benchmark

## Dataset Summary
PRE-HAL is a visual question answering (VQA) dataset designed to evaluate and mitigate hallucination in Multimodal Large Language Models (MLLMs). It focuses on testing the model's ability to distinguish between visual perception and parametric knowledge, specifically targeting various hallucination types.

### Data Instances
Each instance represents a multiple-choice question associated with an image.

### Data Fields
* **index**: Unique identifier for the sample.
* **image**: The input image file path.
* **question**: The query text related to the image.
* **A, B, C, D**: The multiple-choice options.
* **answer**: The correct option key (e.g., A, B, C, or D).
* **hallucination_type**: The specific category of hallucination being tested (e.g., reasoning relation, perception instance).

## Usage
This dataset is intended for:
1. Evaluating the reliability of MLLMs.
2. Benchmarking hallucination rates across different model architectures.
3. Fine-tuning models to improve grounding and reduce fabrication.

## Citation
If you use this dataset in your research, please cite it as follows:

```bibtex
@article{huang2025visual,
  title={Visual hallucination detection in large vision-language models via evidential conflict},
  journal = {International Journal of Approximate Reasoning},
  volume = {186},
  pages = {109507},
  year = {2025},
  doi = {https://doi.org/10.1016/j.ijar.2025.109507},
  author = {Tao Huang and Zhekun Liu and Rui Wang and Yang Zhang and Liping Jing},
  publisher={Elsevier}
}