HHVD

# HHVD: Human Hallucination Verification Dataset ## Introduction **HHVD** is **H**uman **H**allucination **V**erification **D**ataset for multimodal hallucination verifiability, constructed from 4,470 time-constrained human responses to image–text pairs and designed to evaluate obvious and elusive hallucinations based on how easily humans can verify them. <p align="center"> <img src="intro.png" alt="outline" width="400"/> </p> ## Data Construction Our data construction pipeline builds a human-centered benchmark for multimodal hallucination verifiability by combining AMBER-based image–text pair generation with time-constrained human annotation. The collected responses enable us to distinguish obvious, elusive, and neutral hallucinations according to human identification accuracy and verification effort. <p align="center"> <img src="pipeline.png" alt="outline" width="1000"/> </p> ## QuickStart ### Image Download Get the images from this [LINK](https://github.com/junyangwang0410/AMBER). ### Responses Generation The format of responses is: ```json [ { "response": "Yes", "No" or "Uncertain" }, ...... { "response": "Yes", "No" or "Uncertain" } ] ``` ## Citation If you find our work useful, please cite our paper: ```bibtex @article{pang2026steering, title={Steering the Verifiability of Multimodal AI Hallucinations}, author={Pang, Jianhong and Cheng, Ruoxi and Ye, Ziyi and Ma, Xingjun and Wu, Zuxuan and Huang, Xuanjing and Jiang, Yu-Gang}, journal={arXiv preprint arXiv:2604.06714}, year={2026} } ```