OmniHalluBench

OmniHalluBench is a unified evaluation benchmark for omni-modal hallucination, built from six existing datasets spanning text, image, audio, and video, and covering both judgment and free-form generation. It contains 3,540 evaluation examples in total.

This release provides OmniHalluBench in a unified, portable directory structure.

Paper: OmniConfess: Eliciting Token Confessions to Mitigate Omni-Modal Hallucination
Code: RongHuiQiang/OmniConfess

Layout

OmniHalluBench/
├── phd/
│   ├── annotations/phd.json              # 906 samples (image + text)
│   └── media/images/
├── cmm/
│   ├── annotations/cmm.json              # 721 samples (video + audio)
│   └── media/reorg_raw_files/
├── pubmedqa/
│   └── annotations/pubmedqa.json         # 468 samples (text)
├── ragtruth/
│   └── annotations/ragtruth.json         # 435 samples (text)
├── haloquest/
│   ├── annotations/haloquest.json        # 365 samples (image)
│   └── media/images/
└── avhbench/
    ├── annotations/avhbench.json         # 645 samples (video + audio)
    └── media/videos/

Usage

Download the benchmark from Hugging Face:

pip install -U huggingface_hub
hf download Eric1231/OmniHalluBench \
  --repo-type dataset \
  --local-dir OmniHalluBench

Media paths in each annotations/<ds>.json file are relative to the corresponding dataset directory, such as media/images/val2014/COCO_...jpg.

Set the evaluation paths as follows, replacing <ds> with the subset name:

--data_path OmniHalluBench/<ds>/annotations/<ds>.json
--data_root OmniHalluBench/<ds>/

PubMedQA and RAGTruth contain their text passages directly in the JSON files and require no media directory.

For evaluation instructions, please refer to the OmniConfess evaluation guide.

Source Datasets

Subset	Modality	Samples	Paper	Source
PhD	Image + text	906	CVPR 2025	GitHub
CMM	Video + audio	721	Paper	GitHub
PubMedQA	Text	468	EMNLP 2019	GitHub
RAGTruth	Text	435	ACL 2024	GitHub
HaloQuest	Image	365	ECCV 2024	GitHub
AVHBench	Video + audio	645	ICLR 2025	GitHub

License

OmniHalluBench is compiled from six existing datasets. All source data remain subject to their original licenses and terms of use. Please refer to the source repositories for details.

Citation

If you use OmniHalluBench, please cite OmniConfess and the relevant source datasets listed above.

@misc{rong2026omniconfess,
  title         = {OmniConfess: Eliciting Token Confessions to Mitigate Omni-Modal Hallucination},
  author        = {Huiqiang Rong and Haoran Luo and Hui Feng and Zhonghong Ou and Kaiwen Xue and Guoxin Zhang and Yifan Zhu},
  year          = {2026},
  eprint        = {2610.02999},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.02999}
}