Dataset Card for AVHalluBench

• The dataset is for benchmarking hallucination levels in audio-visual LLMs. It consists of 175 videos and each video has hallucination-free audio and visual descriptions. The statistics are provided in the figure below, and more information can be found in our paper.
• Paper: CrossCheckGPT: Universal Hallucination Ranking for Multimodal Foundation Models
• Multimodal Hallucination Leaderboard: https://huggingface.co/spaces/scb10x/multimodal-hallucination-leaderboard

Dataset Summary

• The videos can be found and downloaded at https://huggingface.co/datasets/potsawee/avhallubench/tree/main/videos. Each video can be identified using video_id.
• Model-generated outputs can be compared against the provided audio and visual descriptions.

Dataset Structure

Each instance consists of:

• video_id: ID for each video
• source: Data source of each video
• audio_description: hallucination-free manual audio description
• visual_description: hallucination-free manual video description

Citation Information

@misc{sun2024crosscheckgpt,
      title={CrossCheckGPT: Universal Hallucination Ranking for Multimodal Foundation Models}, 
      author={Guangzhi Sun and Potsawee Manakul and Adian Liusie and Kunat Pipatanakul and Chao Zhang and Phil Woodland and Mark Gales},
      year={2024},
      eprint={2405.13684},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}