bigai-nlco/VideoHallucer
收藏资源简介:
VideoHallucer是第一个用于大型视频-语言模型(LVLMs)幻觉检测的综合基准。该数据集将幻觉分为内在和外在两类,并进一步细分为对象关系、时间、语义细节、外在事实和外在非事实幻觉。数据集采用了对抗性二元视频问答方法进行评估,其中包含基本问题和幻觉问题的配对。数据统计显示,每种类型的幻觉问题有400个,对应的视频数量分别为183、165、400、200和200个。
VideoHallucer is the first comprehensive benchmark for hallucination detection in large video-language models (LVLMs). The dataset categorizes hallucinations into two main types: intrinsic and extrinsic, offering further subcategories for detailed analysis, including object-relation, temporal, semantic detail, extrinsic factual, and extrinsic non-factual hallucinations. The dataset adopts an adversarial binary VideoQA method for comprehensive evaluation, where pairs of basic and hallucinated questions are crafted strategically. Data statistics show that there are 400 questions for each type of hallucination, with corresponding video counts of 183, 165, 400, 200, and 200 respectively.
VideoHallucer 数据集概述
数据集描述
- 任务类别: 问答 (question-answering)
- 语言: 英语 (en)
- 数据规模: 1K<n<10K
- 许可证: MIT
数据统计
| 幻觉类型 | 对象关系幻觉 | 时间幻觉 | 语义细节幻觉 | 外部事实幻觉 | 外部非事实幻觉 |
|---|---|---|---|---|---|
| 问题数量 | 400 | 400 | 400 | 400 | 400 |
| 视频数量 | 183 | 165 | 400 | 200 | 200 |




