---
title: EvaOmni-Dataset
canonical_url: "https://www.modelscope.cn/datasets/VisionXLab/EvaOmni-Dataset"
md_url: "https://www.modelscope.cn/datasets/VisionXLab/EvaOmni-Dataset.md"
repository: VisionXLab/EvaOmni-Dataset
chinese_name: "证据驱动的音视理解评测数据集"
last_updated: 2026-10-21
license: apache-2.0
storage_size: "12 GB"
downloads: 1513
stars: 0
---

# EvaOmni-Dataset

> EvaOmni-Dataset - VisionXLab 在 ModelScope 开源的数据集。EvaOmni: An Evidence-Grounded Benchmark for Audio-Visual Reasoning

VisionXLab/EvaOmni-Dataset 是 ModelScope 魔搭社区上的数据集，存储大小 12 GB，采用 apache-2.0 许可。

- **Repository**: VisionXLab/EvaOmni-Dataset
- **License**: apache-2.0
- **Storage size**: 12 GB
- **Downloads**: 1513
- **Stars**: 0
- **Last updated**: 2026-10-21

Source: https://www.modelscope.cn/datasets/VisionXLab/EvaOmni-Dataset

---

# EvaOmni: An Evidence-Grounded Benchmark for Audio-Visual Reasoning

[GitHub](https://github.com/VisionXLab/EvaOmni) · [Paper](https://arxiv.org/abs/2610.28444) · [Hugging Face Dataset](https://huggingface.co/datasets/VisionXLab/EvaOmni-Dataset) · [ModelScope Dataset](https://modelscope.cn/datasets/VisionXLab/EvaOmni-Dataset)

## Dataset Overview

**EvaOmni** is an evidence-grounded benchmark for evaluating audio-visual reasoning in omni-modal large language models. It assesses whether models can integrate audio and visual information to answer questions and locate the evidence supporting their answers. Models must return both an answer and timestamped audio and visual evidence intervals, enabling joint evaluation of answer correctness, evidence coverage, and temporal localization.

The dataset contains **392 videos and 1,027 human-annotated question-answer pairs**, with video durations ranging from **1 to 12 minutes** across diverse real-world scenarios. Each question includes temporal intervals and descriptions of supporting audio and visual evidence. Dataset construction combines expert annotation, human cross-validation, and model-assisted screening under text-only, audio-only, and visual-only conditions to reduce textual and unimodal shortcuts.

## Capability Taxonomy

EvaOmni covers **15 fine-grained capabilities** organized into three levels:

- **Level 1: Retrieval and Grounding.** Temporal alignment, event binding, and entity binding.
- **Level 2: Basic Understanding.** Causal reasoning, emotional intelligence, intent understanding, atmosphere understanding, implicit event reasoning, semantic disambiguation, temporal understanding, spatial understanding, and conditional counting.
- **Level 3: Complex Reasoning.** Multi-hop reasoning, long-range reasoning, and multi-evidence chains.

## Evaluation

For each question, a model predicts an answer and supporting evidence intervals for audio and vision separately. The main metrics are:

- **Accuracy (Acc.):** Measures answer correctness.
- **Coverage Grounded Score (C.G.):** Combines answer correctness with the mean fraction of annotated evidence segments hit in each modality. An annotated segment counts as hit when it has positive-duration overlap with a predicted interval.
- **Bottleneck Grounded Score (B.G.):** Combines answer correctness with the lower of the audio and visual temporal IoUs, measuring evidence localization in the weaker modality.

Both C.G. and B.G. incorporate answer correctness; an incorrect answer receives zero on both metrics. Evaluation code and usage instructions are available in the [GitHub repository](https://github.com/VisionXLab/EvaOmni).

## Dataset Files

- `test_file.json`: Question-answer annotations, including question IDs, capability levels and categories, video IDs, source URLs, questions, options, answers, and audio and visual evidence.
- `rawvideo/`: Original videos.
- `resized_video/`: Preprocessed videos. Use this directory to reproduce the paper's results.

## Citation

If you find EvaOmni useful for your research, please cite:

```bibtex
@article{hu2026evaomni,
  title={EvaOmni: An Evidence-Grounded Benchmark for Audio-Visual Reasoning},
  author={Hu, Xiaoxing and Yao, Wenxi and Shi, Chaoyuan and Ye, Ziqi and Guo, Xiaorong and Li, Peixuan and Sun, Zhuoran and Gong, Ziyang and Zhang, Peiyuan and Liu, Mingxin and Ma, Shuran and Liu, Hongbo and Yu, Xiaomin and Tian, Yu and Wang, Wentao and Yang, Zhenjie and Hu, Jian and Shen, Yunhang and Cao, Haoyu and Yang, Xue},
  journal={arXiv preprint arXiv:2610.28444},
  year={2026}
}
```