README.md

---
license: apache-2.0
language:
- en
- zh
library_name: transformers
tags:
- avbench
- audio-text
- video-text
- audio-video
base_model:
- Qwen/Qwen2-Audio-7B-Instruct
- Qwen/Qwen2.5-Omni-7B
---

# AVBench Models

This repository hosts the evaluator models used in **AVBench**, a benchmark for text-to-audio-video generation quality and cross-modal consistency.

[![Hugging Face Dataset](https://img.shields.io/badge/HuggingFace-Dataset-FFD21E?logo=huggingface&logoColor=black)](https://huggingface.co/datasets/iiiiii123/AVBench)
[![AVBench Models](https://img.shields.io/badge/HuggingFace-AVBench__model-FFD21E?logo=huggingface&logoColor=black)](https://huggingface.co/iiiiii123/AVBench_model)

## AVBench in brief

AVBench evaluates generated content on two splits:
- **Normal split**: common, easier samples.
- **Hard split**: challenging samples with stronger cross-modal requirements.

It covers cross-modal alignment (Audio-Text / Video-Text / Audio-Video) and generation quality dimensions.

Dataset link:
- https://huggingface.co/datasets/iiiiii123/AVBench

## Model zoo used by AVBench

| Model | Use in AVBench | Trained / merged from |
|---|---|---|
| `Qwen2-Audio-7B-AudioTextMatching-Merged` | Audio-Text consistency scoring (AT) | `Qwen/Qwen2-Audio-7B-Instruct` |
| `Qwen2.5-Omni-7B-VideoTextMatching-Merged` | Video-Text consistency scoring (VT) | `Qwen/Qwen2.5-Omni-7B` |
| `Qwen2.5-Omni-7B-AudioVideoMatching-Merged` | Audio-Video consistency scoring (AV) | `Qwen/Qwen2.5-Omni-7B` |

## Notes

These models are released for AVBench evaluation and analysis.