Hugging Face's logo

Datasets:
KinghtH
/
VL-RouterBench

Modalities:
Image
Formats:
imagefolder
Size:
< 1K
ArXiv:
Libraries:
Datasets
Dataset card Data Studio Files
xet
Community
1
VL-RouterBench
File size: 2,760 Bytes 
6636348
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
 
<div align="center">

<p align="center">
  <img src="assets/icon.png" width="220" alt="VL-RouterBench logo" />
</p>

### VL-RouterBench: A Benchmark for Vision–Language Model Routing

[![arXiv](https://img.shields.io/badge/arXiv-2512.23562-red.svg)](https://arxiv.org/abs/2512.23562)
[![GitHub](https://img.shields.io/badge/GitHub-Repository-black.svg)](https://github.com/K1nght/VL-RouterBench)

</div>

## Overview

We provides a clean, reproducible implementation of **VL-RouterBench**, a benchmark and toolkit for **routing across a pool of Vision–Language Models (VLMs)** under both **performance** and **performance–cost** objectives.

<p align="center">
  <img src="assets/pipeline.png" width="900" alt="VL-RouterBench pipeline" />
</p>

## 📦 Data Preparation

VL-RouterBench converts [**VLMEvalKit**](https://github.com/open-compass/VLMEvalKit) outputs into a unified routing benchmark.

To make data setup easier, we provide a pre-packaged archive **`vlm_router_data.tar.gz`** that contains everything needed to run the pipeline. You can download it from any of the following channels and extract it under the repo root:

- **Google Drive**: [vlm_router_data.tar.gz](https://drive.google.com/file/d/1Va18MW8nJqvatxDXQDQq0t9NAqr93hMg/view?usp=sharing)
- **Baidu Netdisk**: [vlm_router_data.tar.gz](https://pan.baidu.com/s/1D_P8YwY_E5kDA5dUB-ovng) (code: xb1s)
- **Hugging Face**: [vlm_router_data.tar.gz](https://huggingface.co/datasets/KinghtH/VL-RouterBench)

After downloading, extract it as:

```bash

tar -xzf vlm_router_data.tar.gz

```

By default, the pipeline expects the following directories (relative to repo root):

```bash

vlm_router_data/

  VLMEvalKit_evaluation/   # required (for is_correct / evaluation)

  VLMEvalKit_inference/    # required for accurate output-token counting (Step 2)

  TSV_images/              # optional (for TSV-packed image datasets)

```

Notes:
- **`VLMEvalKit_evaluation/`** is used by Step 1 & 4 (contains correctness signals).

- **`VLMEvalKit_inference/`** is used by Step 2 (extract real model outputs to count output tokens).
- **`TSV_images/`** is used by routers for training and inference to make routing decisions.



## 📝 Citation



If you find this benchmark useful, please cite:



```bibtex

@misc{huang2025vlrouterbenchbenchmarkvisionlanguagemodel,

      title={VL-RouterBench: A Benchmark for Vision-Language Model Routing}, 

      author={Zhehao Huang and Baijiong Lin and Jingyuan Zhang and Jingying Wang and Yuhang Liu and Ning Lu and Tao Li and Xiaolin Huang},

      year={2025},

      eprint={2512.23562},

      archivePrefix={arXiv},

      primaryClass={cs.LG},

      url={https://arxiv.org/abs/2512.23562}, 

}

```



---