metadata
dataset_info:
- config_name: vision_bench_0701
features:
- name: question_id
dtype: string
- name: instruction
dtype: string
- name: image
dtype: image
- name: language
dtype: string
splits:
- name: test
num_bytes: 1654009592
num_examples: 500
download_size: 1653981819
dataset_size: 1654009592
- config_name: vision_bench_0617
features:
- name: question_id
dtype: string
- name: instruction
dtype: string
- name: image
dtype: image
- name: language
dtype: string
splits:
- name: test
num_bytes: 1193682526
num_examples: 500
download_size: 1193578497
dataset_size: 1193682526
configs:
- config_name: vision_bench_0701
data_files:
- split: test
path: vision_bench_0701/test-*
- config_name: vision_bench_0617
data_files:
- split: test
path: vision_bench_0617/test-*
WildVision-Bench
We have two versions of Wildvision-Bench data
vision_bench_0617: the selected 500 examples that best simulates the vision-arena elo ranking, same data in the paper.vision_bench_0701: the further filter and selected 500 examples by NSFW and manual selection. Leaderboard are still preparing.
Evaluation
Please refer to our Github for evaluation
If you want to evaluate your model, please use the vision_bench_0617 version to fairly compare the performance with other models in the following leaderboard.
Leaderboard (vision_bench_0717)
| Model | Score | 95% CI | Win Rate | Reward | Much Better | Better | Tie | Worse | Much Worse | Avg Tokens |
|---|---|---|---|---|---|---|---|---|---|---|
| gpt-4o | 89.15 | (-1.9, 1.5) | 80.6% | 56.4 | 255.0 | 148.0 | 14.0 | 72.0 | 11.0 | 142 |
| gpt-4-vision-preview | 79.78 | (-2.9, 2.2) | 71.8% | 39.4 | 182.0 | 177.0 | 22.0 | 91.0 | 28.0 | 138 |
| Reka-Flash | 64.65 | (-2.6, 2.7) | 58.8% | 18.9 | 135.0 | 159.0 | 28.0 | 116.0 | 62.0 | 168 |
| claude-3-opus-20240229 | 62.03 | (-3.7, 2.8) | 53.0% | 13.5 | 103.0 | 162.0 | 48.0 | 141.0 | 46.0 | 105 |
| yi-vl-plus | 55.05 | (-3.4, 2.3) | 52.8% | 7.2 | 98.0 | 166.0 | 29.0 | 124.0 | 83.0 | 140 |
| liuhaotian/llava-v1.6-34b | 51.89 | (-3.4, 3.8) | 49.2% | 2.5 | 90.0 | 156.0 | 26.0 | 145.0 | 83.0 | 153 |
| claude-3-sonnet-20240229 | 50.0 | (0.0, 0.0) | 0.2% | 0.1 | 0.0 | 1.0 | 499.0 | 0.0 | 0.0 | 114 |
| claude-3-haiku-20240307 | 37.83 | (-2.6, 2.8) | 30.6% | -16.5 | 54.0 | 99.0 | 47.0 | 228.0 | 72.0 | 89 |
| gemini-pro-vision | 35.57 | (-3.0, 3.2) | 32.6% | -21.0 | 80.0 | 83.0 | 27.0 | 167.0 | 143.0 | 68 |
| liuhaotian/llava-v1.6-vicuna-13b | 33.87 | (-2.9, 3.3) | 33.8% | -21.4 | 62.0 | 107.0 | 25.0 | 167.0 | 139.0 | 136 |
| deepseek-ai/deepseek-vl-7b-chat | 33.61 | (-3.3, 3.0) | 35.6% | -21.2 | 59.0 | 119.0 | 17.0 | 161.0 | 144.0 | 116 |
| THUDM/cogvlm-chat-hf | 32.01 | (-2.2, 3.0) | 30.6% | -26.4 | 75.0 | 78.0 | 15.0 | 172.0 | 160.0 | 61 |
| liuhaotian/llava-v1.6-vicuna-7b | 26.41 | (-3.3, 3.1) | 27.0% | -31.4 | 45.0 | 90.0 | 36.0 | 164.0 | 165.0 | 130 |
| idefics2-8b-chatty | 23.96 | (-2.2, 2.4) | 26.4% | -35.8 | 44.0 | 88.0 | 19.0 | 164.0 | 185.0 | 135 |
| Qwen/Qwen-VL-Chat | 18.08 | (-1.9, 2.2) | 19.6% | -47.9 | 42.0 | 56.0 | 15.0 | 155.0 | 232.0 | 69 |
| llava-1.5-7b-hf | 15.5 | (-2.4, 2.4) | 18.0% | -47.8 | 28.0 | 62.0 | 25.0 | 174.0 | 211.0 | 185 |
| liuhaotian/llava-v1.5-13b | 14.43 | (-1.7, 1.6) | 16.8% | -52.5 | 28.0 | 56.0 | 19.0 | 157.0 | 240.0 | 91 |
| BAAI/Bunny-v1_0-3B | 12.98 | (-2.0, 2.1) | 16.6% | -54.4 | 23.0 | 60.0 | 10.0 | 164.0 | 243.0 | 72 |
| openbmb/MiniCPM-V | 11.95 | (-2.4, 2.1) | 13.6% | -57.5 | 25.0 | 43.0 | 16.0 | 164.0 | 252.0 | 86 |
| bczhou/tiny-llava-v1-hf | 8.3 | (-1.6, 1.2) | 11.0% | -66.2 | 16.0 | 39.0 | 15.0 | 127.0 | 303.0 | 72 |
| unum-cloud/uform-gen2-qwen-500m | 7.81 | (-1.3, 1.7) | 10.8% | -68.5 | 16.0 | 38.0 | 11.0 | 115.0 | 320.0 | 92 |
Citation
@article{lu2024wildvision,
title={WildVision: Evaluating Vision-Language Models in the Wild with Human Preferences},
author={Lu, Yujie and Jiang, Dongfu and Chen, Wenhu and Wang, William Yang and Choi, Yejin and Lin, Bill Yuchen},
journal={arXiv preprint arXiv:2406.11069},
year={2024}
}