File size: 4,111 Bytes
3fdb5fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
82dfbe9
3fdb5fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
82dfbe9
 
 
 
 
3fdb5fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
---
license: cc-by-nc-4.0
task_categories:
  - audio-classification
  - video-text-to-text
  - question-answering
language:
  - en
size_categories:
  - 1K<n<10K
tags:
  - benchmark
  - evaluation
  - audio-visual
  - synchronization
  - vggsound
  - multimodal
  - mllm
pretty_name: "VGGSync-3K: out-of-domain audio-visual sync benchmark"
configs:
  - config_name: default
    data_files:
      - split: test
        path: eval/test_3k.jsonl
---

# VGGSync-3K · out-of-domain audio-visual sync benchmark

Out-of-domain evaluation set used in the paper
**[When Vision Speaks for Sound](https://arxiv.org/abs/2605.16403)**.

Derived from [VGGSoundSync](https://www.robots.ox.ac.uk/~vgg/research/avs/),
this 3,000-clip slice tests whether a video-capable MLLM can detect
audio temporal offsets on **everyday sound events** outside the THUD
in-domain training distribution.

Each item is one VGGSound clip in one of three conditions:

| Condition | Count | `gt_synced` | `gt_direction` | `gt_offset_sec` |
|---|---:|---|---|---|
| Audio aligned (no shift) | 1,000 | `true`  | `none`  | `0.0` |
| Audio shifted **early**  | 1,000 | `false` | `early` | shift amount |
| Audio shifted **delay**  | 1,000 | `false` | `delay` | shift amount |

Shift magnitudes span 4 difficulty levels (`very_easy`, `easy`, `medium`,
`hard`); synced clips are tagged `difficulty=synced`.

## What's in this repo

| File | Description |
|---|---|
| `eval/test_3k.jsonl` | Test records — uid / label / paths / ground truth |
| `media.zip` | All referenced videos + audios |

## How to use

```bash
# Download
hf download Rakancorle1/vggsync-3k --repo-type=dataset --local-dir vggsync-3k
cd vggsync-3k

# Unzip media — paths in JSONL resolve automatically
unzip -q media.zip
```

```python
from datasets import load_dataset
ds = load_dataset("Rakancorle1/vggsync-3k")
```

## Record schema

```jsonc
{
  "uid": "-r3nM90RCNs_medium_early_0.5s",
  "ytid": "-r3nM90RCNs",
  "label": "sharpen knife",
  "difficulty": "very_easy" | "easy" | "medium" | "hard" | "synced",
  "video_path": "videos/<uid>.mp4",
  "audio_path": "audios/<uid>.wav",
  "gt_synced": true | false,
  "gt_direction": "none" | "early" | "delay",
  "gt_offset_sec": 0.0          // > 0 for shifted; sign implied by direction
}
```

`uid` follows the pattern `<ytid>_<difficulty>_<direction>_<offset>s`
for shifted clips, and `<ytid>` alone for synced clips.

## Reference eval scripts

Paper evaluation code lives in the
[wvs-code repository](https://github.com/rakanWen/wvs-code) on GitHub.

The JSONL is self-contained — `video_path`, `audio_path`, and the `gt_*`
fields are all you need to compare any model's output against ground truth.

## Citation

```bibtex
@article{wen2026whenvisionspeaksforsound,
  title     = {When Vision Speaks for Sound},
  author    = {Xiaofei Wen and Wenjie Jacky Mo and Xingyu Fu and Rui Cai and
               Tinghui Zhu and Wendi Li and Yanan Xie and Muhao Chen and Peng Qi},
  year      = {2026},
  url       = {https://arxiv.org/abs/2605.16403}
}
```

Please also cite the original VGGSoundSync work:

```bibtex
@InProceedings{Chen21b,
  title        = "Audio-Visual Synchronization in the Wild",
  author       = "Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, Andrew Zisserman",
  booktitle    = "BMVC",
  year         = "2021"}
```

## Related releases

- **[Rakancorle1/thud-eval](https://huggingface.co/datasets/Rakancorle1/thud-eval)** — in-domain audio-visual Clever Hans benchmark (sync / mute / swap)
- **[Rakancorle1/hans-10k](https://huggingface.co/datasets/Rakancorle1/hans-10k)** — DPO training data
- **[Rakancorle1/hans-sft-4k](https://huggingface.co/datasets/Rakancorle1/hans-sft-4k)** — SFT training data
- **[Collection](https://huggingface.co/collections/Rakancorle1/when-vision-speaks-for-sound)** — everything in one place

## License

This subset is released under **CC-BY-NC-4.0** for the annotations and
shifted-clip pairings. The underlying audio/video content is derived
from VGGSound — please follow VGGSound's licensing terms when using or
redistributing.