Datasets:
File size: 4,336 Bytes
b20d107 | 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 | ---
license: mit
task_categories:
- audio-classification
- image-to-text
- text-to-audio
- audio-to-audio
language:
- en
pretty_name: Audio-Centric Multimodal Benchmark
configs:
- config_name: composed_audio_retrieval_queries
data_files:
- split: test
path: data/composed_audio_retrieval/queries/*.parquet
- config_name: composed_audio_retrieval_candidates
data_files:
- split: test
path: data/composed_audio_retrieval/candidates/*.parquet
- config_name: audio_visual_queries
data_files:
- split: test
path: data/audio_visual/queries/*.parquet
- config_name: audio_visual_candidates
data_files:
- split: test
path: data/audio_visual/candidates/*.parquet
---
# Audio-Centric Multimodal Benchmark (ACM)
This dataset packages the ACM benchmark introduced by
[Efficient and High-Fidelity Omni Modality Retrieval](https://hmchuong.github.io/omniret/).
The paper is available at [arXiv:2603.02098](https://arxiv.org/abs/2603.02098).
Dataset repo: `chuonghm/ACM`
ACM contains four HuggingFace subsets. Each subset uses a single `test` split
so query and candidate tables can keep their natural schemas:
- `composed_audio_retrieval_queries`: flattened AT2A query rows.
- `composed_audio_retrieval_candidates`: AT2A audio candidate pool.
- `audio_visual_queries`: paired audio-visual query IDs.
- `audio_visual_candidates`: audio/video/image candidate pool.
## Files
- `data/composed_audio_retrieval/queries/*.parquet`: 4,251 flattened AT2A query
rows with source audio, target audio, modification text, and captions.
- `data/composed_audio_retrieval/candidates/*.parquet`: 5,480 audio candidates.
- `data/audio_visual/queries/*.parquet`: 1,292 paired audio-visual query IDs.
- `data/audio_visual/candidates/*.parquet`: 5,480 audio/video/image candidates.
- Optional `media.zip`: a single archive holding every raw media file
(`media/audios/*.wav`, `media/videos/*.mp4`, `media/images/*.jpg`), uploaded
with `--upload-media`. Download and unzip it (it restores the `media/`
folder) to resolve the `*_path` columns. It is shipped as one archive rather
than tens of thousands of loose files so the Dataset Viewer can build.
## Loading
```python
from datasets import load_dataset
at2a_queries = load_dataset(
"chuonghm/ACM",
"composed_audio_retrieval_queries",
split="test",
)
at2a_candidates = load_dataset(
"chuonghm/ACM",
"composed_audio_retrieval_candidates",
split="test",
)
audio_visual_queries = load_dataset(
"chuonghm/ACM",
"audio_visual_queries",
split="test",
)
audio_visual_candidates = load_dataset(
"chuonghm/ACM",
"audio_visual_candidates",
split="test",
)
```
The first preview rows in each split embed HuggingFace `Audio`, `Image`, and
`Video` bytes so the Dataset Viewer can render playable audio, images, and
videos without duplicating every media file inside Parquet. Later rows keep
empty media cells and provide `*_path` columns. Download `media.zip` and unzip
it (it restores the `media/` folder) to resolve those paths to real files.
## License and Source Data
The ACM text information and benchmark metadata in this repository are released
under the MIT license.
The media files are derived from VGG-Sound and remain subject to the
VGG-Sound/source-video terms. The original media source is
[Loie/VGGSound](https://huggingface.co/datasets/Loie/VGGSound), which mirrors
the [VGG-Sound dataset](https://www.robots.ox.ac.uk/~vgg/data/vggsound/)
introduced in the [VGG-Sound paper](https://arxiv.org/abs/2004.14368).
## Citation
```bibtex
@article{huynh2026omniret,
title = {Efficient and High-Fidelity Omni Modality Retrieval},
author = {Huynh, Chuong and Luong, Manh and Shrivastava},
journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2026}
}
```
## Notes
The paper/project page should be cited when using this benchmark. Media source
files are derived from VGG-Sound. The text information and benchmark metadata
are released under MIT, while media follows the VGG-Sound/source-video terms.
See the original HuggingFace dataset source
[Loie/VGGSound](https://huggingface.co/datasets/Loie/VGGSound), the
[VGG-Sound homepage](https://www.robots.ox.ac.uk/~vgg/data/vggsound/), and
the [VGG-Sound paper](https://arxiv.org/abs/2004.14368).
|