README.md

---
pretty_name: SMP-FSAVC Splits and Semantic Prompts
language:
  - en
license: other
task_categories:
  - video-classification
  - audio-classification
size_categories:
  - 10K<n<100K
tags:
  - audio-visual
  - few-shot-learning
  - semantic-prompting
  - video-captioning
configs:
  - config_name: ave
    default: true
    data_files:
      - split: source_train
        path: viewer/AVE/pretrain.parquet
      - split: source_test
        path: viewer/AVE/pretrain_test.parquet
      - split: target_train_pool
        path: viewer/AVE/fewshot.parquet
      - split: target_test
        path: viewer/AVE/fewshot_test.parquet
  - config_name: kinetics-sounds
    data_files:
      - split: source_train
        path: viewer/Kinetics-Sounds/pretrain.parquet
      - split: source_test
        path: viewer/Kinetics-Sounds/pretrain_test.parquet
      - split: target_train_pool
        path: viewer/Kinetics-Sounds/fewshot.parquet
      - split: target_test
        path: viewer/Kinetics-Sounds/fewshot_test.parquet
  - config_name: vggsound100
    data_files:
      - split: source_train
        path: viewer/VGGSound100/pretrain.parquet
      - split: source_test
        path: viewer/VGGSound100/pretrain_test.parquet
      - split: target_train_pool
        path: viewer/VGGSound100/fewshot.parquet
      - split: target_test
        path: viewer/VGGSound100/fewshot_test.parquet
---

# SMP-FSAVC splits and semantic prompts

[中文说明](README.zh-CN.md) · [SMP code](https://github.com/DennisHgj/SMP_FSAVC) · [Paper](https://ieeexplore.ieee.org/abstract/document/11352954) · [DOI](https://doi.org/10.1109/TASLPRO.2026.3654246)

This repository releases the source/target class partitions, train/test sample
lists, and mPLUG-2-generated semantic prompts used by **Semantic Modulated
Prompting for Few-Shot Audio-Visual Classification** (SMP-FSAVC). It is ready
to host on both GitHub and the Hugging Face Hub: the original headerless CSV
files are preserved for the training code, while equivalent Parquet files
power the Hugging Face Dataset Viewer.

No video, audio, or extracted frame is distributed here. Users must obtain
media from the respective upstream datasets and follow their terms.

## What is included

| Dataset | Released rows | Source classes | Target classes | Notes |
| --- | ---: | ---: | ---: | --- |
| AVE | 4,051 | 16 | 12 | 28 defined classes |
| Kinetics-Sounds | 22,908 | 19 | 13 | 32 defined classes |
| VGGSound100 | 59,795 | 60 defined / 59 available | 40 | Source label 14 has no obtainable media in this snapshot |
| **Total** | **86,754** |  |  |  |

The VGGSound100 files named `train_caption.csv` and `test_caption.csv` in the
research workspace are not included because they are exact unions of the four
released split files and would duplicate content.

## Split semantics

| File | Hugging Face split | Meaning |
| --- | --- | --- |
| `pretrain.csv` | `source_train` | Source-class pretraining set |
| `pretrain_test.csv` | `source_test` | Source-class evaluation set |
| `fewshot.csv` | `target_train_pool` | Target-class pool from which N-way K-shot support sets are sampled |
| `fewshot_test.csv` | `target_test` | Target-class query/evaluation set |

`fewshot.csv` is a sampling pool, not one fixed K-shot episode. The SMP
protocol repeatedly samples support examples from it.

## Repository layout

```text
.
|-- csv/                       # original headerless files used by SMP
|   |-- AVE/
|   |-- Kinetics-Sounds/
|   `-- VGGSound100/
|-- viewer/                    # equivalent Parquet files for Dataset Viewer
|-- metadata/
|   |-- dataset_statistics.json
|   |-- generation_config.yaml
|   |-- label_maps.json
|   `-- checksums.sha256
|-- scripts/
|   |-- build_release.py
|   `-- verify_release.py
|-- LICENSE_DATA.md
|-- LICENSE_CODE
`-- DATA_REMOVAL.md
```

## Data schema

The raw files have no header. AVE rows contain three fields:

```text
clip_id,integer_label,semantic_prompt
```

Kinetics-Sounds and VGGSound100 rows contain a fourth field:

```text
clip_id,integer_label,semantic_prompt,class_name
```

CSV quoting is used where captions or class names contain commas. The Parquet
mirrors expose the same four named columns for every dataset:

| Column | Type | Description |
| --- | --- | --- |
| `clip_id` | string | Upstream clip identifier; no media URL or media payload |
| `label` | int64 | Dataset-specific integer class label |
| `semantic_prompt` | string | mPLUG-2-generated video caption used as the semantic prompt |
| `class_name` | string | Human-readable class name; derived from the AVE identifier suffix for AVE |

Complete source/target label mappings and availability flags are in
[`metadata/label_maps.json`](metadata/label_maps.json).

## Loading the data

The Parquet files require no custom loading script. After cloning this
repository:

```python
from datasets import load_dataset

ave = load_dataset(
    "parquet",
    data_files={
        "source_train": "viewer/AVE/pretrain.parquet",
        "source_test": "viewer/AVE/pretrain_test.parquet",
        "target_train_pool": "viewer/AVE/fewshot.parquet",
        "target_test": "viewer/AVE/fewshot_test.parquet",
    },
)
```

When this tree is pushed to a Hugging Face dataset repository, the YAML
configuration at the top of this card automatically exposes the `ave`,
`kinetics-sounds`, and `vggsound100` configurations in Dataset Viewer.

For the [SMP training repository](https://github.com/DennisHgj/SMP_FSAVC),
point the source annotation root at a dataset's `csv/<dataset>/` directory for
pretraining. Point the few-shot root at the same directory for target training;
the scripts select the required filenames.

## Semantic prompts

The released `semantic_prompt` values were generated from visual frames with
[mPLUG-2](https://github.com/X-PLUG/mPLUG-2), using its MSVD-finetuned video
captioning checkpoint. The retained settings use 16 RGB frames at 224×224,
ViT-L/14, beam size 5, output lengths 4–20, and seed 42. See
[`metadata/generation_config.yaml`](metadata/generation_config.yaml) for the
full reconstructed configuration and reproducibility caveats.

SMP also supports static prompts. To test that setting, replace each sample's
prompt with:

```python
static_prompt = "a video of [label]"
```

`[label]` is literal text: it is not substituted with the sample label or
class name. Every sample receives exactly the same string.

The published CSVs retain the generated captions; the static alternative is
not duplicated as another set of files.

## Integrity and rebuilding

Verify the checked-in release artifacts with Python's standard library:

```bash
python scripts/verify_release.py
```

To validate the raw CSVs and rebuild all Parquet mirrors and metadata:

```bash
python -m pip install -r requirements-build.txt
python scripts/build_release.py
```

The build script checks row counts, column counts, blank fields, duplicate clip
IDs, label coverage, class-name consistency, and leakage across the four files.
It never rewrites the original CSVs.

## Limitations and responsible use

- Semantic prompts are model-generated descriptions, not ground-truth
  captions. They may hallucinate, omit audible/visible events, encode social
  biases, or disagree with the class label.
- The released sample counts reflect the processed media snapshot available
  during the research. YouTube-hosted source clips can disappear over time, so
  later downloads may not reproduce the same media inventory.
- VGGSound100 source label `14` (`subway, metro`) has no available example in
  this release because the original videos could not be obtained. The numeric
  label space remains `0..59` for compatibility.
- The original frame loader did not explicitly sort stored frame filenames
  before uniform selection. The retained settings therefore document the
  procedure but do not guarantee bit-for-bit caption regeneration on a new
  filesystem snapshot.
- These annotations are intended for research on audio-visual learning and
  few-shot classification. Inspect generated captions and upstream media for
  suitability before using them in sensitive applications.

See [DATA_REMOVAL.md](DATA_REMOVAL.md) for correction or removal requests.

## Licensing

The release contains derived annotations and identifiers, but no media. See
[LICENSE_DATA.md](LICENSE_DATA.md) for the layered licensing terms. In short,
release-authored split definitions, generated prompts, and metadata are
provided under CC BY 4.0 to the extent the authors hold rights; upstream
identifiers and labels remain subject to their source terms, and video
copyright remains with the original owners.

## Citation

If you use these splits or semantic prompts, please cite the SMP article and
the upstream datasets relevant to your experiment.

```bibtex
@ARTICLE{11352954,
  author={Huang, Guanjie and Cui, Yawen and Tsang, Danny H.K. and Wang, Wenwu and Liu, Li},
  journal={IEEE Transactions on Audio, Speech and Language Processing},
  title={Semantic Modulated Prompting for Few-Shot Audio-Visual Classification},
  year={2026},
  volume={34},
  pages={723-736},
  doi={10.1109/TASLPRO.2026.3654246}
}
```

Additional BibTeX entries for AVE, Kinetics/Kinetics-Sounds, VGGSound, and
mPLUG-2 are provided below.

<details>
<summary>Upstream dataset and caption-model citations</summary>

```bibtex
@inproceedings{tian2018ave,
  author={Tian, Yapeng and Shi, Jing and Li, Bochen and Duan, Zhiyao and Xu, Chenliang},
  title={Audio-Visual Event Localization in Unconstrained Videos},
  booktitle={European Conference on Computer Vision},
  pages={247--263},
  year={2018}
}

@inproceedings{carreira2017quo,
  author={Carreira, Joao and Zisserman, Andrew},
  title={Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  pages={6299--6308},
  year={2017}
}

@inproceedings{arandjelovic2017look,
  author={Arandjelovic, Relja and Zisserman, Andrew},
  title={Look, Listen and Learn},
  booktitle={IEEE International Conference on Computer Vision},
  pages={609--617},
  year={2017}
}

@inproceedings{chen2020vggsound,
  author={Chen, Honglie and Xie, Weidi and Vedaldi, Andrea and Zisserman, Andrew},
  title={VGGSound: A Large-Scale Audio-Visual Dataset},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing},
  pages={721--725},
  year={2020}
}

@inproceedings{xu2023mplug2,
  author={Xu, Haiyang and Ye, Qinghao and Yan, Ming and Shi, Yaya and Ye, Jiabo and Xu, Yuanhong and Li, Chenliang and Bi, Bin and Qian, Qi and Wang, Wei and Xu, Guohai and Zhang, Ji and Huang, Songfang and Huang, Fei and Zhou, Jingren},
  title={mPLUG-2: A Modularized Multi-modal Foundation Model Across Text, Image and Video},
  booktitle={Proceedings of the 40th International Conference on Machine Learning},
  volume={202},
  pages={38728--38748},
  year={2023}
}
```

</details>