---
title: TASLE
canonical_url: "https://www.modelscope.cn/datasets/L67plus/TASLE"
md_url: "https://www.modelscope.cn/datasets/L67plus/TASLE.md"
repository: L67plus/TASLE
last_updated: 2026-08-21
license: "Apache License 2.0"
storage_size: "56 GB"
downloads: 318160
stars: 0
---

# TASLE

> TASLE - L67plus 在 ModelScope 开源的数据集。TASLE is a dataset for detecting and temporally localizing AIGC-manipulated segments in long videos. It contains real and manipulated videos, temporal boundaries of manipulated segments, and related semantic annotations.…

L67plus/TASLE 是 ModelScope 魔搭社区上的数据集，存储大小 56 GB，采用 Apache License 2.0 许可。

- **Repository**: L67plus/TASLE
- **License**: Apache License 2.0
- **Storage size**: 56 GB
- **Downloads**: 318160
- **Stars**: 0
- **Last updated**: 2026-08-21

Source: https://www.modelscope.cn/datasets/L67plus/TASLE

---

# TASLE

[![Project Page](https://img.shields.io/badge/Project-Page-3273DC)](https://debby-0527.github.io/TASLE/)
[![MSLoc Code](https://img.shields.io/badge/GitHub-MSLoc-181717?logo=github)](https://github.com/67Pluuus/MSLoc)

TASLE is a dataset for detecting and temporally localizing AIGC-manipulated segments in long videos. It contains real and manipulated videos, temporal boundaries of manipulated segments, and related semantic annotations. The dataset supports video authenticity detection, manipulated-segment localization, and temporal boundary analysis.

## Dataset Splits

| Annotation file | Samples | Description |
| --- | ---: | --- |
| `train.json` | 11,179 | Default training split: 5,506 manipulated and 5,673 real videos |
| `test.json` | 1,293 | Default test split: 631 manipulated and 662 real videos |
| `test_add.json` | 1,442 | Optional OOD-extended test split; not used by the default evaluation |
| `test_segment_classification.json` | 1,000 | Optional annotations for segment-level classification evaluation |
| `train_trace_full_video.json` | 11,179 | Optional full-video temporal localization annotations |

The default training and evaluation workflow uses `train.json` and `test.json`.

## Directory Structure

Download the repository contents into a directory named `MSLoc_assets` so that the released code can resolve the default paths directly:

```text
MSLoc_assets/
├── data/Tasle-CoT-10K/
│   ├── annos/                 # JSON annotations
│   └── videos/                # Video files
├── DeMamba/                   # Stage-1 pretrained weights and released results
└── Trace/                     # Stage-2 pretrained weights and released assets
```

## Annotation Format

The main annotation fields are shown below:

```json
{
  "source": "video source",
  "video_path": "relative/path/to/video.mp4",
  "duration": 10.0,
  "type": "real or fake",
  "annotations": [
    {
      "segment": [1.2, 6.8],
      "label": "object category",
      "model": "generation model"
    }
  ]
}
```

`segment` is expressed in seconds and specifies the start and end times of a manipulated segment. For real videos, `annotations` is usually an empty list.

## Download

```bash
pip install modelscope
modelscope download --dataset L67plus/TASLE --local_dir ./MSLoc_assets
```

The dataset is large, so please ensure that sufficient disk space is available before downloading it.

## Project Resources

- [Project page](https://debby-0527.github.io/TASLE/)
- [MSLoc code](https://github.com/67Pluuus/MSLoc)

Clone the code into a directory named `MSLoc` and place it beside the downloaded `MSLoc_assets` directory:

```text
workspace/
├── MSLoc/
└── MSLoc_assets/
```

The directory names are intentional: `MSLoc` contains the source code, while `MSLoc_assets` contains the dataset and released checkpoints. Please refer to the README in the MSLoc repository for environment setup, inference instructions, and the expected file paths.

## License

This repository is released under the Apache License 2.0. Third-party source videos remain subject to the licenses and terms of use of their respective original datasets.