PPTAnimation_Test Dataset Card

1. Overview

PPTAnimation_Test comprises 1,000 synthetic short videos (< 15 s each) of PowerPoint slide animations, paired one-to-one with human-written captions.

• Videos/: MP4 clips named video_0001.mp4 … video_1000.mp4.
• Captions/: Plain-text files (video_0001.txt, etc.) that store the ground-truth natural-language descriptions of each video.

The dataset is ideal for vision–language model (VLM) fine-tuning or evaluation, and it also supports tasks such as video captioning and video understanding.

2. Directory Structure

PPTAnimation_Test/
├── Videos/
│   ├── video_0001.mp4
│   ├── video_0002.mp4
│   └── ...
└── Captions/
    ├── video_0001.txt
    ├── video_0002.txt
    └── ...

3. Tasks & Applications

• Vision–language alignment: video–text retrieval, cross-modal understanding

• Video caption generation: produce textual descriptions from animation videos

• VLM fine-tuning / benchmarking: assess a model’s ability to understand PPT animations

4. License

This dataset is released under the Apache License 2.0. Please comply with its terms of use.

5. Citation

If you use the dataset in academic work, please cite the following paper:

@misc{jiang2025animationneedsattentionholistic,
      title        = {Animation Needs Attention: A Holistic Approach to Slides Animation Comprehension with Visual-Language Models},
      author       = {Yifan Jiang and Yibo Xue and Yukun Kang and Pin Zheng and Jian Peng and Feiran Wu and Changliang Xu},
      year         = {2025},
      eprint       = {2507.03916},
      archivePrefix= {arXiv},
      primaryClass = {cs.AI},
      url          = {https://arxiv.org/abs/2507.03916},
}

6. Usage Example

from modelscope.msdatasets import MsDataset

dataset = MsDataset.load(
    dataset_name='jyf9774/PPTAnimation_Test',
    namespace='jyf9774',
    split='train'  # no official split; use 'train' or None
)

sample = dataset[0]
print(sample['text'])        # Caption text
sample['video'].display()    # Preview the video in a notebook