Datasets:

Forceless
/

Zenodo10K

task_categories:
  - image-text-to-text
license: other
dataset_info:
  features:
    - name: filename
      dtype: string
    - name: size
      dtype: int64
    - name: url
      dtype: string
    - name: license
      dtype: string
    - name: title
      dtype: string
    - name: created
      dtype: string
    - name: updated
      dtype: string
    - name: doi
      dtype: string
    - name: checksum
      dtype: string
  splits:
    - name: pptx
      num_bytes: 3925161
      num_examples: 10448
  download_size: 2028492
  dataset_size: 3925161
configs:
  - config_name: default
    data_files:
      - split: pptx
        path: data/pptx-*

PPTAgent/Zenodo10K

This is the dataset used in PPTAgent: Generating and Evaluating Presentations Beyond Text-to-Slides and also used for evaluation in VLM-SlideEval: Evaluating VLMs on Structured Comprehension and Perturbation Sensitivity in PPT. It was crawled from zenodo. To the best of our knowledge, it is the largest presentation dataset currently available, comprising over 10,000 PowerPoint (.pptx) files, all distributed under a clear and compliant license.

For more information regarding the PPTAgent project, please visit its github repo.

VLM-SlideEval Abstract

Vision-language models (VLMs) are increasingly used to evaluate multimodal content, including presentation slides, yet their slide-specific understanding remains underexplored {despite their growing role as critics in agentic, model-forward pipelines}. We introduce VLM-SlideEval, an evaluation framework that probes VLMs along three axes: (1) element-level extraction from slide images aligned to ground truth; (2) robustness to controlled perturbations in geometry, style, and text; and (3) higher-level comprehension, such as recovering a deck's narrative order from shuffled slides. Using publicly available decks from Zenodo ( this https URL ), we standardize ground-truth element metadata from PowerPoint XML and live renderings into a unified, verifiable schema. Empirically, VLMs underperform on pixel-accurate extraction and show non-trivial agreement, fidelity, and consistency under controlled perturbations, while performing better on single-slide content understanding; however, they do not reliably capture narrative structure across slides. These results highlight the limits of current VLMs for slide evaluation and motivate calibrated, critic-in-the-loop evaluators that drive iterative refinement and selection in agentic pipelines.

Sample Usage

dirname = f"zenodo-pptx/pptx/{task['license']}/{task['created'][:4]}/"
basename = f"{task['checksum'][4:]}-{task['filename']}"
filepath = dirname + basename
try:
    open('/tmp/'+basename,'wb').close()
except:
    filepath = dirname + basename[:240] + ".pptx"

Citation

If you find this project helpful, please use the following to cite it:

@misc{zheng2025pptagentgeneratingevaluatingpresentations,
      title={PPTAgent: Generating and Evaluating Presentations Beyond Text-to-Slides}, 
      author={Hao Zheng and Xinyan Guan and Hao Kong and Jia Zheng and Hongyu Lin and Yaojie Lu and Ben He and Xianpei Han and Le Sun},
      year={2025},
      eprint={2501.03936},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2501.03936}, 
}

If you use this dataset for VLM evaluation, please also cite:

@article{xu2025vlm,
  title={VLM-SlideEval: Evaluating VLMs on Structured Comprehension and Perturbation Sensitivity in PPT},
  author={Xu, Charles and Li, Qiyang and Luo, Jianlan and Levine, Sergey},
  journal={arXiv preprint arXiv:2510.22045},
  year={2025}
}