VFIG-Bench Dataset Card

Overview

VFIG-Bench is the evaluation benchmark accompanying VFIG-Data, used to assess vision-language models on the task of vectorizing complex academic figures into SVG. It comprises two complementary test sets:

• VFIG-Bench — 400 in-distribution figure–SVG pairs (held-out from VFIG-Data).
• VFIG-Bench-OOD — 198 out-of-distribution figures (image-only, no SVG ground truth) sourced from well-known academic figures, used to probe generalization.

Config	Split	Rows	Schema
VFIG-Bench	test	400	{filename: string, svg: string}
VFIG-Bench-OOD	test	198	{filename: string, image: Image}

• License: ODC-BY 1.0
• Paper: VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models

Usage

from datasets import load_dataset

# In-distribution test (figure-SVG pairs)
ds = load_dataset("QijiaHe/VFIG-Bench", "VFIG-Bench", split="test")
print(ds[0]["filename"], len(ds[0]["svg"]))

# Out-of-distribution test (image only)
ds_ood = load_dataset("QijiaHe/VFIG-Bench", "VFIG-Bench-OOD", split="test")
print(ds_ood[0]["filename"], ds_ood[0]["image"].size)

Notes on the OOD config

The VFIG-Bench-OOD config contains only PNG images — there is no SVG ground truth. It is intended for evaluating a model's ability to vectorize unseen, real-world academic figures.

License

This dataset is licensed under ODC-BY 1.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Citation

@misc{he2026vfigvectorizingcomplexfigures,
      title={VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models},
      author={Qijia He and Xunmei Liu and Hammaad Memon and Ziang Li and Zixian Ma and Jaemin Cho and Jason Ren and Daniel S Weld and Ranjay Krishna},
      year={2026},
      eprint={2603.24575},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.24575},
}