---
license: apache-2.0
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: filename
    dtype: string
  - name: difficulty
    dtype: string
  - name: svg_code
    dtype: string
  splits:
  - name: train
    num_bytes: 1394385
    num_examples: 300
  - name: easy
    num_bytes: 105183
    num_examples: 100
  - name: medium
    num_bytes: 501313
    num_examples: 100
  - name: hard
    num_bytes: 787889
    num_examples: 100
  download_size: 1529634
  dataset_size: 2788770
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: easy
    path: data/easy-*
  - split: medium
    path: data/medium-*
  - split: hard
    path: data/hard-*
language:
- en
pretty_name: SVGenius
size_categories:
- n<1K
tags:
- svg-understanding
- svg-editing
- svg-generation
---

# **SVGenius: Benchmarking LLMs in SVG Understanding, Editing and Generation**

<!-- Provide a quick summary of the dataset. -->
<a href="https://arxiv.org/abs/2506.03139" target="_blank">
    <img alt="arXiv" src="https://img.shields.io/badge/arXiv-SVGenius-red?logo=arxiv" height="20" />
</a>
<a href="https://huggingface.co/datasets/xiaoooobai/SVGenius" target="_blank">
    <img alt="SVGenius" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Benchmark-SVGenius-ffc107?color=ffc107&logoColor=white" height="20" />
</a>
<a href="https://zju-real.github.io/SVGenius/" target="_blank">
    <img alt="Webpage" src="https://img.shields.io/badge/%F0%9F%8C%8E_Website-SVGenius-green.svg" height="20" />
</a>

We introduce SVGenius, the first large-scale complexity-stratified benchmark accessing (M)LLMs capabilities across three progressive dimensions: **Understanding** (perceptua and semantic QA), **Editing** (bug fixing, code optimization, style editing), and **Generation** (text-to-SVG, image-to-SVG, style transfer).  Built on real-world data from **24** application domains with systematic complexity stratification, SVGenius evaluates models through **8** task categories and **18** metrics. We assess **22** mainstream models spanning different scales, architectures, training paradigms, and accessibility levels.
<img alt="SVGenius Dataset Overview" 
     src="https://cdn.jsdelivr.net/gh/ZJU-REAL/SVGenius@main/docs/static/images/overview.jpg" 
     style="width: 100%; max-width: 1000px;" />


The dataset contains the following fields:
| Field Name | Description |
| :--------- | :---------- |
| `id` | Unique identifier for each SVG icon sample |
| `filename` | Original filename of the SVG file, preserving the source naming convention with category and identifier information |
| `difficulty` | Complexity level of the SVG icon, includes 3 distinct categories: `easy`, `medium`, and `hard` for evaluating different levels of SVG processing capabilities |
| `svg_code` | Complete SVG markup code containing the vector graphics definition, including all paths,styles, and attributes |
- **Language(s) (NLP):** en, zh
- **License:** mit


## Uses

```py
from datasets import load_dataset
ds = load_dataset("xiaoooobai/SVGenius")
```


## Citation

```bibtex
@misc{chen2025svgeniusbenchmarkingllmssvg,
      title={SVGenius: Benchmarking LLMs in SVG Understanding, Editing and Generation}, 
      author={Siqi Chen and Xinyu Dong and Haolei Xu and Xingyu Wu and Fei Tang and Hang Zhang and Yuchen Yan and Linjuan Wu and Wenqi Zhang and Guiyang Hou and Yongliang Shen and          Weiming Lu and Yueting Zhuang},
      year={2025},
      eprint={2506.03139},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.03139}, 
}
```