---
license: apache-2.0
task_categories:
- image-to-text
- visual-question-answering
language:
- zh
- en
tags:
- art
size_categories:
- 1K<n<10K
---



# 🧠 CalliReader: Contextualizing Chinese Calligraphy via an Embedding-aligned Vision Language Model
<div align="center">
  <a href="https://github.com/LoYuXr/CalliReader">📂 Code</a>
  <a href="https://arxiv.org/pdf/2503.06472">📄 Paper</a>
</div>

**CalliBench** is aimed to comprehensively evaluate VLMs' performance on the recognition and understanding of Chinese calligraphy. 



## 📦 Dataset Summary

* **Samples**: 3,192 image–annotation pairs
* **Tasks**: **Full-page recognition** and **Contextual VQA** (choice of author/layout/style, bilingual interpretation, and intent analysis).
* **Annotations**:
  
  * Metadata of author, layout, and style.
  * Fine-grained annotations of **character-wise bounding boxes and labels**.
  * Certain samples include **contextual VQA**.

## 🧪 How To Use
All **.parqeut** files of different tiers can be found in the sub-folders of **data**. **Pandas** can be used to parse and further process those files.

For example, to load a sample and convert its image into a .jpg file: 
```
import pandas as pd
import io
from PIL import Image
df = pd.read_parquet('./data/train.parquet')
image_data = df.iloc[0]['image'] 
image = Image.open(io.BytesIO(image_data['bytes']))
image.save('output_image.jpg') 
```

## 🤗 License

Apache 2.0 – open for research and commercial use.