README.md

---
license: apache-2.0
---

# GlyphByT5 Pretraining


This is the pretraining data for [Glyph-ByT5](https://glyph-byt5.github.io/).

## Dataset Details

> [**Glyph-ByT5: A Customized Text Encoder for Accurate Visual Text Rendering**](https://glyph-byt5.github.io/)            
> [Zeyu Liu](https://github.com/lzy-tony), [Weicong Liang](https://scholar.google.com/citations?user=QvHDIygAAAAJ&hl=zh-CN), [Zhanhao Liang](https://github.com/RockeyCoss), [Chong Luo](https://www.microsoft.com/en-us/research/people/cluo/), [Ji Li](https://sites.google.com/a/usc.edu/jili/), [Gao Huang](https://www.gaohuang.net/), [Yuhui Yuan](https://www.microsoft.com/en-us/research/people/yuyua/)
> Microsoft Research Asia; Tsinghua University; Peking University; The Australian National University  
> Preprint

### Dataset Description

- **Repository:** [https://github.com/AIGText/Glyph-ByT5]
- **Paper:** [https://arxiv.org/abs/2403.09622]
- **Project Page:** [https://glyph-byt5.github.io/]

## Dataset Structure

This dataset contains a json file containing the annotation file needed for Glyph-ByT5 pretraining. The json file is organized as a list of dicts. Each dict contains the following fields:

**Note: As font family and color are randomly selected during pretraining, 'color' and 'font-family' fields are None and later filled randomly in the corresponding dataset code.**

Currently, we release a subset of fonts that we use, containing 100 with free commercial use license, see [repo](https://github.com/AIGText/Glyph-ByT5) for more details.

```
{
  'texts': [
    text_1,
    ...
    text_k,
  ],
  'styles': [
    {
      'color': color_1,
      'font-family': 'font_1',
    },
    ...
    {
      'color': color_k,
      'font-family': 'font_k',
    },
  ],
  'bbox': [
    [x1, y1, w1, h1, a1],
    ...
    [xk, yk, wk, hk, ak],
  ]
}
```