---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: question
    dtype: string
  - name: text
    dtype: string
  - name: is_vertical
    dtype: bool
  - name: num_columns
    dtype: int32
  - name: noise_type
    dtype: int32
  splits:
  - name: train
    num_bytes: 28489001997
    num_examples: 17952
  download_size: 28461996472
  dataset_size: 28489001997
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- image-text-to-text
language:
- ja
size_categories:
- 10K<n<100K
license: cc-by-4.0
tags:
- ocr
---

# Synth-JDoc

[Paper](https://link.springer.com/chapter/10.1007/978-3-032-36033-5_15) | [Code](https://github.com/llm-jp/synth-jdoc)

**Synth-JDoc** is a dataset of synthetic Japanese document images generated using HTML/CSS.

We generate embedded images, captions, and titles from prepared text, and use these elements to synthesize document images featuring diverse multi-column layouts in both vertical and horizontal writing.

Because the document images are synthesized directly from text, this dataset is completely free from OCR errors.

![Construction Pipeline](./figs/overview_method.jpg)


## Dataset details

- `id`
  - Image ID
- `image`
  - PIL Image
- `question`
  - Prompt for VLM
- `text`
  - The text written within the synthetic document image
- `is_vertival`
  - `True` if the layout uses vertical writing
- `num_columns`
  - Number of columns
- `noise_type`
  - 0: No noise, 1: scan like noise, 2: Noise by [Augraphy](https://github.com/sparkfish/augraphy)


This dataset was introduced in our paper: [Synth-JDoc: Synthesizing a Japanese Document Image Dataset for OCR with Diverse Layouts and Embedded Images](https://link.springer.com/chapter/10.1007/978-3-032-36033-5_15).

The code used to generate this dataset is available at our GitHub repository: [https://github.com/llm-jp/synth-jdoc](https://github.com/llm-jp/synth-jdoc)

Before releasing the dataset, we manually checked images flagged by an NSFW image detection model and removed those confirmed to be NSFW.
Therefore, please note that this dataset is not identical to the one used for the experiments in our paper.
Additionally, please be aware that this NSFW filtering process is not perfect.

## License

This dataset is released under [Creative Commons Attribution 4.0 License](https://creativecommons.org/licenses/by/4.0/legalcode).

## Citation
```
@InProceedings{10.1007/978-3-032-36033-5_15,
author="Sasagawa, Keito
and Kurita, Shuhei
and Kawahara, Daisuke",
editor="Fink, Gernot A.
and Forn{\'e}s, Alicia
and Kise, Koichi
and Lopresti, Daniel",
title="Synth-JDoc: Synthesizing a Japanese Document Image Dataset for OCR with Diverse Layouts and Embedded Images",
booktitle="Document Analysis and Recognition -- ICDAR 2026",
year="2027",
publisher="Springer Nature Switzerland",
address="Cham",
pages="245--260",
abstract="The ability of Large Vision Language Models (LVLMs) to read text within document images is crucial, as it enables various applications such as Document Visual Question Answering. To enhance the text-reading capabilities of LVLMs, high-quality OCR datasets are essential. This need is particularly critical for Japanese documents, which often feature vertically written text alongside horizontally written text. Current LVLMs demonstrate considerably lower performance on vertically written Japanese text than on horizontally written text, necessitating specialized OCR datasets to bridge this gap. However, manually constructing OCR datasets is expensive and difficult to scale. Alternatively, constructing datasets by extracting text from existing document images using OCR models introduces challenges, such as text recognition errors and the prerequisite of sourcing document images.",
isbn="978-3-032-36033-5"
}
```