MTHv2_Plus / README.md
AnonymousGAOSA's picture
Upload 4 files
0bc48c2 verified
|
Raw History Blame Contribute Delete
5.58 kB
---
license: cc-by-nc-sa-4.0
language:
- lzh
- zh
- en
pretty_name: MTHv2_Plus - VLM SFT data for vertical-layout historical Chinese OCR
tags:
- ocr
- historical-documents
- tripitaka
- buddhist-texts
- vision-language-model
- sft
- vertical-text
size_categories:
- 1K<n<10K
task_categories:
- image-to-text
- text-generation
---
# MTHv2_Plus — VLM SFT Data for Vertical-Layout Historical Chinese OCR
**中文说明见下半部分 | Chinese documentation follows the English section below.**
---
## English
Message-level SFT (supervised fine-tuning) data for training **vision-language
models** to transcribe vertical-layout historical Chinese documents — built on
top of **MTHv2** (Tripitaka Koreana in Han, TKH + Multiple Tripitaka in Han,
MTH; HCIILAB, South China University of Technology; ICFHR 2020).
This repository contains **annotations only** — no images. Each sample is one
page: a fixed user instruction with an `<image>` placeholder, and the expected
full-page transcription as the assistant response.
## Data format
Single JSON file `1016v2_TKHMTH2200.json`, a JSON array of 2,919 entries:
```json
[
{
"messages": [
{ "role": "user", "content": "<image>\n<transcription instructions>" },
{ "role": "assistant", "content": "full-page transcription, columns separated by \\n" }
],
"images": ["TKHMTH2200/<subset>/img/<name>.png"]
}
]
```
- One page image per entry, 2,919 total: 731 `.png` + 2,188 `.jpg`, spanning
three MTHv2 subsets — **TKH 999 / MTH1000 846 / MTH1200 1,074 pages**
(reflected in the `images` path prefix: `TKHMTH2200/TKH|MTH1000|MTH1200/img/<name>`).
- `user` content — one fixed instruction (identical across all 2,919 entries)
describing the OCR contract: read columns **right-to-left, top-to-bottom**;
keep original punctuation, titles and interlinear notes (`[...]`); skip
page furniture; output `□` for illegible characters, `<SKIP>` for empty
pages.
- `assistant` content — the ground-truth transcription, 9–1,424 characters
(median 338), unique per page (no duplicates).
- Image basenames are the **original MTHv2 filenames** (e.g.
`07-V011P0015.png`, `QL_25_524.jpg`, `0001_007_26_01.jpg`), so annotations
join 1:1 with the upstream image set — no mapping file needed.
## Getting the images
Download MTHv2 from the
[official release repository](https://github.com/HCIILAB/MTHv2_Datasets_Release)
(Google Drive / Baidu Drive links in its README), then join by basename:
the `images` path prefix (`TKHMTH2200/<subset>/img/`) is a relative
layout hint only — rewrite it to wherever you placed the upstream files.
## Files
| File | Description |
|---|---|
| `1016v2_TKHMTH2200.json` | SFT message data, 2,919 entries |
| `LICENSE` | CC-BY-NC-SA-4.0 |
| `NOTICE.md` | upstream attribution and usage notes |
| `README.md` | this card |
## Licensing
- **Upstream images**: MTHv2 is "free to the academic community for research
purpose usage only" (no OSI/SPDX license).
- **These annotations**: **CC-BY-NC-SA-4.0** — chosen to stay within the
upstream restriction: **academic research, non-commercial use only,
share-alike**. For commercial use, contact the MTHv2 authors
(scutmaweihong@gmail.com / eelwjin@scut.edu.cn per the upstream README).
## Citation
```bibtex
@inproceedings{ma2020joint,
title={Joint Layout Analysis, Character Detection and Recognition for Historical Document Digitization},
author={Ma, Weihong and Zhang, Hesuo and Jin, Lianwen and Wu, Sihang and Wang, Jiapeng and Wang, Yongpan},
booktitle={International Conference on Frontiers in Handwriting Recognition (ICFHR)},
year={2020}
}
```
Please also cite this repository when you use the annotations.
---
## 中文说明
基于 **MTHv2**(TKH 高丽大藏经 + MTH 多种大藏经;华南理工大学 HCIILAB;ICFHR 2020)构建的**视觉语言模型(VLM)SFT 微调数据**,用于竖排历史文档整页 OCR。本仓库**仅含标注,不含图片**。
### 数据格式
单文件 `1016v2_TKHMTH2200.json`,JSON 数组,共 2,919 条,结构同上方英文部分(`messages` + `images`,每条一页一图)。
- 共 2,919 页:731 张 `.png` + 2,188 张 `.jpg`,覆盖三个子集——**TKH 999 / MTH1000 846 / MTH1200 1,074 页**。
- `user` 指令:全数据集**唯一一条固定指令**(竖排从右到左、每列自上而下;保留标点/标题/夹注;忽略版心等页面构件;不可辨识字 `□`、无字页 `<SKIP>`)。
- `assistant`:整页转写真值,9–1,424 字(中位数 338),逐页唯一。
- 图片文件名即 **MTHv2 原始文件名**,可与上游图片直接按名对齐,无需映射文件。
### 获取图片
从[官方发布仓库](https://github.com/HCIILAB/MTHv2_Datasets_Release)下载(内有 Google Drive / 百度网盘链接),按文件名对齐即可;JSON 内的 `images` 路径前缀(`TKHMTH2200/<子集>/img/`)只是相对布局提示,请改写为你的实际路径。
### 文件
| 文件 | 说明 |
|---|---|
| `1016v2_TKHMTH2200.json` | SFT 消息数据,2,919 条 |
| `LICENSE` | CC-BY-NC-SA-4.0 |
| `NOTICE.md` | 上游署名与使用说明 |
| `README.md` | 本说明 |
### 许可
- 上游图片:仅限学术研究用途。
- 本标注集:**CC-BY-NC-SA-4.0**——为符合上游限制,仅限非商业、相同方式共享的学术研究使用;商用请联系 MTHv2 原作者(scutmaweihong@gmail.com / eelwjin@scut.edu.cn)。
### 引用
请同时引用上游论文与本仓库(BibTeX 见上方英文部分)。