JIA244601/Handwritten-Historical-Archive-Image-Dataset-of-Modern-China_1840_1949

📜 Chinese Modern Era (1840–1949) Handwritten Historical Archive Dataset

中国近代史 (1840–1949) 手写历史档案数据集

---

📖 Dataset Description | 数据集描述

🎯 Purpose & Motivation | 目的与动机

To address the recognition difficulties and generalization bottlenecks faced by existing Optical Character Recognition (OCR) models when processing handwritten historical archives from modern Chinese history (1840–1949), a joint student research team from Capital Normal University (China) constructed and released this dataset. This dataset aims to significantly enhance model robustness against modern handwritten Chinese character variants, strike-through corrections, and mixed horizontal/vertical layouts through a combination of synthetic pre-training and real-world domain adaptation fine-tuning.

针对现有光学字符识别（OCR）模型在处理中国近代史（1840–1949 年）手写历史档案时面临的识别困难与泛化瓶颈，来自首都师范大学（中国）的联合研究团队（学生）构建并发布了本数据集。本数据集旨在通过合成预训练与真实场景微调相结合的方式，显著提升模型对于近代手写汉字变体、涂抹修正及混合排版的鲁棒性。

---

🗂️ Dataset Composition | 数据构成

The dataset consists of two complementary parts, covering the entire pipeline from large-scale synthetic pre-training to fine-grained real-world fine-tuning.
本数据集由两个互补部分构成，涵盖从大规模合成预训练到真实场景精细微调的全流程需求。

1. 🔬 Synthetic Pre-training Dataset | 合成预训练数据集

Volume: 53,500 images, containing a total of over 2.35 million characters.
Generation Logic: Directly generated from modern Chinese historical texts, comprehensively simulating the characteristics of real handwritten archives using Python scripts.
Image Features:
- Layout: Randomly assigned horizontal or vertical writing direction (50% ratio each).
- Cropping: Images are variably sized and cropped at the sentence level.
- Fonts: Random combination of 8 distinct imitation-handwriting fonts.
- Character Set: Mixed Chinese and foreign languages; mixed traditional and simplified Chinese characters.
- Noise Simulation: To simulate typos and corrections found in authentic manuscripts, random noise is systematically added at a rate of 1%.

数据量：53,500 张图像，共计超过 235 万 字符。
生成逻辑：直接源自中国近代史文献语料，利用 Python 脚本全真模拟真实历史档案中的手写体特征。
图像特征：
- 版面布局：随机设定横排与竖排书写方向（各占 50% 比例）。
- 切割方式：图像尺寸不固定，以句子级别进行切割。
- 字体混合：随机组合 8 种 仿手写体字体。
- 字符集：中英文混杂，繁简体汉字混合。
- 噪声模拟：为模拟真实手稿中的笔误与涂改痕迹，系统以 1% 的概率随机添加噪声干扰。

2. 🏛️ Real-world Domain Adaptation Fine-tuning Dataset | 真实域适应微调数据集

Volume: 9,452 images.
Source: Directly derived from authentic archival compilations, spanning from June 11, 1919, to February 25, 1949, covering handwritten archival images from various periods and multiple organizations.
Annotation Protocol & Workflow:
1. Tool: PaddleLabel annotation platform.
2. Pre-annotation: Initial recognition performed by the built-in PP-OCRv3 model.
3. Manual Proofreading: Meticulous character-by-character verification conducted by annotators proficient in modern Chinese history.
4. Handling Rules:
- Illegibility/Occlusion: Text completely illegible, over 15% damaged, or obscured is uniformly annotated as `*` and treated as noise.
- Typographical Errors: If the intended correct form can be determined, it is directly annotated in its correct form; otherwise classified as noise.
- Variant Characters: Proofread according to the Dictionary of Variant Characters and annotated directly as standard orthographic characters.
5. Output: Post-annotation, images and corresponding annotation results are cropped into small single-line or single-column patches.

数据量：9,452 张图像。
数据来源：直接截取自真实档案汇编，时间跨度为 1919 年 6 月 11 日至 1949 年 2 月 25 日，涵盖多个时期与组织机构的手写档案影像。
标注规范与流程：
1. 工具：采用 PaddleLabel 标注平台。
2. 预标注：内置 PP-OCRv3 模型进行初步识别。
3. 人工精校：由熟悉中国近代史的标注员进行逐字校对。
4. 特殊处理规则：
- 污损遮挡：完全无法辨认、破损面积超过 15% 或被遮挡的文字，统一标注为 `*` 并视为噪声。
- 笔误修正：若能判定书写者意图的正确字形，则直接标注为正确字；否则归为噪声。
- 异体字处理：依据《异体字字典》进行校对，直接标注为规范正体字。
5. 输出格式：标注完成后，图像与对应的标注结果被切割为单行或单列的小尺寸切片。

---

⚠️ Usage Instructions & Disclaimer | 使用须知与免责声明

License: This dataset is released under the MIT License. In no event shall the data authors be liable for any claim, damages, or other liability arising from or in connection with the use of this data, whether in contract, tort, or otherwise.

许可证：本数据集遵循 MIT License 开源协议。数据作者在任何情况下均不对因使用或无法使用本数据而导致的任何索赔、损害或其他责任负责，无论该责任是基于合同、侵权或其他法律理论。

This repository provides the original archival images (both real and synthetic), the raw annotation files exported from the PaddleLabel platform, and the annotation files in JSONL format. Please download and use them as needed.
本仓库提供数据集的原始档案图像（包含真实数据与合成数据）、PaddleLabel 平台导出的原始标注文件以及便于处理的 JSONL 格式标注文件，请按需下载使用。

🔔 Important Notes | 重要提示

1. Before using the dataset, you may need to modify the image path configurations within the dataset files.
   使用数据集前，您可能需要根据本地环境修改数据集内部的图像路径配置。
2. After downloading, please refrain from changing folder names or image file names (codes). The image file names maintain a one-to-one correspondence with the entries in the annotation files; alteration will invalidate the annotations.
   下载后请尽量避免更改文件夹名称或图像名称（编码），尤其是图像文件名。图像文件名与标注文件中的信息存在 一一对应 的映射关系，改动将导致标注失效。

---

📚 Citation | 引用格式

If this dataset is helpful for your research, please cite it using the following format:
若本数据集对您的研究有所帮助，请按以下格式引用：

English Citation (BibTeX)

@misc{cnu_chinese_modern_handwritten_archive_1840_1949,
  title        = {Chinese Modern Era (1840–1949) Handwritten Historical Archive Dataset},
  author       = {Jingkai Jia, Jiayi Zhang, Ruizhu Li, Zichong Tian},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{}},
  note         = {MIT Licensed. Contains 53.5k synthetic images and 9.4k real archival images for OCR fine-tuning.},
}