AncientDoc_Plus / README.md
AnonymousGAOSA's picture
Upload 5 files
8d0d861 verified
|
Raw History Blame Contribute Delete
6.86 kB
metadata
license: cc-by-4.0
language:
  - lzh
  - zh
  - en
pretty_name: AncientDoc_Plus - VLM SFT data for vertical-layout ancient Chinese OCR
tags:
  - ocr
  - ancient-chinese
  - classical-chinese
  - historical-documents
  - vision-language-model
  - sft
  - vertical-text
size_categories:
  - 1K<n<10K
task_categories:
  - image-to-text
  - text-generation

AncientDoc_Plus — VLM SFT Data for Vertical-Layout Ancient 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 ancient Chinese books — built on top of the AncientDoc benchmark (ByteDance, upstream license CC0-1.0).

This repository contains annotations only — no images. Each sample is one page: a fixed system-style user instruction with an <image> placeholder, and the expected full-page transcription as the assistant response.

Data format

Single JSON file 1016v2_AncientDoc.json, a JSON array of 2,198 entries:

[
  {
    "messages": [
      { "role": "user",      "content": "<image>\n<transcription instructions>" },
      { "role": "assistant", "content": "full-page transcription, columns separated by \\n" }
    ],
    "images": ["AncientDoc/<书名>page_NN.png"]
  }
]
  • images — exactly one page image per entry (2,198 total, all .png).
  • user content — one fixed instruction (identical across all 2,198 entries) describing the OCR contract: read columns right-to-left, top-to-bottom; keep original punctuation, titles and interlinear notes ([...]); skip page furniture (版心/页眉/页码/印章 etc.); output □ for illegible characters, <SKIP> for empty pages.
  • assistant content — the ground-truth transcription, 31–792 characters (median 301), unique per page (no duplicates).

Getting the images

The images paths in the JSON are relative flat names (AncientDoc/<书名>page_NN.png); the actual page images live in the upstream CC0 repository under a two-level category layout (imgs/<category>/<book>/page_NN.png). A mapping file source_path_map.json is provided (see Files below) that resolves every flat name to its exact upstream path: {"<书名>page_NN.png": "imgs/<category>/<book>/page_NN.png"}. All 2,198 mappings were verified 1:1 against the upstream repository tree (14 categories, e.g. 诗文评类/艺术类/别集类/兵家类/儒家类 …).

Download the upstream images and rewrite the images field accordingly:

import json
m = json.load(open("source_path_map.json", encoding="utf-8"))
# e.g. "楚辭新註page_17.png" -> "imgs/楚辞类/楚辭新註/page_17.png"
pip install -U huggingface_hub
python -c "from huggingface_hub import snapshot_download; snapshot_download('ByteDance/AncientDoc', repo_type='dataset', local_dir='AncientDoc')"

Files

File Description
1016v2_AncientDoc.json SFT message data, 2,198 entries
source_path_map.json flat filename → upstream imgs/<category>/<book>/page_NN.png (2,198 verified mappings)
LICENSE CC-BY-4.0 (this annotation set)
NOTICE.md upstream attribution and usage notes
README.md this card

Why "Plus"

The upstream AncientDoc benchmark ships images plus label.csv, but is not packaged for VLM fine-tuning. We contribute: (a) full-page transcription annotations in a ready-to-use messages/images SFT layout with a fixed OCR instruction prompt; (b) a verified path map so the annotations can be joined back to upstream images regardless of local directory layout.

Licensing

  • Upstream images & benchmark: ByteDance/AncientDoc, CC0-1.0.
  • These annotations (SFT data + path map): CC-BY-4.0 — our own original transcriptions of public-domain ancient texts; the CC0 upstream imposes no additional restriction. Attribution: link to this repository.

Citation

@misc{ancientdoc2024,
  title={AncientDoc: A Benchmark for Chinese Ancient Document Understanding},
  author={{ByteDance}},
  howpublished={\url{https://huggingface.co/datasets/ByteDance/AncientDoc}},
  note={CC0}
}

Please also cite this repository when you use the annotations.


中文说明

基于 AncientDoc 基准(字节跳动,上游许可 CC0-1.0)构建的视觉语言模型(VLM)SFT 微调数据,用于训练竖排古籍整页 OCR。本仓库仅含标注,不含图片。

数据格式

单文件 1016v2_AncientDoc.json,JSON 数组,共 2,198 条:

[
  {
    "messages": [
      { "role": "user",      "content": "<image>\n<转写规则说明>" },
      { "role": "assistant", "content": "整页转写文本,列间以 \\n 分隔" }
    ],
    "images": ["AncientDoc/<书名>page_NN.png"]
  }
]
  • images:每条恰好一张页面图(共 2,198 张,全部 .png)。
  • user 指令:全数据集唯一一条固定指令,约定 OCR 契约——普通正文从右到左、每列自上而下;保留原标点、标题与双列夹注([...]);忽略版心/页眉页脚/页码/印章等;模糊不可辨识字输出 □,无字页输出 <SKIP>。
  • assistant:整页转写真值,31–792 字(中位数 301),逐页唯一,无重复。

获取图片

JSON 中的 images 是相对扁平名(AncientDoc/<书名>page_NN.png),而上游 CC0 仓库的实际目录为 imgs/<类别>/<书名>/page_NN.png。为此提供 **source_path_map.json**(2,198 条映射,已逐一与上游目录树核对): {"<书名>page_NN.png": "imgs/<类别>/<书名>/page_NN.png"}。下载上游图片后按 映射改写 images 字段即可对齐。

文件

文件 说明
1016v2_AncientDoc.json SFT 消息数据,2,198 条
source_path_map.json 扁平文件名 → 上游路径(2,198 条已核验映射)
LICENSE 本标注集适用 CC-BY-4.0
NOTICE.md 上游署名与使用说明
README.md 本说明

为什么叫 "Plus"

上游 AncientDoc 基准只有图片 + label.csv,并非 VLM 微调格式。我们补充:(a) 现成的 messages/images SFT 布局与固定 OCR 指令的整页转写标注;(b) 经核验的路径映射,使标注可与任意本地目录布局下的上游图片对齐。

许可

  • 上游图片与基准:CC0-1.0。
  • 本标注集(SFT 数据 + 路径映射):CC-BY-4.0——公有领域古籍文本的我方原创转写;使用请署名并链接本仓库。

引用

请同时引用上游基准与本仓库(BibTeX 见上方英文部分)。