AncientDoc: A Benchmark for Chinese Ancient Document Understanding

AncientDoc is the first comprehensive benchmark dataset specifically designed for Chinese Ancient Document Understanding. It covers multi-task evaluation ranging from OCR to knowledge reasoning, aiming to promote research on the recognition, understanding, and reasoning capabilities of multimodal large models in the scenario of ancient documents.

Dataset Overview

• Data Scale: 2,973 pages
• Number of Literatures: Approximately 100 books
• Types of Literatures: 14 categories (e.g., collected works, Chu Ci-style poems, poetry and prose criticism, encyclopedic books, catalogs, etc.)
• Time Span: Spanning from the Warring States period to the Qing Dynasty, covering multiple important historical periods
• Task Types:
  1. Page-level OCR: Full-page text recognition (including complex scenarios such as vertical typesetting, variant characters, and annotations)
  2. Vernacular Translation: Intralingual translation from classical Chinese to modern Chinese
  3. Reasoning-based QA: Implicit reasoning QA based on the meaning of the text
  4. Knowledge-based QA: QA based on textual facts and background knowledge
  5. Linguistic Variant QA: QA related to literary genres, rhetoric, and linguistic styles

Data Distribution

Distribution by Dynasty

• Ming Dynasty: 1,148 pages
• Qing Dynasty: 778 pages
• Song Dynasty: 540 pages
• Tang Dynasty: 208 pages
• Han Dynasty: 110 pages
• Yuan Dynasty: 69 pages
• Southern and Northern Dynasties: 54 pages
• Jin Dynasty: 42 pages
• Warring States period: 24 pages

Distribution by Category (Top 3 by Page Count)

1. Astronomy and Mathematics (238 pages)
2. Confucianism (232 pages)
3. Art (234 pages)

Distribution by Script Style

• Regular script: Approximately 97%
• Cursive script: Approximately 3%

Data Format

Data is provided in the form of images + CSV annotations: