Ancient-Bench: A Comprehensive Multi-millennial, Multi-medium, and Multi-script Benchmark for Ancient Chinese Artifact Text Recognition

English | 中文简体

---

🚨 Crucial Notice

Strictly for Academic Research Use Only.
The images in this dataset are curated from publicly accessible archives of official cultural heritage institutions (e.g., museums and libraries). The authors claim no copyright over the original images. Users must strictly comply with the corresponding institutional terms and licenses. Any commercial use is strictly prohibited.

---

📥 Dataset Access

Platform	Link	Extraction Code / Notes
Baidu Netdisk (百度网盘)	🔗 Click to Download from Baidu Netdisk	54su
Hugging Face	🤗 Dataset Repository	Direct Access

You can also download directly via Python using huggingface_hub:

from huggingface_hub import hf_hub_download

zip_path = hf_hub_download(
    repo_id="hiuyi/Ancient-Bench",
    filename="Ancient-Bench.zip",
    repo_type="dataset",
    local_dir="./",
)
print(f"Dataset downloaded to: {zip_path}")

---

📌 Overview

Ancient-Bench (Accepted by EMNLP 2026) is the first comprehensive, multi-millennial, multi-medium, and multi-script benchmark tailored for ancient Chinese artifact text recognition. It comprises 2,700 meticulously annotated high-resolution images, spanning 3,000+ years of Chinese script evolution, covering 9 heterogeneous artifact categories from 14 national-level cultural institutions, and encompassing 7 historical script forms.

To establish an objective and unified evaluation across heterogeneous media, Ancient-Bench defines three standardization protocols: Symbol Standardization, Character Standardization, and Parsing Standardization.

✨ Key Features

1. Multi-millennial (3,000+ Years): Spans the entire evolutionary trajectory of Chinese writing, divided into three major historical eras:
   • Early Civilization Period (1200 – 221 BCE): Oracle bones, ritual bronzes, early bamboo slips, and silk manuscripts.
   • Imperial Medieval & Stone Inscription Age (221 BCE – 900 CE): Official seals, stone steles, epitaphs, and cliff inscriptions.
   • Early Modern Printing & Artistic Maturity Period (900 – 1911 CE): Woodblock editions and paper calligraphy works.
2. Multi-medium (9 Artifact Categories): Oracle Bones, Bronzes, Bamboo/Wooden Slips, Silk Manuscripts, Seals, Steles, Cliff Inscriptions, Ancient Editions, and Calligraphy.
3. Multi-script (7 Mainstream Scripts): Oracle Bone Script, Bronze Script, Seal Script, Clerical Script, Regular Script, Cursive Script, and Running Script.
4. Authoritative & Diverse Provenance: Sourced from 14 national-level cultural heritage institutions and authentic in-the-wild cliff inscriptions.

📊 Dataset Statistics

Medium	Images	Resolution Range (px)	Chars / Image	Institutions	Script Types Covered
Oracle (甲骨)	300	600×800	1–68	1	Oracle Bone
Bronze (金文)	250	41×80 – 1024×1024	1–77	2	Bronze
Slip (简牍)	300	29×224 – 700×6820	1–107	3	Seal, Clerical
Silk (帛书)	300	46×124 – 1330×2067	1–784	2	Seal, Clerical
Seal (印章)	350	100×150 – 1333×1341	1–40	4	Oracle, Bronze, Seal
Steles (碑帖)	300	85×155 – 5795×16745	4–816	4	Seal, Clerical, Regular, Running
Cliff (摩崖)	200	183×113 – 7300×5760	2–216	In-the-wild	Seal, Clerical, Regular, Running
Edition (刻本)	300	751×506 – 970×2817	4–776	1	Seal, Clerical, Regular, Cursive, Running
Calligraphy (书法)	400	303×509 – 4802×13385	2–1666	4	Bronze, Seal, Clerical, Regular, Cursive, Running
Total	2,700	41×80 – 5795×16745	1–1666	14 + In-the-wild	All 7 Scripts

📐 Standardization Protocols

• Symbol Standardization: Preserves historical typographic marks as-is (= and - for repetition/ligatures; ~ for transpositions; □ for unreadable damaged characters; <unrecognizable> for unencoded characters).
• Character Standardization: Uses scholarly clerical transcription (隶定) with traditional characters as reference; strictly complies with the “what you see is what you get” (WYSIWYG) principle to avoid semantic distortion from modern simplification.
• Parsing Standardization: Preserves reading hierarchy and column-by-column orders; explicitly tags double-line interlinear notes using （）, non-transcription areas (center seam, book ears) using <ignore>, and margin notes using <note>.

---

⚖️ Ethical Statement & Copyright

1. Academic Use Only: Ancient-Bench is constructed strictly for academic research and algorithm benchmarking. Commercial use in any form is prohibited.
2. Image Copyright: Images in Ancient-Bench are collected from publicly accessible online repositories of museums, libraries, and cultural heritage institutions. The authors do not claim or transfer any copyright over these artifacts or source images.
3. Compliance: Researchers downloading this dataset must strictly comply with the terms of service and licensing of the respective cultural institutions.

---

📖 Citation

If you find Ancient-Bench helpful in your research, please cite our EMNLP 2026 paper:

@inproceedings{cheng2026ancientbench,
  title={Ancient-Bench: A Comprehensive Multi-millennial, Multi-medium, and Multi-script Benchmark for Ancient Chinese Artifact Text Recognition},
  author={Cheng, Hiuyi and Xu, Nuo and Zhang, Yuyi and Zheng, Xuhan and Pan, Wei and Zhang, Jing and Peng, Dezhi and Liao, Minghui and Teng, Yihua and Wu, Jihao and Ren, Haoyu and Jin, Lianwen},
  booktitle={Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2026},
  url={https://arxiv.org/abs/2608.27169}
}