Elamur Research⌕
☰
home / papers / 2026 / EvaHan 2026 Ancient Books Multimodal OCR and Layout Analysis System Technical Report

EvaHan 2026 Ancient Books Multimodal OCR and Layout Analysis System Technical Report

Paper · 2026 · Chenrui Zheng

Summary

This system report describes one team's entry to EvaHan 2026, a shared task on OCR for ancient Chinese books with multimodal LLMs. The task has three parts: printed text recognition (Task A, CER), layout element detection of text, image, book_edge and seal (Task B, mAP@[.5:.95]), and handwritten text recognition (Task C, CER). Under the Closed Modality rule, the author fine-tuned Qwen2.5-VL-7B-Instruct.

The approach combines three things:

  • Cleaner data: punctuation was removed, repetition marks were swapped for a placeholder token and restored afterwards, and Task B annotations were corrected by hand.
  • Synthetic data: hard-character samples and recombined layouts were generated.
  • Staged training: a curriculum ordered by difficulty, mainly with LoRA, and with full-parameter and GRPO variants tested for Task B.

On a local validation split, staged training clearly beat plain SFT for both character recognition and layout detection. GRPO destabilised training, and the hardest layouts caused repetition loops. For historical-text digitisation, the main lesson is that data cleaning and curriculum design matter at least as much as the training algorithm. Code and synthetic data are released on GitHub.

Key points

  • Data split: each official dataset has about 5,000 image–text pairs, split 9:1 into 4,500 training and 500 validation samples. For Task B, 480 images with unreliable or overly complex annotations were dropped.
  • Synthetic data:
    • 1,000 hard-character samples for Tasks A and C, built from characters that Qwen2.5-VL-72B misread.
    • 2,500 recombined layout images for Task B.
  • Task B curriculum: K-Means clustering on element count, density and overlap produced three difficulty levels: Easy (2,400), Medium (949) and Hard (723).
  • Validation CER:
    • Task A fell from 10.0% to 2.0% (Stage 1) and was 3.0% after Stage 2.
    • Task C fell from 22.0% to 10.0% after Stage 2.
  • Task B results:
    • The paper reports a gain of 10.9 percentage points in mAP over its baseline.
    • The submitted model, Stage1-LoRA + Stage2-Full + Stage3-Full, reached roughly 0.51 mAP, compared with 0.47 for direct SFT. These figure values come from a garbled plot.
    • GRPO did not help and made training unstable.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Sun Yat-sen UniversityGuangzhou · China72 works · 9,502 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 worksEngineering · 188 worksMedia Technology · 42 works

🧭 Topics 1Research subjects and languages the work is about.Vision & language · 811 works

Cite

Cite as Chenrui Zheng. 2026. EvaHan 2026 Ancient Books Multimodal OCR and Layout Analysis System Technical Report. Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3n7j6t59tuhs

@inproceedings{zheng2026evahan,
  title = {{EvaHan 2026 Ancient Books Multimodal OCR and Layout Analysis System Technical Report}},
  author = {Chenrui Zheng},
  booktitle = {Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/3n7j6t59tuhs},
  url = {https://doi.org/10.63317/3n7j6t59tuhs}
}

References

11 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. B. Wang, C. Xu, X. Zhao, et al., “MinerU: An Open-Source Solution for Precise Document Content Extraction,” arXiv:2409.18839, 2024.
  2. Baidu. PP-Structure: Document Structure Analysis. https://github.com/PaddlePaddle/ PaddleOCR, 2024. ↗
  3. Y. Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” in Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pp. 41-48, 2009.
  4. F. Borisyuk, A. Gordo, and V. Sivakumar, “Rosetta: Large scale system for text detection and recognition in images,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 71-79, 2018. find
  5. C. Cui, T. Sun, M. Lin, et al., “PaddleOCR 3.0 Technical Report,” arXiv preprint arXiv:2507.05595v1, 2025. — [arXiv:2507.05595](https://arxiv.org/abs/2507.05595) find
  6. H. Li, Y. Liu, W. Liao, et al., “Optical Character Recognition in the Era of Large Models: Status and Prospect,” Journal of Image and Graphics, vol. 30, no. 6, pp. 2023-2050, 2025. find
  7. J. Niu, Z. Liu, Z. Gu, et al., “MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing,” arXiv:2509.22186, 2025.
  8. J. Poznanski, A. Rangapur, J. Borchardt, et al., “olmOCR: Unlocking Trillions of Tokens in PDFs with Vision Language Models,” arXiv:2502.18443, 2025.
  9. R. Saini et al., “ICDAR 2019 Historical Document Reading Challenge on Large Structured Chinese Family Records,” in IEEE International Conference on Document Analysis and Recognition (ICDAR), 2019. find
  10. P. Soviany, R. T. Ionescu, P. Rota, and N. Sebe, “Curriculum Learning: A Survey,” International Journal of Computer Vision, vol. 130, pp. 877– 902, 2022. find
  11. X. Wang, Y. Chen, and W. Zhu, “A Survey on Curriculum Learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 4555-4576, 2021. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zheng-2026-evahan-2026-ancient-books-multimodal/