Eski Uygur OCR: VLM’lerin İnce Ayarının Yeniden Üretilmesi Yoluyla Devam Eden İlk Çalışma

Year 2025, Volume: 9 Issue: 2, 313 - 318, 18.08.2025

Mehmet Oguz Derin , Erdem Uçar

https://doi.org/10.35236/jots.1762980

Abstract

Old Uyghur manuscripts remain largely inaccessible due to the absence of optical character recognition (OCR) systems aligned with modern scholarly practices. To address this, we present a work‑in‑progress OCR approach using fine-tuning of vision–language models (VLMs) for Old Uyghur transliteration. Specifically, LLaMA‑3.2‑11B‑Vision was fine‑tuned with 4‑bit Unsloth quantization and LoRA on 525 manually annotated pages from Radloff and Malov’s edition of Altun Yaruk Sudur, following Unicode‑based transliteration guidelines. On a held‑out test set of 16 pages from one bölük (part) of the fifth tägzinč, the approach yielded a character error rate (CER) of 5.46% and normalized edit distance (NED) of 0.286. These results demonstrate feasibility; however, limitations are evident owing to binarized scans, a single woodblock style, small training data, and lack of positional supervision. Consequently, findings provide a baseline and a reproducible path toward robust Old Uyghur OCR.

Keywords

OCR , Vision–Language Models , Old Uyghur , Unicode‑based transliteration , Low‑resource , Reproducibility

Old Uyghur OCR: The First Work-in-Progress via Reproducing Fine-tuning of VLMs

Abstract

Old Uyghur manuscripts remain largely inaccessible due to the absence of optical character recognition (OCR) systems aligned with modern scholarly practices. To address this, we present a work‑in‑progress OCR approach using fine-tuning of vision–language models (VLMs) for Old Uyghur transliteration. Specifically, LLaMA‑3.2‑11B‑Vision was fine‑tuned with 4‑bit Unsloth quantization and LoRA on 525 manually annotated pages from Radloff and Malov’s edition of Altun Yaruk Sudur, following Unicode‑based transliteration guidelines. On a held‑out test set of 16 pages from one bölük (part) of the fifth tägzinč, the approach yielded a character error rate (CER) of 5.46% and normalized edit distance (NED) of 0.286. These results demonstrate feasibility; however, limitations are evident owing to binarized scans, a single woodblock style, small training data, and lack of positional supervision. Consequently, findings provide a baseline and a reproducible path toward robust Old Uyghur OCR.

Keywords

OCR , Vision–Language Models , Old Uyghur , Unicode‑based transliteration , Low‑resource , Reproducibility

References

• Chung, Y. H. M., & Choi, D. (2025). Fine‑tuning vision‑language models as OCR systems for low‑resource languages: A case study of Manchu. arXiv: 2507.06761.
• Derin, M. O. & Harada, T. (2021). Universal Dependencies for Old Turkish. Proceedings of the Fifth Workshop on Universal Dependencies (pp. 129–141). Association for Computational Linguistics.
• Han, D. et al. (2023). Unsloth: Finetune LLMs 2× faster with 70% less memory. (https://github.com/unslothai/unsloth)
• Radlov, V. V. & Malov, S. E. (1913-1917). Suvarṇaprabhāsa (Sutra zolotogo bleska), Tekst’ Uygurskoy Redaktsiy, izdali I-IV. St. Petersburg: Imperatorskoy Akad. Nauk.
• Uçar, E. (2023). Altun Yaruq Sudur’un St. Petersburg Yazması’ndaki Yapraklar ile Radlov-Malov Neşrindeki Sayfaların Tekabül Listesi (Deneme). Babür, 2(2), 35-63.
• Uçar, E. & Derin, M. O. (2025). Die digitale Zukunft der sogdischen und altuigurischen Schrift: Entwicklung und Etablierung von Standards für Transliteration, Transkription und Unicode‑Kodierung. Journal of Old Turkic Studies, 9(2), 535-552.
• Zieme, P. (2015). Local Literatures: Uighur. In Silk J. A. et al. (Eds.), Brill’s Encyclopedia of Buddhism. Vol. I: Literature and Languages (pp. 871-882). Leiden-Boston: Brill.