Elamur Research⌕
☰
home / papers / 2026 / Ketaba-OCR at AR-MS NakbaNLP 2026: Efficient Adaptation of Vision-Language Models for Handwritten Recognition

Ketaba-OCR at AR-MS NakbaNLP 2026: Efficient Adaptation of Vision-Language Models for Handwritten Recognition

Paper · 2026 · Hassan Barmandah, Fatimah Emad Eldin, Khloud Al Jallad et al.

Summary

The paper presents Ketaba-OCR-LoRA, a system for Subtask 2 of the NakbaNLP 2026 shared task. The task is transcribing line images from the Omar Al-Saleh Memoir Collection (1951–1965), Palestinian testimonies handwritten in historical Ruq'ah and Naskh scripts. The output must be Unicode text that keeps the original orthography and diacritics. The authors did not adapt a general vision-language model. Instead they started from a public handwriting-recognition model, a 4B Qwen2.5-VL variant already trained on Muharaf and IAM. They tuned it on the shared-task data with QLoRA, which cut memory needs from about 40GB to about 8GB. Six fine-tuned variants were then merged with a new weighting scheme called Linear+Boost. On the blind test the system came first when scores were computed line by line (CER 0.0819, WER 0.2588), and it ranked third on the official leaderboard, which pools scores across the whole corpus (CER 0.0938, WER 0.2996). The study shows that a handwriting-specialised backbone with cheap adapters beats generalist models. It also gives a reproducible, low-cost recipe for digitising heritage documents.

Key points

  • Data: roughly 15,962 training, 1,774 development and 2,095 test line images. The official ranking used a hidden CodaBench set of 2,671 images. Fine-tuning used only the shared-task data.
  • Backbone: sherif1313/Arabic-English-handwritten-OCR-v3, based on Qwen2.5-VL. It was adapted with 4-bit NF4 QLoRA (r = 32, α = 64) plus DoRA and RSLoRA. Adapters sit on the attention projections (q, k, v, o) and the MLP projections (gate, up, down).
  • Domain adaptation lowered CER from about 0.58 (untuned) to about 0.08, a relative gain of roughly 86%.
  • Linear+Boost weighting gives each model a weight of (1 − CER), plus a 0.5 bonus when CER is below 0.15. It was chosen from 30 tested strategies. It moved blind-test CER from 0.0884 (inverse-CER weighting) to 0.0819, a relative gain of 7.4%.
  • Error profiles of the baselines differed:
    • base VLMs produced over 40,000 substitutions;
    • the MSA-oriented Qari models over-segmented badly, with more than 200,000 insertions.

Interesting ideas

  • I1. Begin from a model already specialised for a nearby domain and add small adapters, rather than adapting a broad generalist; this gave large gains at a fraction of the memory.
  • I2. When combining systems, use bounded linear weights from their error rate plus a fixed bonus for those past a quality threshold; this favours strong members without letting one dominate as inverse-error weighting can.
  • I3. An ensemble can act as a guard against typical generative failures such as repetition loops and context drift, using voting and edit-distance consensus across candidate outputs.
  • I4. Scoring the same output both per line and over the pooled corpus can change the ranking, so reporting both shows how robust a system really is.
  • I5. Splitting errors into substitutions and insertions showed that different models fail in different ways, for example over-segmentation versus misreading.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tree): gparse could be adapted to a new UD dialect or genre by training low-rank adapters on top of a parser already trained on a related treebank, avoiding full retraining.
  • I2 (learn): witness parsers voting on silver trees could be weighted this way, so strong parsers lead while weaker ones still help break ties.
  • I3 (learn): when LLM teacher tiers produce labels or analyses, edit-distance consensus over several samples could filter out looping or drifting answers before they enter training data.
  • I4 (learn): evaluation could report both per-sentence averaged LAS and pooled LAS, exposing whether errors cluster in a few long or odd sentences.
  • I5 (tokens): breaking tokenizer and tagger errors down by type (extra splits vs. wrong labels) would show whether fixes should target segmentation rules or the tagging model.

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

🎓Umm al-Qura UniversityMakkah al Mukarramah · Saudi Arabia21 works · 274 citations🎓Tuwaiq AcademyRiyadh · Saudi Arabia11 works · 7 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksEngineering · 133 worksMedia Technology · 34 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksVision & language · 657 works

Cite

Cite as Hassan Barmandah, Fatimah Emad Eldin, Khloud Al Jallad and Omer Nacar. 2026. Ketaba-OCR at AR-MS NakbaNLP 2026: Efficient Adaptation of Vision-Language Models for Handwritten Recognition. Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3inc2znes52o

@inproceedings{barmandah2026ketaba,
  title = {{Ketaba-OCR at AR-MS NakbaNLP 2026: Efficient Adaptation of Vision-Language Models for Handwritten Recognition}},
  author = {Hassan Barmandah, Fatimah Emad Eldin, Khloud {Al Jallad} and Omer Nacar},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/3inc2znes52o},
  url = {https://doi.org/10.63317/3inc2znes52o}
}

References

18 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mehreen Aladmani, Erin Woertz, David Juen, Aashish Agarwal, Stephen Houdek, Gerik Jager, Razan Abuaita, and Nizar Habash. 2024. Muharaf: Manuscripts of handwritten Arabic dataset for cursive text recognition. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 6651–6661, Torino, Italy. find
  2. Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966. ↗
  3. Tim Dettmers, Mike Lewis, Sam Shleifer, and Luke Zettlemoyer. 2022. 8-bit optimizers via block-wise quantization. In International Conference on Learning Representations. ↗
  4. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient fine-tuning of quantized llms. In Advances in Neural Information Processing Systems, volume 36. find
  5. Hadi Hamoud, Ahmad Ali Chamseddine, Bilal Shalash, Firas Ben Abid, Mustafa Jarrar, Chadi Abou Chakra, Bernard Ghanem, and Fadi A. Zaraket. 2026. find
  6. NAKBA NLP 2026: Shared Task on Arabic Handwritten Manuscript Understanding (Palestine Memory–Omar Al-Saleh Memoir). In Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources (Nakba-NLP 2026), co-located with the Language Resources and Evaluation Conference (LREC 2026), Palma, Mallorca, Spain.
  7. Khalil Hennara, Muhammad Hreden, Mohamed Motasim Hamed, Ahmad Bastati, Zeina Aldallal, Sara Chrouf, and Safwan AlModhayan. 2025. Baseer: A vision-language model for arabic document-to-markdown OCR. arXiv preprint arXiv:2509.18174. ↗
  8. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  9. Damjan Kalajdzievski. 2023. A rank stabilization scaling factor for fine-tuning with lora. arXiv preprint arXiv:2312.03732. ↗
  10. Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. 2023. Trocr: Transformer-based optical character recognition with pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 13094–13102. find
  11. Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. DoRA: Weight-decomposed low-rank adaptation. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 32100– 32121. PMLR. ↗
  12. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations. find
  13. Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, Benjamin Bossan, and Marian Tietz. 2022. PEFT: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft. ↗
  14. Urs-Viktor Marti and Horst Bunke. 2002. The IAM-database: An English sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1):39–46. ↗
  15. Ahmed Sherif. 2025. Arabic-english handwritten ocr v3. https://huggingface.co/sherif1313/Arabic-English-handwritten-OCR-v3. Vision-Language Transformer model fine-tuned for Arabic and English handwritten text recognition. Apache-2.0 License. Accessed: 2026-01-10. ↗
  16. Ahmed Wasfy, Omer Nacar, Abdelakreem Elkhateb, Mahmoud Reda, Omar Elshehy, Adel Ammar, and Wadii Boulila. 2025. Qari-ocr: High-fidelity arabic text recognition through multimodal large language model adaptation. arXiv preprint arXiv:2506.02295. ↗
  17. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45. Association for Computational Linguistics.
  18. Zai-org. 2026. Glm-ocr: A multimodal large language model for ocr and document understanding. https://github.com/zai-org/GLM-OCR. Accessed: 2026-02-19. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/barmandah-2026-ketaba-ocr-ar-ms-nakbanlp/