---
title: "Misraj AI at AR-MS NAKBA-NLP 2026: A State-of-the-Art VLM in"
description: "The paper describes the winning system of the AR-MS NAKBA-NLP 2026 shared task, which asks for line-level transcription of… 15 references."
authors: ["Khalil Hennara", "Muhammad Hreden", "Zeina Aldallal", "Sara Chrouf", "Safwan AlModhayan"]
year: 2026
venue: "Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026"
source: "https://aclanthology.org/2026.nakbanlp-1.27/"
doi: "10.63317/2uur73vrcmk5"
licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)"
licence_url: "https://creativecommons.org/licenses/by-nc/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/hennara-2026-misraj-ai-ar-ms-nakba/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/hennara-2026-misraj-ai-ar-ms-nakba/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/) / Misraj AI at AR-MS NAKBA-NLP 2026: A State-of-the-Art VLM in Arabic Handwritten Text Recognition

# Misraj AI at AR-MS NAKBA-NLP 2026: A State-of-the-Art VLM in Arabic Handwritten Text Recognition

**Authors:** [Khalil Hennara](https://md.research.elamur.ai/free/en/authors/khalil-hennara/), [Muhammad Hreden](https://md.research.elamur.ai/free/en/authors/muhammad-hreden/), [Zeina Aldallal](https://md.research.elamur.ai/free/en/authors/zeina-aldallal/), [Sara Chrouf](https://md.research.elamur.ai/free/en/authors/sara-chrouf/) and [Safwan AlModhayan](https://md.research.elamur.ai/free/en/authors/safwan-almodhayan/) · **Year:** 2026 · **Venue:** [Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026](https://md.research.elamur.ai/free/en/venues/nakbanlp/) ([International Workshop on Nakba Narratives as Language Resources @ LREC 2026](https://md.research.elamur.ai/free/en/venues/nakbanlp/2026/)) · **Publisher:** [European Language Resources Association (ELRA)](https://md.research.elamur.ai/free/en/publishers/european-language-resources-association-elra/) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.nakbanlp-1.27/) · [PDF](https://aclanthology.org/2026.nakbanlp-1.27.pdf) · [DOI 10.63317/2uur73vrcmk5](https://doi.org/10.63317/2uur73vrcmk5) · [OpenAlex](https://openalex.org/W7166519090)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper describes the winning system of the AR-MS NAKBA-NLP 2026 shared task, which asks for line-level transcription of handwritten Arabic in the Omar Al-Saleh memoir collection. These are 20th-century archival pages with faded ink and aged paper. The team started from Baseer, a 3B-parameter Arabic document VLM, and trained it with standard next-token prediction. The loss was computed only on the transcription tokens.

Two choices mattered most. First, they used the Muharaf corpus of historical manuscripts as an intermediate training stage, converted to grayscale so it looks like the competition images. Second, they unfroze the vision encoder and gave it a much smaller learning rate than the text decoder. They then averaged the weights of two epochs.

The final model reached CER 0.08 and WER 0.24 on the hidden test set and took first place. A zero-shot frontier model (Gemini 3.1 Pro) did clearly worse. The authors conclude that cursive historical Arabic needs dedicated fine-tuning, and they release the weights and inference code.

## Key points
- **Data:**
  - Nakba training set (NTD): 15,962 lines.
  - Nakba dev set (NDD): 2,095 lines. Images are very wide and short, with a mean aspect ratio near 11.75 (about 614×52 px).
  - Muharaf (MD): 21,196 lines, mostly Ruq'ah script.
- **Decoder-only tuning on the test set (CER / WER):**
  - ND alone: 0.19 / 0.42.
  - ND and MD mixed jointly: 0.18 / 0.42.
  - MD first, then ND: 0.16 / 0.40.
- **Full encoder-decoder tuning (decoder LR 1e-4, vision LR 9e-6):**
  - Both stages with encoder and decoder trained: 0.10 / 0.29.
  - Merging epochs 1 and 5 of that run: 0.08 / 0.24.
- **Leaderboard:**
  - The official baseline scored CER 0.3683 / WER 0.6905.
  - The runner-up WER was 3reeq at 0.2996.
  - The runner-up CER was Oblevit at 0.0925.
- **Zero-shot comparison:** Gemini 3.1 Pro scored CER 0.22 / WER 0.47 on the dev set. A 3-epoch Baseer fine-tune on ND alone scored 0.19 / 0.42.
- **Training setup:** 5 epochs, AdamW with weight decay 0.01, cosine schedule, batch size 128, 644×644 input, 2 H100 GPUs.

## Interesting ideas
- Using auxiliary data as a separate earlier stage worked better than mixing it with the target data, because the model first learns the general skill and then specializes.
- Unfreezing the pretrained encoder but giving it a learning rate roughly ten times smaller than the head lets it adapt without losing what it already knows.
- Averaging the weights of an early checkpoint and a late checkpoint gave a further clear gain at no extra training cost, by smoothing over overfitting to the training set.
- The auxiliary corpus was altered so its surface statistics match the target set, here by converting images to grayscale.
- A cheap, short fine-tune was used as a reality check against a large zero-shot model, and zero-shot exploration was dropped once the small model won.

## For Mova

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

- **I1** (learn): Train the gparse or perceptron parser first on a large out-of-domain treebank, then on the target-domain silver data, and compare this against simply pooling the two.
- **I2** (tree): When fine-tuning the gparse transformer encoder on a new treebank, use a separate, much lower learning rate for the encoder than for the arc and label heads, which should limit forgetting.
- **I3** (learn): Average gparse checkpoints from different epochs and check LAS on held-out data before choosing a single best epoch.
- **I4** (tree): Convert auxiliary treebanks to the target UD dialect and tokenization before mixing them in, so the parser does not learn annotation artifacts.
- **I5** (learn): Before adding a new LLM teacher tier, compare it with a small in-domain trained model on the same held-out set, and keep the teacher only if it actually beats that 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.*

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works · [⚙️ Engineering](https://md.research.elamur.ai/free/en/shelves/engineering/) · 188 works · [⚙️ Media Technology](https://md.research.elamur.ai/free/en/shelves/engineering/media-technology/) · 42 works

🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works

## Cite

Cite as: Khalil Hennara, Muhammad Hreden, Zeina Aldallal, Sara Chrouf and Safwan AlModhayan. 2026. Misraj AI at AR-MS NAKBA-NLP 2026: A State-of-the-Art VLM in Arabic Handwritten Text 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.

```bibtex
@inproceedings{hennara2026misraj,
  title = {{Misraj AI at AR-MS NAKBA-NLP 2026: A State-of-the-Art VLM in Arabic Handwritten Text Recognition}},
  author = {Khalil Hennara, Muhammad Hreden, Zeina Aldallal, Sara Chrouf and Safwan AlModhayan},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/2uur73vrcmk5},
  url = {https://doi.org/10.63317/2uur73vrcmk5}
}
```

- [← previous MIRTH: Mutual-Information Reasoning with Temporal Hubs for Vision-Language-Action Agents](https://md.research.elamur.ai/free/en/papers/sun-2026-mirth-mutual-information-reasoning-temporal/)
- [next → MIThinker: A Plug-and-Play Policy-Optimized Thinker For Motivational Interviewing Counseling](https://md.research.elamur.ai/free/en/papers/yang-2026-mithinker-plug-play-policy-optimized/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/)

## References

15 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. 2025. [Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities](https://md.research.elamur.ai/free/en/papers/comanici-2025-gemini-2-5-pushing-frontier/). arXiv preprint arXiv:2507.06261. [↗](https://arxiv.org/abs/2507.06261)
2. Google. 2026. Gemini (3.1 pro). Large language model. [find](https://scholar.google.com/scholar?q=Google%2E+2026%2E+Gemini+%283%2E1+pro%29%2E+Large+language+model%2E)
3. Hadi Hamoud, Ahmad Ali Chamseddine, Bilal Shalash, Firas Ben Abid, Mustafa Jarrar, Chadi Abou Chakra, Bernard Ghanem, and Fadi A. Zaraket. 2026. 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. [find](https://scholar.google.com/scholar?q=Hadi+Hamoud%2C+Ahmad+Ali+Chamseddine%2C+Bilal+Shalash%2C+Firas+Ben+Abid%2C+Mustafa+Jarrar%2C+Chadi+Abou+Chakra%2C+Bernard+Ghanem%2C+and+Fadi+A%2E+Zaraket%2E+2026%2E+Nakba+nlp+2026%3A+Shared+task+on+arabic+handwritten+manus)
4. 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](https://md.research.elamur.ai/free/en/papers/hennara-2025-baseer-vision-language-model-arabic/). arXiv preprint arXiv:2509.18174. [↗](https://arxiv.org/abs/2509.18174)
5. Abdeljalil Elmajjodi Haitame Bouanane Imane Momayiz, Soufiane Ait Elaouad. 2025. Atlasocr: Open-source ocr for moroccan darija with vision–language models. [https://huggingface.co/atlasia/AtlasOCR](https://huggingface.co/atlasia/AtlasOCR). [↗](https://huggingface.co/atlasia/AtlasOCR)
6. Zhang Li, Yuliang Liu, Qiang Liu, Zhiyin Ma, Ziyang Zhang, Shuo Zhang, Zidun Guo, Jiarui Zhang, Xinyu Wang, and Xiang Bai. 2025. Monkeyocr: Document parsing with a structure-recognition-relation triplet paradigm. arXiv preprint arXiv:2506.05218. [↗](https://arxiv.org/abs/2506.05218)
7. Liana Lorigo and Venu Govindaraju. 2006. Offline arabic handwriting recognition: A survey. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28:712 – 724. [find](https://scholar.google.com/scholar?q=Liana+Lorigo+and+Venu+Govindaraju%2E+2006%2E+Offline+arabic+handwriting+recognition%3A+A+survey%2E+Pattern+Analysis+and+Machine+Intelligence%2C+IEEE+Transactions+on%2C+28%3A712+%E2%80%93+724%2E)
8. Mehreen Saeed, Adrian Chan, Anupam Mijar, Joseph Moukarzel, Georges Habchi, Carlos Younes, Amin Elias, Chau-Wai Wong, and Akram Khater. 2025. Muharaf: Manuscripts of handwritten arabic dataset for cursive text recognition. [find](https://scholar.google.com/scholar?q=Mehreen+Saeed%2C+Adrian+Chan%2C+Anupam+Mijar%2C+Joseph+Moukarzel%2C+Georges+Habchi%2C+Carlos+Younes%2C+Amin+Elias%2C+Chau%2DWai+Wong%2C+and+Akram+Khater%2E+2025%2E+Muharaf%3A+Manuscripts+of+handwritten+arabic+dataset+for+cur)
9. Mahmoud Salaheldin Kasem, Mohamed Mahmoud, and Hyun-Soo Kang. 2025. Advancements and challenges in arabic optical character recognition: A comprehensive survey. ACM Computing Surveys, 58(4):1–37. [find](https://scholar.google.com/scholar?q=Mahmoud+Salaheldin+Kasem%2C+Mohamed+Mahmoud%2C+and+Hyun%2DSoo+Kang%2E+2025%2E+Advancements+and+challenges+in+arabic+optical+character+recognition%3A+A+comprehensive+survey%2E+ACM+Computing+Surveys%2C+58%284%29%3A1%E2%80%9337%2E)
10. Hunyuan Vision Team, Pengyuan Lyu, Xingyu Wan, Gengluo Li, Shangpin Peng, Weinong Wang, Liang Wu, Huawen Shen, Yu Zhou, Canhui Tang, et al. 2025. [Hunyuanocr technical report](https://md.research.elamur.ai/free/en/papers/team-2025-vision-hunyuanocr-technical-report/). arXiv preprint arXiv:2511.19575. [↗](https://arxiv.org/abs/2511.19575)
11. 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. [find](https://scholar.google.com/scholar?q=Ahmed+Wasfy%2C+Omer+Nacar%2C+Abdelakreem+Elkhateb%2C+Mahmoud+Reda%2C+Omar+Elshehy%2C+Adel+Ammar%2C+and+Wadii+Boulila%2E+2025%2E+Qari%2Docr%3A+High%2Dfidelity+arabic+text+recognition+through+multimodal+large+language+model+)
12. Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, et al. 2024. General ocr theory: Towards ocr-2.0 via a unified end-to-end model. arXiv preprint arXiv:2409.01704. [↗](https://arxiv.org/abs/2409.01704)
13. Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. 2025. [Qwen-image technical report](https://md.research.elamur.ai/free/en/papers/wu-2025-qwen-image-technical-report/). arXiv preprint arXiv:2508.02324. [↗](https://arxiv.org/abs/2508.02324)
14. Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. 2026. [Model merging in llms, mllms, and beyond: Methods, theories, applications, and opportunities](https://md.research.elamur.ai/free/en/papers/yang-2024-model-merging-llms-mllms-methods/). ACM Computing Surveys, 58(8):1– 41.
15. Fadi Zaraket, Bilal Shalash, Hadi Hamoud, Ahmad Chamseddine, Firas Ben Abid, Mustafa Jarrar, Chadi Abou Chakra, and Bernard Ghanem. 2026. Ar-ms: Arabic manuscript understanding. 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. [find](https://scholar.google.com/scholar?q=Fadi+Zaraket%2C+Bilal+Shalash%2C+Hadi+Hamoud%2C+Ahmad+Chamseddine%2C+Firas+Ben+Abid%2C+Mustafa+Jarrar%2C+Chadi+Abou+Chakra%2C+and+Bernard+Ghanem%2E+2026%2E+Ar%2Dms%3A+Arabic+manuscript+understanding%2E+In+Proceedings+of+the+)

*Updated 2 November 2026*