---
title: "Sadeed: Advancing Arabic Diacritization Through Small Language Model — Aldallal 2025"
description: "The authors build Sadeed by fine-tuning Kuwain 1.5B, a small decoder-only model pre-trained on Arabic… 7.19 points. 32 references, cited by 3 papers here."
authors: ["Zeina Aldallal", "Sara Chrouf", "Khalil Hennara", "Mohamed Motaism Hamed", "Muhammad Hreden", "Safwan AlModhayan"]
year: 2025
venue: "arXiv:2504.21635"
source: "https://arxiv.org/abs/2504.21635"
doi: "10.48550/arxiv.2504.21635"
arxiv: "2504.21635"
licence: "CC BY-NC-SA 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by-nc-sa/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/aldallal-2025-sadeed-advancing-arabic-diacritization-through/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/aldallal-2025-sadeed-advancing-arabic-diacritization-through/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/2/) / Sadeed: Advancing Arabic Diacritization Through Small Language Model

# Sadeed: Advancing Arabic Diacritization Through Small Language Model

**Authors:** [Zeina Aldallal](https://md.research.elamur.ai/free/en/authors/zeina-aldallal/), [Sara Chrouf](https://md.research.elamur.ai/free/en/authors/sara-chrouf/), [Khalil Hennara](https://md.research.elamur.ai/free/en/authors/khalil-hennara/), [Mohamed Motaism Hamed](https://md.research.elamur.ai/free/en/authors/mohamed-motaism-hamed/), [Muhammad Hreden](https://md.research.elamur.ai/free/en/authors/muhammad-hreden/) and [Safwan AlModhayan](https://md.research.elamur.ai/free/en/authors/safwan-almodhayan/) · **Year:** 2025 · **Venue:** arXiv:2504.21635 · **Licence:** [CC BY-NC-SA 4.0 (arXiv)](http://creativecommons.org/licenses/by-nc-sa/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2504.21635) · [PDF](https://arxiv.org/pdf/2504.21635) · [DOI 10.48550/arxiv.2504.21635](https://doi.org/10.48550/arxiv.2504.21635) · [arXiv 2504.21635](https://arxiv.org/abs/2504.21635) · [OpenAlex](https://openalex.org/W4417525073)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · 3 in our library, of the 140,884 papers whose bibliographies we read

## Summary
Arabic is usually written without short-vowel marks (diacritics). Restoring them is hard because one consonant skeleton can stand for several words, and the right choice often depends on the whole sentence and even on punctuation. The authors build Sadeed by fine-tuning Kuwain 1.5B, a small decoder-only model pre-trained on Arabic, on diacritized text from the Tashkeela Corpus. Before training, that text went through strict cleaning and normalization. They also audit existing benchmarks (Abbad, CATT) and find mistakes in splitting, labeling and content, plus heavy overlap between the Fadel and Abbad sets, which both come from Tashkeela. In response they release SadeedDiac-25, an expert-reviewed benchmark that mixes Classical Arabic and Modern Standard Arabic. On this benchmark Sadeed trails only the big proprietary models and clearly beats the open Arabic LLMs. Its main weakness is hallucinated words, and it is weaker on MSA because little MSA training data exists. The work shows that a small task-specific model trained on clean data can be competitive, and that dataset leakage can make earlier results look better than they are.

## Key points
- Base model: Kuwain 1.5B (decoder-only SLM). Training data: a cleaned Tashkeela-derived set of about 1 million examples and about 53 million words, compatible with the Fadel test set.
- SadeedDiac-25: a new open benchmark balancing CA and MSA, reviewed by experts, designed to avoid the split and label errors found in Abbad and CATT.
- Ranking on SadeedDiac-25 by DER/WER, with and without case endings:
  - Claude-3-7-Sonnet, GPT-4 and gemini-flash-2.0 lead.
  - Sadeed comes next.
  - Aya-23-8B is the best of the remaining open models.
  - ALLaM, Yehia, jais, gemma and SILMA trail far behind.
- Of Sadeed's WER of about 9.92, roughly 7.19 points come from hallucinated words, which the authors attribute to the small model size.
- Fadel and Abbad overlap strongly, so earlier models trained on one and tested on the other may report inflated scores.

## Interesting ideas
- Report hallucinated output (words the model invented rather than mis-marked) as its own error column, apart from ordinary errors; this shows whether failures come from wrong analysis or from the output drifting away from the input.
- Before trusting a score, check for overlap between training and test sets that come from the same parent corpus, since shared origin quietly inflates results.
- Score each item both with and without its sentence-dependent part (here the case ending), so the easy lexical part and the hard syntax-driven part are measured separately.
- Punctuation alone can flip the grammatical reading of a sentence, so removing or changing it makes a cheap robustness test.
- A small model fine-tuned on one narrow, clean task can beat much larger general models that only roughly understand the task.

## For Mova

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

- **I1** (back): The text → graph → text compressor and generation could track how many output words have no source in the meaning graph, separately from wording errors.
- **I2** (learn): Silver data from witness parsers and LLM teachers should be deduplicated against evaluation treebanks and test fables, and that check should be part of evaluation.
- **I3** (tokens): The tagger and morphology could report accuracy split into context-free features and features that need syntax, which shows where the tree must help.
- **I4** (tree): Run the perceptron parser and gparse on copies of sentences with punctuation removed or changed, and log which attachments flip, as a stress test.
- **I5** (learn): Instead of relying on a big LLM teacher, Mova could distill a small specialist for one narrow step, such as a single repair or labeling task, and keep the teacher only for hard cases.

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

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

Topics: [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · [🐪 Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/)

## Institutions

- [Saad Specialist Hospital](https://md.research.elamur.ai/free/en/hospitals/saad-specialist-hospital/) — Hospital, Khobar, Saudi Arabia · 3 works · 7 citations

## Cite

Cite as: Zeina Aldallal, Sara Chrouf, Khalil Hennara, Mohamed Motaism Hamed, Muhammad Hreden and Safwan AlModhayan. 2025. Sadeed: Advancing Arabic Diacritization Through Small Language Model. arXiv:2504.21635.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{aldallal2025sadeed,
  title = {{Sadeed: Advancing Arabic Diacritization Through Small Language Model}},
  author = {Zeina Aldallal, Sara Chrouf, Khalil Hennara, Mohamed Motaism Hamed, Muhammad Hreden and Safwan AlModhayan},
  year = {2025},
  doi = {10.48550/arxiv.2504.21635},
  eprint = {2504.21635},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2504.21635}
}
```

- [← previous Phonikud: Overcoming Phonetic Underspecification for Hebrew Text-To-Speech](https://md.research.elamur.ai/free/en/papers/kolani-2025-phonikud-overcoming-phonetic-underspecification-hebrew/)
- [next → Speech Act Patterns for Improving Generalizability of Explainable Politeness Detection Models](https://md.research.elamur.ai/free/en/papers/aljanaideh-2025-speech-act-patterns-improving-generalizability/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/2/)

## References

32 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Hamza Abbad and Shengwu Xiong. Multi-components system for automatic arabic diacritization. In Joemon M. Jose, Emine Yilmaz, João Magalhães, Pablo Castells, Nicola Ferro, Mário J. Silva, and Flávio Martins (eds.), Advances in Information Retrieval, pp. 341–355, Cham, 2020. Springer International Publishing. ISBN 978-3-030-45439-5. [find](https://scholar.google.com/scholar?q=Hamza+Abbad+and+Shengwu+Xiong%2E+Multi%2Dcomponents+system+for+automatic+arabic+diacritization%2E+In+Joemon+M%2E+Jose%2C+Emine+Yilmaz%2C+Jo%C3%A3o+Magalh%C3%A3es%2C+Pablo+Castells%2C+Nicola+Ferro%2C+M%C3%A1rio+J%2E+Silva%2C+and+Fl%C3%A1vio+Ma)
2. Ahmed Abdelali, Kareem Darwish, Nadir Durrani, and Hamdy Mubarak. Farasa: A fast and furious segmenter for Arabic. In John DeNero, Mark Finlayson, and Sravana Reddy (eds.), Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pp. 11–16, San Diego, California, June 2016. Association for Computational Linguistics. doi: 10.18653/v1/N16-3003. URL https://aclanthology.org/N16-3003. [↗](https://doi.org/10.18653/v1/n16-3003) [our card](https://md.research.elamur.ai/free/en/papers/abdelali-2016-farasa-fast-furious-segmenter-arabic/)
3. Abdulmohsen Al-Thubaity, Atheer Alkhalifa, Abdulrahman Almuhareb, and Waleed Alsanie. Arabic diacritization using bidirectional long short-term memory neural networks with conditional random fields. IEEE Access, 8:154984–154996, 2020. [find](https://scholar.google.com/scholar?q=Abdulmohsen+Al%2DThubaity%2C+Atheer+Alkhalifa%2C+Abdulrahman+Almuhareb%2C+and+Waleed+Alsanie%2E+Arabic+diacritization+using+bidirectional+long+short%2Dterm+memory+neural+networks+with+conditional+random+fields%2E+I)
4. Faris Alasmary, Orjuwan Zaafarani, and Ahmad Ghannam. Catt: Character-based arabic tashkeel transformer, 2024. URL https://arxiv.org/abs/2407.03236. [↗](https://arxiv.org/abs/2407.03236)
5. Badr AlKhamissi, Muhammad N ElNokrashy, and Mohamed Gabr. Deep diacritization: Efficient hierarchical recurrence for improved arabic diacritization. arXiv preprint arXiv:2011.00538, 2020. [↗](https://arxiv.org/abs/2011.00538)
6. Manar M. Almanea. Automatic methods and neural networks in arabic texts diacritization: A comprehensive survey. IEEE Access, 9:145012–145032, 2021. doi: 10.1109/ACCESS.2021.312297 7. [↗](https://doi.org/10.1109/access.2021.312297)
7. Rehab Alnefaie and Aqil M. Azmi. Automatic minimal diacritization of arabic texts. Procedia Computer Science, 117:169–174, 2017. ISSN 1877-0509. doi: https://doi.org/10.1016/j.procs.20 17.10.106. URL https://www.sciencedirect.com/science/article/pii/S1877050917321634. Arabic Computational Linguistics. [↗](https://doi.org/10.1016/j.procs.20)
8. Viraat Aryabumi, John Dang, Dwarak Talupuru, Saurabh Dash, David Cairuz, Hangyu Lin, Bharat Venkitesh, Madeline Smith, Jon Ander Campos, Yi Chern Tan, et al. Aya 23: Open weight releases to further multilingual progress. arXiv preprint arXiv:2405.15032, 2024. [↗](https://arxiv.org/abs/2405.15032) [our card](https://md.research.elamur.ai/free/en/papers/aryabumi-2024-aya-23-open-weight-releases/)
9. Parnia Bahar, Mattia Di Gangi, Nick Rossenbach, and Mohammad Zeineldeen. Take the hint: Improving arabic diacritization with partially-diacritized text, 2023. URL https://arxiv.org/abs/2306.03557. [↗](https://arxiv.org/abs/2306.03557) [our card](https://md.research.elamur.ai/free/en/papers/bahar-2023-take-hint-improving-arabic-diacritization/)
10. M Saiful Bari, Yazeed Alnumay, Norah A Alzahrani, Nouf M Alotaibi, Hisham A Alyahya, Sultan AlRashed, Faisal A Mirza, Shaykhah Z Alsubaie, Hassan A Alahmed, Ghadah Alabduljabbar, et al. Allam: Large language models for arabic and english. arXiv preprint arXiv:2407.15390, 2024. [↗](https://arxiv.org/abs/2407.15390) [our card](https://md.research.elamur.ai/free/en/papers/bari-2024-allam-large-language-models-arabic/)
11. Kareem Darwish, Hamdy Mubarak, and Ahmed Abdelali. Arabic diacritization: Stats, rules, and hacks. In Nizar Habash, Mona Diab, Kareem Darwish, Wassim El-Hajj, Hend Al-Khalifa, Houda Bouamor, Nadi Tomeh, Mahmoud El-Haj, and Wajdi Zaghouani (eds.), Proceedings of the Third Arabic Natural Language Processing Workshop, pp. 9–17, Valencia, Spain, April 2017a. Association for Computational Linguistics. doi: 10.18653/v 1/W17- 1302. URL https://aclanthology.org/W17-1302. [↗](https://aclanthology.org/W17-1302/) [our card](https://md.research.elamur.ai/free/en/papers/darwish-2017-arabic-diacritization-stats-rules-hacks/)
12. Kareem Darwish, Hamdy Mubarak, and Ahmed Abdelali. Arabic diacritization: Stats, rules, and hacks. In Nizar Habash, Mona Diab, Kareem Darwish, Wassim El-Hajj, Hend Al-Khalifa, Houda Bouamor, Nadi Tomeh, Mahmoud El-Haj, and Wajdi Zaghouani (eds.), Proceedings of the Third Arabic Natural Language Processing Workshop, pp. 9–17, Valencia, Spain, April 2017b. Association for Computational Linguistics. doi: 10.18653/v 1/W17- 1302. URL https://aclanthology.org/W17-1302. [↗](https://aclanthology.org/W17-1302/) [our card](https://md.research.elamur.ai/free/en/papers/darwish-2017-arabic-diacritization-stats-rules-hacks/)
13. Kareem Darwish, Ahmed Abdelali, Hamdy Mubarak, and Mohamed Eldesouki. Arabic diacritic recovery using a feature-rich bilstm model, 2020. URL https://arxiv.org/abs/2002.01207. [↗](https://arxiv.org/abs/2002.01207) [our card](https://md.research.elamur.ai/free/en/papers/darwish-2020-arabic-diacritic-recovery-using-feature/)
14. Mona Diab, Mahmoud Ghoneim, and Nizar Habash. Arabic diacritization in the context of statistical machine translation. In Proceedings of Machine Translation Summit XI: Papers, 2007. [find](https://scholar.google.com/scholar?q=Mona+Diab%2C+Mahmoud+Ghoneim%2C+and+Nizar+Habash%2E+Arabic+diacritization+in+the+context+of+statistical+machine+translation%2E+In+Proceedings+of+Machine+Translation+Summit+XI%3A+Papers%2C+2007%2E)
15. Ali Fadel, Ibraheem Tuffaha, Bara’ Al-Jawarneh, and Mahmoud Al-Ayyoub. Arabic text diacritization using deep neural networks, 2019a. URL https://arxiv.org/abs/1905.01965. [↗](https://arxiv.org/abs/1905.01965) [our card](https://md.research.elamur.ai/free/en/papers/fadel-2019-arabic-text-diacritization-using-deep/)
16. Ali Fadel, Ibraheem Tuffaha, Bara’ Al-Jawarneh, and Mahmoud Al-Ayyoub. Neural arabic text diacritization: State of the art results and a novel approach for machine translation. In Proceedings of the 6th Workshop on Asian Translation, pp. 215–225. Association for Computational Linguistics, 2019b. doi: 10.18653/v1/d19-5229. URL http://dx.doi.org/10.18653/v1/D19-5229. [↗](https://doi.org/10.18653/v1/d19-5229) [our card](https://md.research.elamur.ai/free/en/papers/fadel-2019-neural-arabic-text-diacritization-state/)
17. Nizar Habash, Anas Shahrour, and Muhamed Al Khalil. Exploiting arabic diacritization for high quality automatic annotation. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pp. 4298–4304, 2016. [our card](https://md.research.elamur.ai/free/en/papers/habash-2016-exploiting-arabic-diacritization-high-quality/)
18. Khalil Hennara, Sara Chrouf, Mohamed Motaism Hamed, Zeina Aldallal, Omar Hadid, and Safwan AlModhayan. Kuwain 1.5 b: An arabic slm via language injection. arXiv preprint arXiv:2504.15120, 2025. [↗](https://arxiv.org/abs/2504.15120) [our card](https://md.research.elamur.ai/free/en/papers/hennara-2025-kuwain-1-5b-arabic-slm/)
19. Ruba Kharsa, Ashraf Elnagar, and Sane Yagi. Bert-based arabic diacritization: A state-of-the-art approach for improving text accuracy and pronunciation. Expert Systems with Applications, 248:123416, 2024. ISSN 0957-4174. doi: https://doi.org/10.1016/j.eswa.2024.123416. URL https://www.sciencedirect.com/science/article/pii/S0957417424002811. [↗](https://doi.org/10.1016/j.eswa.2024.123416)
20. Vladimir Likic. The needleman-wunsch algorithm for sequence alignment. Lecture given at the 7th Melbourne Bioinformatics Course, Bi021 Molecular Science and Biotechnology Institute, University of Melbourne, pp. 1–46, 2008. [find](https://scholar.google.com/scholar?q=Vladimir+Likic%2E+The+needleman%2Dwunsch+algorithm+for+sequence+alignment%2E+Lecture+given+at+the+7th+Melbourne+Bioinformatics+Course%2C+Bi021+Molecular+Science+and+Biotechnology+Institute%2C+University+of+Melb)
21. Mohamed Maamouri, Seth Kulick, and Ann Bies. Diacritic annotation in the Arabic treebank and its impact on parser evaluation. In Nicoletta Calzolari, Khalid Choukri, Bente Maegaard, Joseph Mariani, Jan Odijk, Stelios Piperidis, and Daniel Tapias (eds.), Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Marrakech, Morocco, May 2008. European Language Resources Association (ELRA). URL [http://www.lrec-conf.org/proceedings/lrec2008/pdf/706_paper.pdf](http://www.lrec-conf.org/proceedings/lrec2008/pdf/706_paper.pdf). [↗](http://www.lrec-conf.org/proceedings/lrec2008/pdf/706_paper.pdf) [our card](https://md.research.elamur.ai/free/en/papers/maamouri-2008-diacritic-annotation-arabic-treebank-its/)
22. Mokthar Ali Hasan Madhfar and Ali Mustafa Qamar. Effective deep learning models for automatic diacritization of arabic text. IEEE Access, 9:273–288, 2021. doi: 10.1109/ACCESS.2020.3041676. [↗](https://doi.org/10.1109/access.2020.3041676)
23. Navid-AI. Yehia 7b preview. [https://huggingface.co/Navid-AI/Yehia-7B-preview](https://huggingface.co/Navid-AI/Yehia-7B-preview), 2025. [↗](https://huggingface.co/Navid-AI/Yehia-7B-preview)
24. Rani Nelken and Stuart M. Shieber. Arabic diacritization using weighted finite-state transducers. In Kareem Darwish, Mona Diab, and Nizar Habash (eds.), Proceedings of the ACL Workshop on Computational Approaches to Semitic Languages, pp. 79–86, Ann Arbor, Michigan, June 2005. Association for Computational Linguistics. URL https://aclanthology.org/W05-0711. [↗](https://aclanthology.org/W05-0711/) [our card](https://md.research.elamur.ai/free/en/papers/nelken-2005-arabic-diacritization-weighted-finite-state/)
25. Arfath Pasha, Mohamed Al-Badrashiny, Mona Diab, Ahmed El Kholy, Ramy Eskander, Nizar Habash, Manoj Pooleery, Owen Rambow, and Ryan Roth. MADAMIRA: A fast, comprehensive tool for morphological analysis and disambiguation of Arabic. In Nicoletta Calzolari, Khalid Choukri, Thierry Declerck, Hrafn Loftsson, Bente Maegaard, Joseph Mariani, Asuncion Moreno, Jan Odijk, and Stelios Piperidis (eds.), Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), pp. 1094–1101, Reykjavik, Iceland, May 2014. European Language Resources Association (ELRA). URL [http://www.lrec-conf.org/procee](http://www.lrec-conf.org/procee) dings/lrec2014/pdf/593_Paper.pdf. [↗](http://www.lrec-conf.org/procee) [our card](https://md.research.elamur.ai/free/en/papers/pasha-2014-madamira-fast-comprehensive-tool-morphological/)
26. Yasser Sabtan. Arabic part-of-speech tagging using a combined rule-based and data-driven approach. Digital Scholarship in the Humanities, 36(3):719–735, 2021. [find](https://scholar.google.com/scholar?q=Yasser+Sabtan%2E+Arabic+part%2Dof%2Dspeech+tagging+using+a+combined+rule%2Dbased+and+data%2Ddriven+approach%2E+Digital+Scholarship+in+the+Humanities%2C+36%283%29%3A719%E2%80%93735%2C+2021%2E)
27. Neha Sengupta, Sunil Kumar Sahu, Bokang Jia, Satheesh Katipomu, Haonan Li, Fajri Koto, William Marshall, Gurpreet Gosal, Cynthia Liu, Zhiming Chen, Osama Mohammed Afzal, Samta Kamboj, Onkar Pandit, Rahul Pal, Lalit Pradhan, Zain Muhammad Mujahid, Massa Baali, Xudong Han, Sondos Mahmoud Bsharat, Alham Fikri Aji, Zhiqiang Shen, Zhengzhong Liu, Natalia Vassilieva, Joel Hestness, Andy Hock, Andrew Feldman, Jonathan Lee, Andrew Jackson, Hector Xuguang Ren, Preslav Nakov, Timothy Baldwin, and Eric Xing. Jais and jais-chat: Arabic-centric foundation and instruction-tuned open generative large language models, 2023. URL https://arxiv.org/abs/2308.16149. [↗](https://arxiv.org/abs/2308.16149) [our card](https://md.research.elamur.ai/free/en/papers/sengupta-2023-jais-jais-chat-arabic-centric/)
28. Abderrahman Skiredj and Ismail Berrada. Arabic text diacritization in the age of transfer learning: Token classification is all you need, 2024. URL https://arxiv.org/abs/2401.04848. [↗](https://arxiv.org/abs/2401.04848) [our card](https://md.research.elamur.ai/free/en/papers/skiredj-2024-arabic-text-diacritization-age-transfer/)
29. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. [↗](https://arxiv.org/abs/2408.00118)
30. Silma Team. Silma. 2024. URL [https://www.silma.ai](https://www.silma.ai). [↗](https://www.silma.ai)
31. Jezia Zakraoui, Moutaz Saleh, Somaya Al-Maadeed, and Jihad Mohamed Alja’am. Arabic machine translation: A survey with challenges and future directions. IEEE Access, 9:161445–161468, 2021. [find](https://scholar.google.com/scholar?q=Jezia+Zakraoui%2C+Moutaz+Saleh%2C+Somaya+Al%2DMaadeed%2C+and+Jihad+Mohamed+Alja%E2%80%99am%2E+Arabic+machine+translation%3A+A+survey+with+challenges+and+future+directions%2E+IEEE+Access%2C+9%3A161445%E2%80%93161468%2C+2021%2E)
32. Taha Zerrouki and Amar Balla. Tashkeela: Novel corpus of arabic vocalized texts, data for auto-diacritization systems. Data in Brief, 11, 02 2017. doi: 10.1016/j.dib.2017.01.011. [↗](https://doi.org/10.1016/j.dib.2017.01.011)