---
title: "Silah at QIAS 2026 — paper by Kurdi et al. (2026)"
description: "The paper describes team Silah's entry to the QIAS 2026 shared task, where a system reads an Arabic… 18 references, cited by 2 papers here."
authors: ["Ghader Kurdi", "Hanan Justanieah", "Hala Justanieah"]
year: 2026
venue: "The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks"
source: "https://aclanthology.org/2026.osact-1.27/"
doi: "10.63317/4iyrxakdovsm"
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/kurdi-2026-silah-qias-2026-fine-tuning/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/kurdi-2026-silah-qias-2026-fine-tuning/)

[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/9/) / Silah at QIAS 2026: Fine-Tuning vs. Retrieval-Augmented Generation for Islamic Inheritance Reasoning

# Silah at QIAS 2026: Fine-Tuning vs. Retrieval-Augmented Generation for Islamic Inheritance Reasoning

**Authors:** [Ghader Kurdi](https://md.research.elamur.ai/free/en/authors/ghader-kurdi/), Hanan Justanieah and Hala Justanieah · **Year:** 2026 · **Venue:** [The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks](https://md.research.elamur.ai/free/en/venues/osact/) ([Workshop on Open-Source Arabic Corpora and Processing Tools with Shared Tasks 2026](https://md.research.elamur.ai/free/en/venues/osact/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **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.osact-1.27/) · [PDF](https://aclanthology.org/2026.osact-1.27.pdf) · [DOI 10.63317/4iyrxakdovsm](https://doi.org/10.63317/4iyrxakdovsm) · [OpenAlex](https://openalex.org/W7166580230)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [2](https://md.research.elamur.ai/free/en/papers/kurdi-2026-silah-qias-2026-fine-tuning/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper describes team Silah's entry to the QIAS 2026 shared task, where a system reads an Arabic description of a deceased person's surviving relatives and must produce a full structured JSON solution of the Islamic inheritance case. This solution spans which heirs are eligible or blocked, the fixed portions they receive, the decision on whether to apply awl or radd, the base of the case (asl al-mas'ala), and the final allocation after ta'sil. The authors compare three strategies on four open-source LLMs:

- **RAG-P:** a five-stage pipeline that injects hand-curated rules into the prompt at each stage.
- **SFT:** plain LoRA fine-tuning on the 12,000-instance training set.
- **RAG-FT:** a fine-tuned model that receives retrieved rules only at inference.

Fine-tuning wins across the board. Fanar-1-9B-Instruct with SFT reaches the best MIR-E of 0.83. The authors conclude that, with current retrieval, models absorb procedural legal reasoning better from examples than from rules pasted into the prompt. They argue that rule retrieval needs to become more precise and selective before it can help.

## Key points
- **Data:** the QIAS 2026 dataset (Bouchekif et al., 2026) has 12,000 training items, 200 of which are held out for development, plus 500 test items. Each item contains a question, a reasoning trace, a category label and gold JSON.
- **Class imbalance:** awl cases make up 4.81% of the training set and radd cases 2.87%. In the test set the figures are 7.80% and 1.00%. The authors applied no rebalancing.
- **Models compared:** Qwen2.5-3B-Instruct, Meta-Llama-3-8B-Instruct, DeepSeek-R1-Distill-Qwen-7B and the Arabic-oriented Fanar-1-9B-Instruct.
- **Training setup:** LoRA with rank 16 and alpha 32 on all attention and MLP projections, one epoch, learning rate 2e-4, effective batch size 64, maximum sequence length 2048.
- **Rule base:** JSON files built from the authors' lecture slides, grouped as special cases (22), fixed shares (26), asaba (11), blocking (22), asl al-mas'ala (7), adjustment (4) and awl/radd (13).
- **Metric:** MIR-E is a weighted sum of stage scores. Heirs are weighted 0.30, shares 0.20, the awl/radd decision 0.10 and the final distribution 0.20. The awl/radd stage only counts as correct when the earlier stages are also correct.

## Interesting ideas
- Each rule record carries machine-checkable trigger fields (required_heirs, optional_heirs_any, heirs, conditions), so retrieval is a deterministic set-membership test rather than embedding similarity, and the reason a rule fired is always traceable.
- The task is split into ordered stages, each taking the previous stage's structured output plus its own small rule subset, instead of one giant prompt.
- The metric is staged and gated: a later decision earns credit only if the earlier stages are right, which exposes lucky guesses.
- The negative finding is useful: adding rules only at inference to a model trained without them did not help, suggesting train-time and test-time inputs must match.
- Before retrieval, heir mentions are normalised: compound phrases are split and lexical variants are mapped to one canonical singular form.

## For Mova

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

- **I1** (world): Compiled knowledge entries (verb classes, measure rules) could declare which entity types or states must be present before they apply, making every activation explainable.
- **I2** (answer): Maths word problem and school-test readers could chain explicit intermediate structures (quantities, then relations, then the computation), each with its own rule slice and its own check.
- **I3** (learn): Evaluation could score tree, frames, world and answer as weighted stages, with later credit conditional on earlier correctness, so a right answer built on a wrong world model is not rewarded.
- **I4** (learn): When an LLM teacher tier or student model is given extra context (retrieved frames, ontology facts), that context should also appear in training, or the gain should be checked against a no-context control.
- **I5** (tokens): Lemma and morphology output could feed a canonicalisation step that splits coordinated noun phrases and singularises kin and role nouns before rule or knowledge lookup.

*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 University](https://md.research.elamur.ai/free/en/universities/umm-al-qura-university/) — University, Makkah al Mukarramah, Saudi Arabia · 21 works · 274 citations

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,772 works · [🤝 Political Science and International Relations](https://md.research.elamur.ai/free/en/shelves/social-sciences/political-science-and-international-relations/) · 182 works

🧭 Topics (3): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,426 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works · [🔍 Information retrieval & search](https://md.research.elamur.ai/free/en/topics/information-retrieval/) · 473 works

## Cite

Cite as: Ghader Kurdi, Hanan Justanieah and Hala Justanieah. 2026. Silah at QIAS 2026: Fine-Tuning vs. Retrieval-Augmented Generation for Islamic Inheritance Reasoning. The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks.

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

```bibtex
@inproceedings{kurdi2026silah,
  title = {{Silah at QIAS 2026: Fine-Tuning vs. Retrieval-Augmented Generation for Islamic Inheritance Reasoning}},
  author = {Ghader Kurdi, Hanan Justanieah and Hala Justanieah},
  booktitle = {The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks},
  year = {2026},
  doi = {10.63317/4iyrxakdovsm},
  url = {https://doi.org/10.63317/4iyrxakdovsm}
}
```

- [← previous SHAPE: Unifying Safety, Helpfulness and Pedagogy for Educational LLMs](https://md.research.elamur.ai/free/en/papers/zhao-2026-shape-unifying-safety-helpfulness-pedagogy/)
- [next → SILO-BENCH: A Scalable Environment for Evaluating Distributed Coordination in Multi-Agent LLM Systems](https://md.research.elamur.ai/free/en/papers/zhang-2026-silo-bench-scalable-environment-evaluating/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/9/)

## References

18 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ummar Abbas, Mohammad Shahmeer Ahmad, Firoj Alam, Enes Altinisik, Ehsannedin Asgari, Yazan Boshmaf, Sabri Boughorbel, Sanjay Chawla, Shammur Chowdhury, Fahim Dalvi, Kareem Darwish, Nadir Durrani, Mohamed Elfeky, Ahmed Elmagarmid, Mohamed Eltabakh, Masoomali Fatehkia, Anastasios Fragkopoulos, Maram Hasanain, Majd Hawasly, Mus’ab Husaini, Soon-Gyo Jung, Ji Kim Lucas, Walid Magdy, Safa Messaoud, Abubakr Mohamed, Tasnim Mohiuddin, Basel Mousi, Hamdy Mubarak, Ahmad Musleh, Zan Naeem, Mourad Ouzzani, Dorde Popovic, Amin Sadeghi, Husrev Taha Sencar, Mohammed Shinoy, Omar Sinan, Yifan Zhang, Ahmed Ali, Yassine El Kheir, Xiaosong Ma, and Chaoyi Ruan. 2025. [Fanar: An arabic-centric multimodal generative ai platform](https://md.research.elamur.ai/free/en/papers/abbas-2025-fanar-arabic-centric-multimodal-generative/). arXiv preprint arXiv:2501.13944. [↗](https://arxiv.org/abs/2501.13944)
2. Nouar AlDahoul and Yasir Zaki. 2025. [Nyuad at qias shared task: Benchmarking the legal reasoning of llms in arabic islamic inheritance cases](https://md.research.elamur.ai/free/en/papers/aldahoul-2025-nyuad-qias-shared-task-benchmarking/). In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 861–866.
3. Abdessalam Bouchekif, Shahd Gaben, Samer Rashwani, Somaya Eltanbouly, Mutaz Al-Khatib, Heba Sbahi, Mohammed Ghaly, and Emad Mohamed. 2026. [MAWARITH: A dataset and benchmark for legal inheritance reasoning with LLMs](https://md.research.elamur.ai/free/en/papers/bouchekif-2026-mawarith-dataset-benchmark-legal-inheritance/). arXiv preprint arXiv:2603.07539. [↗](https://arxiv.org/abs/2603.07539)
4. Abdessalam Bouchekif, Samer Rashwani, Emad Soliman Ali Mohamed, Mutaz Alkhatib, Heba Sbahi, Shahd Gaben, Wajdi Zaghouani, Aiman Erbad, and Mohammed Ghaly. 2025a. [QIAS 2025: Overview of the shared task on islamic inheritance reasoning and knowledge assessment](https://md.research.elamur.ai/free/en/papers/bouchekif-2025-qias-2025-overview-shared-task/). In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 851–860.
5. Abdessalam Bouchekif, Samer Rashwani, Heba Sbahi, Shahd Gaben, Mutaz Al-Khatib, and Mohammed Ghaly. 2025b. [Assessing large language models on islamic legal reasoning: Evidence from inheritance law evaluation](https://md.research.elamur.ai/free/en/papers/bouchekif-2025-assessing-large-language-models-islamic/). In Proceedings of The Third Arabic Natural Language Processing Conference, pages 246–257.
6. DeepSeek-AI. 2025. [Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning](https://md.research.elamur.ai/free/en/papers/deepseek-ai-2025-deepseek-r1-incentivizing-reasoning-capability/).
7. Eman Elrefai, Mohamed Lotfy Elrefai, and Aml Hassan Esmail. 2025. [Gumball at qias 2025: Arabic llm automated reasoning in islamic inheritance](https://md.research.elamur.ai/free/en/papers/elrefai-2025-gumball-qias-2025-arabic-llm/). In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 953–959.
8. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. [The llama 3 herd of models](https://md.research.elamur.ai/free/en/papers/grattafiori-2024-llama-3-herd-models/). arXiv preprint arXiv:2407.21783. [↗](https://arxiv.org/abs/2407.21783)
9. Mohamed Motasim Hamed, Nada Ghneim, and Riad Sonbol. 2025. [Hiast at qias 2025: Retrieval-augmented llms with top-hit web evidence for arabic islamic reasoning qa](https://md.research.elamur.ai/free/en/papers/hamed-2025-hiast-qias-2025-retrieval-augmented/). In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 883–891.
10. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. 2022. [Lora: Low-rank adaptation of large language models](https://md.research.elamur.ai/free/en/papers/hu-2021-lora-low-rank-adaptation-large/). Iclr, 1(2):3.
11. Hala Justanieah and Ghader Kurdi. 2025. An evaluation of chatgpt’s responses to islamic inheritance law questions: Insights and challenges. In Proceedings of the Conference on Artificial Intelligence and Driving Change in Education, Saudi Arabia. [find](https://scholar.google.com/scholar?q=Hala+Justanieah+and+Ghader+Kurdi%2E+2025%2E+An+evaluation+of+chatgpt%E2%80%99s+responses+to+islamic+inheritance+law+questions%3A+Insights+and+challenges%2E+In+Proceedings+of+the+Conference+on+Artificial+Intelligence+)
12. Ghader Reda Kurdi. 2026. A systematic mapping study: Exploring islamic inheritance in computing research. International Journal of Informatics and Communication Technology (IJ-ICT). Accepted for publication (in press). [↗](https://doi.org/10.11591/ijict.v15i2.pp597-606)
13. Ghader Reda Kurdi and Hala Mohammad Justanieah. 2025. Enhancing evaluation practices for islamic inheritance calculation systems: Toward a standardized benchmark. Indonesian Journal of Electrical Engineering and Computer Science, 40(3):1549–1566. [↗](https://doi.org/10.11591/ijeecs.v40.i3.pp1549-1566)
14. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. [Retrieval-augmented generation for knowledge-intensive nlp tasks](https://md.research.elamur.ai/free/en/papers/lewis-2020-retrieval-augmented-generation-knowledge-intensive/). Advances in neural information processing systems, 33:9459–9474.
15. AL-Smadi Mohammad. 2025. [Qu-nlp at qias 2025 shared task: A two-phase llm fine-tuning and retrieval-augmented generation approach for islamic inheritance reasoning](https://md.research.elamur.ai/free/en/papers/smadi-2025-qu-nlp-qias-2025-shared/). In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 892–898.
16. Mohamed Lamine Mouhoub and Chahinez BOUCHEKIF. 2026. [Psl at qias 2026: Which models perform better in arabic inheritance reasoning?](https://md.research.elamur.ai/free/en/papers/mohammed-2026-psl-qias-2026-which-models/) In Proceedings of the 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7), Palma, Spain.
17. Wassim Swaileh, Mohammed-En-Nadhir Zighem, Hichem Telli, Salah Eddine Bekhouche, Abdellah Zakaria Sellam, and Fadi Dornaika. 2026. Cvpd at qias 2026: Rag-guided llm reasoning for al-mawarith share calculation and heir allocation. In Proceedings of the 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7), Palma, Spain. [find](https://scholar.google.com/scholar?q=Wassim+Swaileh%2C+Mohammed%2DEn%2DNadhir+Zighem%2C+Hichem+Telli%2C+Salah+Eddine+Bekhouche%2C+Abdellah+Zakaria+Sellam%2C+and+Fadi+Dornaika%2E+2026%2E+Cvpd+at+qias+2026%3A+Rag%2Dguided+llm+reasoning+for+al%2Dmawarith+share+cal)
18. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671. 10. Appendices 10.1. Examples of Inheritance Rules Example 1: { "id":"share_01", "title":  " ( ék. ð QË@) ðQ ®Ë@ H . Am• @ ", "text":  " . ðQ ®Ë@ H . Am• @ áÓ ék. ð QË@ ", "category":"fixed_share", "madhhab":"general", "heirs":[  " ék. ð P " ], "conditions":[], "lang":"ar" } Example 2: { "id": "special_case_solution_03", "title":   " éK PY» B@ éË A‚ÖÏ @ Ég ", "text":   h. ð QË@ (1 : éË A‚ÖÏ @ è Yë ú ¯ " Xñk. ð ÐY« Y J« ’ JË@ HQK  ñË@ ¨Q ®Ë@ ÐY« Qå„. IÊ  JË@ HQ  K Ð B@ (2 . HP@  ð ¨Q ¯ HP@ Y J« (3 . á JK QÒªË@ áÓ éË  A‚ÖÏ @ àñºK B @ð èñ kB @ XYªK ÐY«ð  I  k B@ ©Ó Ym.Ì '@ ¨AÒJk. @ : Q ¯ I.kA“ Xñk. ð ©Ó é  ®J  ®  ‚Ë@  Ë@ IÊ   €Yƒ ,ú ¯AJ .  K , éÖÞ A ®ÖÏ @) áÓ È@ñk B@ É ’ ¯ @ ú ¯ Yj.ÊË Q ¢ J K Y g AJ ¯ éË É ’ ¯ @ ɾË@ €Y ƒ éË  A‚ÖÏ @ è Yë ú ¯ð ,(ɾË@  I  k B@ Õæ ”  (4 .€Y‚Ë@ Ym.Ì '@ úÍ@ H. B I  k B@ ð @ é  ®J  ®  ‚Ë@ . á J   K B@ ¡k ÉJÓ Q» YÊË AÒî DJ K. Õæ„ ®  K ð ù ®K  . AÒJ ¯ ", "category": "special_case_solution", "madhhab": "general", "heirs":     k@ ", " H. B I  k@ "], [ " h. ð P ", " Ð@ ", " Yg. ", " é ®J ® ƒ  I "required_heirs": [ " h. ð P ", " Ð@ ", " Yg. "], "optional_heirs_any":     k@ ", " H. B I  k@ "], [ " é ®J ® ƒ  I "conditions": {}, "lang": "ar" } 10.2. Fine-Tuning Configuration and Training Details All experiments were implemented using the HuggingFace Transformers library and PyTorch, and executed with GPU acceleration. The training configurations for LoRA and model optimisation are summarised in Tables 2 and 3. Parameter Task type LoRA rank (r) LoRA alpha LoRA dropout Bias Target modules Value CAUSAL_LM 16 32 0.05 none q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj Table 2: LoRA configuration. Parameter Tokeniser Value Same as base model tokeniser 2048 AdamW 2 × 10 −4 0.01 Linear (no warmup) 1 4 1 16 64 eval_loss False FP16 when CUDA is available Enabled Disabled 42 Max sequence length Optimiser Learning rate Weight decay Learning rate scheduler Number of epochs Train batch size per device Eval batch size per device Gradient accumulation steps Effective train batch size Selection metric Greater is better Mixed precision Gradient checkpointing Use cache Seed Table 3: Training configuration. [↗](https://arxiv.org/abs/2407.10671)

*Updated 1 November 2026*