---
title: "Enabling Structured Reasoning in Sindhi with Culturally"
description: "Large language models handle reasoning poorly in Sindhi, an Indo-Aryan language with over 40 million… 13 references, cited by 1 paper here."
authors: ["Mehak Mehak", "Kamyar Zeinalipour", "Pireh Soomro", "Cristiano Chesi", "Marco Gori", "Marco Maggini"]
year: 2026
venue: "Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)"
source: "https://aclanthology.org/2026.loreslm-1.22/"
doi: "10.18653/v1/2026.loreslm-1.22"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/mehak-2026-enabling-structured-reasoning-sindhi-culturally/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/mehak-2026-enabling-structured-reasoning-sindhi-culturally/)

[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/12/) / Enabling Structured Reasoning in Sindhi with Culturally Grounded Instruction Tuning

# Enabling Structured Reasoning in Sindhi with Culturally Grounded Instruction Tuning

**Authors:** Mehak Mehak, [Kamyar Zeinalipour](https://md.research.elamur.ai/free/en/authors/kamyar-zeinalipour/), Pireh Soomro, [Cristiano Chesi](https://md.research.elamur.ai/free/en/authors/cristiano-chesi/), [Marco Gori](https://md.research.elamur.ai/free/en/authors/marco-gori/) and [Marco Maggini](https://md.research.elamur.ai/free/en/authors/marco-maggini/) · **Year:** 2026 · **Venue:** [Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)](https://md.research.elamur.ai/free/en/venues/loreslm/) ([Workshop on Language Models for Low-Resource Languages 2026](https://md.research.elamur.ai/free/en/venues/loreslm/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.loreslm-1.22/) · [PDF](https://aclanthology.org/2026.loreslm-1.22.pdf) · [DOI 10.18653/v1/2026.loreslm-1.22](https://doi.org/10.18653/v1/2026.loreslm-1.22) · [OpenAlex](https://openalex.org/W7140143216)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · 1 in our library, of the 149,439 papers whose bibliographies we read

## Summary

Large language models handle reasoning poorly in Sindhi, an Indo-Aryan language with over 40 million speakers. Earlier Sindhi NLP work focused on basic resources such as word embeddings, POS tagging and sentiment analysis. The authors built Sindhi-Reasoning-Instruct, a reasoning instruction corpus covering deductive, inductive, causal and commonsense tasks. Source texts came from Sindhi and English Wikipedia and from the Sindhi newspaper Daily Kawish. English material was machine-translated and then corrected by native linguists. GPT-4 drafted the reasoning items, and two trained Sindhi linguists checked their logic, fixed the script and adapted cultural details to the Sindh region. Six LLaMA and Mistral models (1B–24B) were then adapted with LoRA. After tuning, all models scored higher on automatic metrics and human ratings, wrote Perso-Arabic Sindhi correctly and followed SOV word order. Reasoning quality grew with model size, and Mistral-Small-24B performed best. The work suggests that a small, carefully curated native dataset can give a low-resource language real reasoning ability, not only translation.

## Key points

- **Corpus:** Sindhi-Reasoning-Instruct has 2,190 instruction–context–output triplets built from 1,200 passages. About 80% are encyclopedic and 20% come from Daily Kawish news. Roughly 12% of GPT-4 drafts were dropped or rewritten for logical errors.
- **Models:** LoRA was applied to LLaMA-3.2-1B, LLaMA-3.2-3B, LLaMA-3.1-8B, Mistral-7B v0.3, Mistral-Nemo and Mistral-Small-24B. The corpus was profiled with Stanza (model sd).
- **Human ratings:** Gains ranged from about 38% to 141%. Mistral-Small-24B improved most, rising from 1.49 to 3.59 (+141%). For the same model, ROUGE-L rose from 0.288 to 0.441 and BLEU from 0.123 to 0.257.
- **Reasoning accuracy (Mistral-Small-24B):** inductive 83%, commonsense 82%, deductive 80%, causal 77%. It led in every reasoning category.
- **Limitations:** The dataset is small, and most of it is translated from English. English-centric tokenizers split Sindhi script inefficiently. Human evaluation relied on only a few experts.

*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 (5): [🖥️ 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 · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 2,869 works

🧭 Topics (2): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,320 works · [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,426 works

## Cite

Cite as: Mehak Mehak, Kamyar Zeinalipour, Pireh Soomro, Cristiano Chesi, Marco Gori and Marco Maggini. 2026. Enabling Structured Reasoning in Sindhi with Culturally Grounded Instruction Tuning. Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026).

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

```bibtex
@inproceedings{mehak2026enabling,
  title = {{Enabling Structured Reasoning in Sindhi with Culturally Grounded Instruction Tuning}},
  author = {Mehak Mehak, Kamyar Zeinalipour, Pireh Soomro, Cristiano Chesi, Marco Gori and Marco Maggini},
  booktitle = {Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loreslm-1.22},
  url = {https://doi.org/10.18653/v1/2026.loreslm-1.22}
}
```

- [← previous Ellat at SemEval-2026 Task 11: Comparing Encoder and Decoder Models for Syllogistic Reasoning](https://md.research.elamur.ai/free/en/papers/memar-2026-ellat-semeval-2026-task-11/)
- [next → Energy Matching based Preference Learning for Diffusion Language Models](https://md.research.elamur.ai/free/en/papers/shankar-2026-energy-matching-based-preference-learning/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/12/)

## References

13 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Wazir Ali, Jay Kumar, Junyu Lu, and Zenglin Xu. 2021. Word embedding-based new corpus for low-resource language: Sindhi. arXiv preprint arXiv:1911.12579. — [arXiv:1911.12579](https://arxiv.org/abs/1911.12579) [find](https://scholar.google.com/scholar?q=Wazir+Ali%2C+Jay+Kumar%2C+Junyu+Lu%2C+and+Zenglin+Xu%2E+2021%2E+Word+embedding%2Dbased+new+corpus+for+low%2Dresource+language%3A+Sindhi%2E+arXiv+preprint+arXiv%3A1911%2E12579%2E+%E2%80%94+%5BarXiv%3A1911%2E12579%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F191)
2. Wazir Ali, Saifullah Tumrani, Jay Kumar, and Tariq Rahim Soomro. 2024. An evaluation of sindhi word embedding in semantic analogies and downstream tasks. arXiv preprint arXiv:2408.15720. — [arXiv:2408.15720](https://arxiv.org/abs/2408.15720) [find](https://scholar.google.com/scholar?q=Wazir+Ali%2C+Saifullah+Tumrani%2C+Jay+Kumar%2C+and+Tariq+Rahim+Soomro%2E+2024%2E+An+evaluation+of+sindhi+word+embedding+in+semantic+analogies+and+downstream+tasks%2E+arXiv+preprint+arXiv%3A2408%2E15720%2E+%E2%80%94+%5BarXiv%3A2408)
3. Abdullatif Köksal, Marion Thaler, Ayyoob Imani, Ahmet Üstün, Anna Korhonen, and Hinrich Schütze. 2024. Muri: High-quality instruction tuning datasets for low-resource languages via reverse instructions. arXiv preprint arXiv:2409.12958. [find](https://scholar.google.com/scholar?q=Abdullatif+K%C3%B6ksal%2C+Marion+Thaler%2C+Ayyoob+Imani%2C+Ahmet+%C3%9Cst%C3%BCn%2C+Anna+Korhonen%2C+and+Hinrich+Sch%C3%BCtze%2E+2024%2E+Muri%3A+High%2Dquality+instruction+tuning+datasets+for+low%2Dresource+languages+via+reverse+instruction)
4. Edoardo Maria Ponti, Goran Glavaš, Olga Majewska, Qianchu Liu, Ivan Vulić, and Anna Korhonen. 2020. [XCOPA: A multilingual dataset for causal commonsense reasoning](https://md.research.elamur.ai/free/en/papers/ponti-2020-xcopa-multilingual-dataset-causal-commonsense/). In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362–2376.
5. Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. [Stanza: A Python natural language processing toolkit for many human languages](https://md.research.elamur.ai/free/en/papers/qi-2020-stanza-python-natural-language-processing/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 101–108. Associat…
6. Mutee U. Rahman Arain and Mohammad Iqbal Bhatti. 2010. Finite state morphology and sindhi noun inflections. In Proceedings of the 24th Pacific Asia Conference on Language, Information and Computation (PACLIC 24), pages 669–676. [find](https://scholar.google.com/scholar?q=Mutee+U%2E+Rahman+Arain+and+Mohammad+Iqbal+Bhatti%2E+2010%2E+Finite+state+morphology+and+sindhi+noun+inflections%2E+In+Proceedings+of+the+24th+Pacific+Asia+Conference+on+Language%2C+Information+and+Computation+)
7. Raghav Ramji and Keshav Ramji. 2025. Inductive linguistic reasoning with large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 22783–22810. [find](https://scholar.google.com/scholar?q=Raghav+Ramji+and+Keshav+Ramji%2E+2025%2E+Inductive+linguistic+reasoning+with+large+language+models%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+ACL+2025%2C+pages+22783%E2%80%9322810%2E)
8. Yu Sheng, Wanting Wen, Linjing Li, and Daniel Zeng. 2025. Evaluating generalization capability of language models across abductive, deductive and inductive logical reasoning. In Proceedings of the 31st International Conference on Computational Linguistics (COLING), pages 4945–4957. [find](https://scholar.google.com/scholar?q=Yu+Sheng%2C+Wanting+Wen%2C+Linjing+Li%2C+and+Daniel+Zeng%2E+2025%2E+Evaluating+generalization+capability+of+language+models+across+abductive%2C+deductive+and+inductive+logical+reasoning%2E+In+Proceedings+of+the+31s)
9. Irum Naz Sodhar, Suriani Sulaiman, and Abdul Hafeez Buller. 2023. Morphology-assisted sindhi text analysis for natural language processing applications. Indian Journal of Science and Technology, 16(35):2894–2901. [find](https://scholar.google.com/scholar?q=Irum+Naz+Sodhar%2C+Suriani+Sulaiman%2C+and+Abdul+Hafeez+Buller%2E+2023%2E+Morphology%2Dassisted+sindhi+text+analysis+for+natural+language+processing+applications%2E+Indian+Journal+of+Science+and+Technology%2C+16%2835)
10. Emily Vaillancourt and Christopher Thompson. 2024. Instruction tuning on large language models to improve reasoning performance. TechRxiv. [find](https://scholar.google.com/scholar?q=Emily+Vaillancourt+and+Christopher+Thompson%2E+2024%2E+Instruction+tuning+on+large+language+models+to+improve+reasoning+performance%2E+TechRxiv%2E)
11. Hongbin Zhang, Kehai Chen, Xuefeng Bai, Yang Xiang, and Min Zhang. 2024. Lingualift: An effective two-stage instruction tuning framework for low-resource language reasoning. arXiv preprint arXiv:2412.12499. [find](https://scholar.google.com/scholar?q=Hongbin+Zhang%2C+Kehai+Chen%2C+Xuefeng+Bai%2C+Yang+Xiang%2C+and+Min+Zhang%2E+2024%2E+Lingualift%3A+An+effective+two%2Dstage+instruction+tuning+framework+for+low%2Dresource+language+reasoning%2E+arXiv+preprint+arXiv%3A2412%2E)
12. Hongbin Zhang, Kehai Chen, Xuefeng Bai, Yang Xiang, and Min Zhang. 2025. Lingualift: An effective two-stage instruction tuning framework for low-resource language reasoning. arXiv preprint arXiv:2412.12499. [find](https://scholar.google.com/scholar?q=Hongbin+Zhang%2C+Kehai+Chen%2C+Xuefeng+Bai%2C+Yang+Xiang%2C+and+Min+Zhang%2E+2025%2E+Lingualift%3A+An+effective+two%2Dstage+instruction+tuning+framework+for+low%2Dresource+language+reasoning%2E+arXiv+preprint+arXiv%3A2412%2E)
13. Guneet Singh Kohli, Arghyadeep Sen, Sambit Sekhar, Shashikanta Sahoo, Shantipriya Parida, Satya Ranjan Dash, and Ondřej Bojar. 2023. Bengaligpt: An instruction following llama model for bengali. arXiv preprint arXiv:2312.12879. — [arXiv:2312.12879](https://arxiv.org/abs/2312.12879) [find](https://scholar.google.com/scholar?q=Guneet+Singh+Kohli%2C+Arghyadeep+Sen%2C+Sambit+Sekhar%2C+Shashikanta+Sahoo%2C+Shantipriya+Parida%2C+Satya+Ranjan+Dash%2C+and+Ondr%CC%8Cej+Bojar%2E+2023%2E+Bengaligpt%3A+An+instruction+following+llama+model+for+bengali%2E+arXi)

*Updated 31 October 2026*