---
base_model: QCRI/Fanar-1-9B
library_name: peft
---

---
license: apache-2.0
language:
- ar
metrics:
- accuracy
base_model:
- QCRI/Fanar-1-9B
pipeline_tag: question-answering
---
# Fanar-1-9B-Islamic-Inheritance-Reasoning

This repository contains the model **Fanar-1-9B-Islamic-Inheritance-Reasoning**, developed by **QU-NLP** for the **QIAS 2025 Shared Task (SubTask 1: Islamic Inheritance Reasoning)**.  
The model is fine-tuned with **LoRA** and integrated into a **Retrieval-Augmented Generation (RAG)** pipeline, enabling mid-scale Arabic LLMs to perform domain-specific reasoning in **Islamic inheritance law (ʿIlm al-Farāʾiḍ)**.

---

## 📖 Description

Islamic inheritance law involves complex reasoning over:
- Identifying eligible heirs  
- Applying Qurʾanic fixed-share rules  
- Handling multiple inheritance scenarios  
- Performing fractional and numerical calculations  

Our approach fine-tuned **Fanar-1-9B** with **LoRA adapters** and integrated it into a **RAG pipeline** to ground responses in authoritative Islamic sources.  

**Results at QIAS 2025**:
- **Overall Accuracy:** 85.8%  
- **Advanced Reasoning:** 97.6% (outperforming Gemini 2.5 and OpenAI o3)  
- Surpassed zero-shot prompting of GPT-4.5, LLaMA, Mistral, and ALLaM.  

---

## 🧩 Example

**السؤال:** مات وترك: ابن ابن عم شقيق و بنت (5) و أم الأم و ابن عم الأب، كم عدد الأسهم التي تحصل عليها بنت (5) قبل تصحيح المسألة؟  

**الخيارات:**  
A) سهمان  
B) 0 سهم  
C) 6 أسهم  
D) 5 أسهم  
E) 4 أسهم  
F) 3 أسهم  

✅ **Model Output: E**  
> بنت (5) تأخذ أربعة أسهم قبل التصحيح.  

![Example Output](31a90e44-8b2e-4919-b38f-1df6b4ce8d95.png)

---



## 📊 Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("QU-NLP/Fanar-1-9B-Islamic-Inheritance-Reasoning")
model = AutoModelForCausalLM.from_pretrained("QU-NLP/Fanar-1-9B-Islamic-Inheritance-Reasoning")

question = "مات وترك: ابن ابن عم شقيق و بنت (5) و أم الأم و ابن عم الأب..."
options = ["سهمان", "0 سهم", "6 أسهم", "5 أسهم", "4 أسهم", "3 أسهم"]

# Prepare a prompt using RAG-retrieved context
prompt = f"السؤال: {question}\n\nالخيارات:\n" + "\n".join([f"{chr(65+i)}) {opt}" for i,opt in enumerate(options)]) + "\n\nاختر الحرف الصحيح:"

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Citation

If you use this model in your research, please cite the following paper:


@inproceedings{QU-NLP-QIAS2025,
  author    = {Mohammad AL-Smadi},
  title     = {QU-NLP at QIAS 2025 Shared Task: A Two-Phase LLM Fine-Tuning and Retrieval-Augmented Generation Approach for Islamic Inheritance Reasoning},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference (ArabicNLP 2025)},
  year      = {2025},
  publisher = {Association for Computational Linguistics},
  note      = {Suzhou, China, Nov 5--9},
  url       = {https://arabicnlp2025.sigarab.org/}
}