Gumball at QIAS 2025: Arabic LLM Automated Reasoning in Islamic Inheritance

Résumé

In this paper, we present a system for solving Islamic inheritance problems using large language models (LLMs), focusing on accurate reasoning in Arabic based on fara'id rules.Our approach is built on the Qwen3-4B model, quantized, and trained using the Unsloth framework for efficiency.We explore multiple training strategies: (1) retrieval-augmented generation (RAG) using fatwas from Islamweb, (2) supervised fine-tuning (SFT) on annotated inheritance datasets, (3) instruction tuning of a base Qwen model followed by GRPO training for multiple choice question solving, and (4) a two-stage pipeline involving SFT on a classical Islamic inheritance book followed by MCQ fine-tuning.Among these, the fourth approach achieved 97.2% accuracy, outperforming all other submissions and ranking our team first in the competition.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0