ISL-NLP at PalmX 2025: Retrieval-Augmented Fine-Tuning for Arabic Cultural Question Answering

Résumé

Cultural understanding is essential for large language models (LLMs), particularly in the Arabic context where many models struggle to capture nuanced cultural elements.To address this gap, we propose a novel approach for Arabic cultural multiple-choice question answering that integrates retrieval-based training data augmentation with parameter-efficient fine-tuning.Our system employs Gemini 1 to retrieve contextual evidence for each question, selects candidate pairs, and adapts NileChat-3B by fine-tuning only three projection layers, reducing trainable parameters by 68.2% while preserving general language proficiency.On the PalmX 2025 Subtask 1 benchmark 2 , our system attains 67.60% accuracy on the blind test set, ranking 6 th overall and outperforming the NileChat-3B baseline by 3% on the development set.The model weights are publicly available at MohamedGomaa30/Ibn-Al-Nafs.

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