MarsadLab at PalmX Shared Task: An LLM Benchmark for Arabic Culture and Islamic Civilization
Year: 2025
Type: conference-paper
Abstract: This paper presents our submission to the PalmX 2025 Shared Task on Arabic cultural and religious knowledge comprehension.We focus on training large language models capable of representing domain-specific cultural and religious knowledge in Arabic.Our approach leverages parameter-efficient fine-tuning of the instruction-tuned Qwen2.5-7Bmodel using Low-Rank Adaptation (LoRA).To address the challenges of limited training data, we apply quantization-aware fine-tuning with 4bit precision, enabling efficient adaptation under constrained resources.The model is further aligned with the multiple-choice evaluation format to enhance task-specific reasoning.Without relying on external data augmentation, our system achieves competitive performance across both the Arabic General Culture and Islamic Culture subtasks, demonstrating the effectiveness of targeted fine-tuning for enriching cultural and religious knowledge representation in LLMs.On the blind test sets, our systems ranked 7 th and 4 th in the cultural and Islamic subtasks, respectively.To ensure reproducibility, we make our full codebase and experimental configurations available at https://github.com/rafiulbiswas/PalmX.
Institution: Northwestern University in Qatar
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037751