Year: 2025
Type: conference-paper
Abstract: Large Language Models (LLMs) have become central to natural language processing, but their performance in low-resource cultural domains remains limited, mainly due to the dominance of English data in training.This limitation is especially evident in open small models.Evaluating and improving LLMs' performance in Arabic culture is therefore necessary.This paper presents Phoenix and PhoenixIs, two models fine-tuned for the Palmx 2025 general culture and Islamic culture subtasks.Phoenix uses the Palmx-GC and Palmx-IC datasets as seed data and applies diverse data augmentation strategies to construct an enriched fine-tuning dataset.Phoenix achieves an accuracy of 71.35% on the general culture subtask, while PhoenixIs reaches 83.82% on the Islamic culture subtask.
Authors: Houdaifa Atou, Issam Ait Yahia, Ismail Berrada
Institution: Université Mohammed VI Polytechnique
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037750