Jenin at AraGenEval Shared Task: Parameter-Efficient Fine-Tuning and Layer-Wise Analysis of Arabic LLMs for Authorship Style Transfer and Classification
Year: 2025
Type: conference-paper
Abstract: We benchmark two adaptation strategies for Arabic LLMs across three tasks in the AraGenEval Shared Task: (1) parameterefficient fine-tuning (LoRA) applied to decoder-based generative models (Gemma, Qwen) for author style transfer, and (2) full fine-tuning applied to encoder-based models (AraBERTv2, AraModernBert) for author classification and humanmachine text detection.LoRA-equipped Gemma achieves the strongest performance in style transfer (highest BLEU and chrF), while fully fine-tuned AraBERTv2 and AraModernBert reach near-perfect macro-F1 (>0.99) in classification and detection.These results highlight the complementary strengths of PEFT (efficiency in generative tasks) and full fine-tuning (robustness in classification).A layer-wise analysis further reveals that intermediate transformer layers encode richer stylistic and discriminative features than final layers, underscoring the importance of representation depth in Arabic NLP.All code and models are available at: https://github.com/mtami/AraG enEval2025.
Institution: Arab American University
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4416037822
FWCI:
Cites:
Cited by:
Related to: