Published in the Proceedings of The Third Arabic Natural Language Processing Conference (ArabicNLP 2025), co-located with EMNLP 2025, Suzhou, China.
The paper introduces Shawarma Chats, 30,000 six-turn dialogues grounded in Wikipedia content and split evenly across Modern Standard, Egyptian and Maghrebi Arabic, built with LLM generation, a self-repair loop and native-speaker checks. LoRA fine-tuning six open LLMs (1B to 24B) on it gave consistent gains, especially for models larger than 7B.
Authors: Kamyar Zeinalipour, Mohamed Zaky Saad, Oumaima Attafi, Marco Maggini, Marco Gori.
Read the paper | Code | Models and data
All the lab's language releases are listed in Language models and datasets.