The Arabic language presents unique challenges for natural language processing due to its complex grammar, diverse dialects, and frequent omission of diacritics. This paper proposes a unified token-free model based on ByT5 that simultaneously performs spelling correction (including Jordanian dialect-to-Modern Standard Arabic (MSA) translation) and diacritization. Our approach uses task-specific prefixes (“correct:” for correction and “diacritize:” for combined correction and diacritization) to enable flexible multi-task learning. The model was fine-tuned on the JODA dataset (Jordanian dialect/MSA pairs) and high-quality Tashkeela subsets (Clean-50 and Clean-400), with synthetic errors injection to enhance robustness. Automatic evaluation showed an overall evaluation score of 78.06% on JODA and 92.45% on the combined test set of JODA and Tashkeela. Manual evaluation of 200 JODA samples revealed a character error rate of 4.41% and diacritic error rate of 1.32%, demonstrating practical efficacy in handling Arabic’s complexities.
