Handwritten Text Recognition (HTR) for Arabic presents unique challenges due to the script’s cursive nature, varying writer styles, and morphological complexity. While modern Vision-Language Models (VLMs) have significantly advanced document parsing, their direct application to highly specific cursive domains requires strategic adaptation. This paper details our submission to the Nakba OCR competition, which adapts a 3B-parameter VLM to recognize historical Arabic manuscripts. We employ a progressive training pipeline that utilizes domain-matched data augmentation to bridge the gap between standard printed Arabic OCR and historical handwritten manuscripts. Moving beyond standard decoder-only Supervised Fine-Tuning (SFT), we fine-tune the entire encoder-decoder architecture using differential learning rates. This approach, followed by a final checkpoint merge, allows the model to better resolve the fine visual details of cursive Arabic script. Our final unified model (submitted under the team name Misraj AI) establishes a new state-of-the-art (SOTA) on the Nakba dataset, achieving a Word Er- ror Rate (WER) of 0.24 and a Character Error Rate (CER) of 0.08, and officially securing first place on the leaderboard.

In this paper, we explore the rich diversity of Arabic dialects by introducing a suite of pioneering models called Lahjawi. The primary model, Lahjawi-D2D, is the first designed for cross-dialect translation among 15 Arabic dialects. Furthermore, we introduce Lahjawi-D2MSA, a model designed to convert any Arabic dialect into Modern Standard Arabic (MSA). Both models are fine-tuned versions of Kuwain-1.5B an in-house built small language model, tailored for Arabic linguistic characteristics. We provide a detailed overview of Lahjawi’s architecture and training methods, along with a comprehensive evaluation of its performance. The results demonstrate Lahjawi’s success in preserving meaning and style, with BLEU scores of 9.62 for dialect-to-MSA and 9.88 for dialect-to- dialect tasks. Additionally, human evaluation reveals an accuracy score of 58% and a fluency score of 78%, underscoring Lahjawi’s robust handling of diverse dialectal nuances. This research sets a foundation for future advancements in Arabic NLP and cross-dialect communication technologies.