The development of Arabic natural language processing (NLP) applications and large language models (LLMs) faces substantial challenges, primarily due to the scarcity of high-quality native Arabic datasets. To address this critical gap, we present DIA2 (a Comprehensive and Diverse Diacritized Modern Standard Arabic Corpus), a novel dataset curated from 28 diverse, carefully selected Arabic sources. DIA2 emphasizes the use of original Arabic text and explicitly avoids machine-translated content. The corpus incorporates substantial amounts of text from books, news articles, and poetry, and employs extensive data preprocessing to support NLP research and LLM development. Our preprocessing pipeline includes rigorous text cleaning, URL- and document-level deduplication, and automatic diacritization, while preserving a gold diacritized subset derived from manually annotated sources. The resulting corpus comprises over 140 GB of high-quality text, containing more than 26 million unique words and 41.9 billion tokens. To evaluate the proposed pipeline, we conducted controlled continued pretraining experiments using Llama3.1-8B on both raw and processed subsets of DIA2. The model trained on processed data consistently outperformed its counterpart across multiple Arabic evaluation benchmarks. These results highlight the positive impact of systematic preprocessing and the utility of DIA2 in empowering native Arabic LLMs and downstream NLP tasks.

In light of the growing adoption of large language models (LLMs) as educational tutors, it is crucial to effectively evaluate their pedagogical capabilities across multiple dimensions. Toward this goal, we address the Mistake Identification sub-task of the BEA 2025 Shared task, aiming to assess the accuracy of tutors in detecting and identifying student errors. We experiment with several LLMs, including GPT-4o-mini, Mistral-7B, and Llama-3.1-8B, evaluating them in both zero-shot and fine-tuned settings. To address class imbalance, we augment the training data with synthetic examples, targeting underrepresented labels, generated by Command R+. Our GPT-4o model finetuned on the full development set achieves a strict macro-averaged F1 score of 71.63%, ranking second in the shared task. Our work highlights the effectiveness of fine-tuning on task-specific data and suggests that targeted data augmentation can further support LLM performance on nuanced pedagogical evaluation tasks.