The introduction of AI technology has enabled both personalized learning methods and content creation in modern educational settings. This paper presents a fine-tuned LLaMA-3.1-8B Instruct model for generating contextualized math problems for French-speaking middle school students aligned with the Moroccan educational curriculum. Our objective is to create adaptive learning content for both educators and learners through a large dataset of 5K exercises that are organized based on academic level and difficulty. For implementation, we used advanced techniques including gradient checkpointing and LoRA to optimize memory usage and training efficiency. Evaluation results confirm that the model generates educational materials with low loss and fast processing times. This work shows how LLMs can be applied to personalized learning and paves the way to future developments for AI learning content generation and related fields.