Personalized learning has become a crucial aspect of modern education, where recommendation systems are important in tailoring educational resources to individual learners. This study explores the fine-tuning of LLaMA, a large language model (LLM), to enhance educational recommendations using K-12 datasets and the OpenEdX learning platform. By leveraging contextual embeddings and transformer-based learning, LLaMA surpasses traditional collaborative filtering (CF) and content-based filtering (CBF) methods in capturing the semantic relationships between students and learning materials. We formalize the problem as an optimization task where the model predicts the most relevant resources based on student profiles, past interactions, and learning preferences. The evaluation, conducted across multiple learning scenarios, demonstrates significant improvements in ranking efficiency, particularly in cold-start and sparse data conditions. Using precision, recall, and NDCG as evaluation metrics, LLaMA achieves a notable increase in recommendation accuracy over hybrid CF + CBF models. The study further discusses computational trade-offs, fairness concerns, and real-time adaptability, providing insights into future research directions for AI-driven education. Our findings highlight the potential of fine-tuned LLaMA models in delivering highly personalized, efficient, and scalable learning experiences.