Large language models have dramatically transformed various fields by enhancing the capabilities of artificial intelligence in understanding and generating human language. The novel concept of predicting critical mixture ratios for continual pre-training introduces a significant advancement in optimizing the performance and efficiency of such models. Through leveraging advanced machine learning techniques, the study develops a predictive framework that systematically varies mixture ratios and evaluates their impact on model performance. The results highlight substantial improvements in accuracy, precision, recall, and F1 score, demonstrating the efficacy of optimized mixture ratios. Furthermore, the study shows that models trained with these optimized ratios exhibit enhanced generalization capabilities, maintaining high performance across unseen datasets. The research also addresses the reduction in training time, translating to operational efficiency gains and cost savings. Despite the promising findings, the study acknowledges several limitations, including the need for more diverse datasets and a deeper exploration of hyperparameter interactions. Future research directions include expanding the scope of data, integrating reinforcement learning techniques, and developing sophisticated validation methods. The study's insights and methodologies provide a valuable foundation for future research and practical implementations, aiming to create more effective and adaptable language models for various applications.