Model Fusion through Bayesian Optimization in Language Model Fine-Tuning