Bridging Model-based Safety and Model-free Reinforcement Learning through System Identification of Low Dimensional Linear Models