Data-adaptive Safety Rules for Training Reward Models