Constructing Highly Inductive Contexts for Dialogue Safety through Controllable Reverse Generation - View it on GitHub