Learning Robust Representations Across Task Oriented And Zero Shot Models