Towards Enabling Meta-Learning from Target Models