Aligning pretrained language models with instruction data generated by themselves.
Expanding natural instructions