Training techniques used to make a model refuse harmful requests and behave responsibly, reducing the risk of misuse.