Learning to summarize from human feedback Nisan Stiennon Long Ouyang Jeff Wu

As language models become more powerful, training and evaluation are increasingly bottlenecked by the data and metrics used for a particular task.