Training is done by partially hiding (masking) text and making a (decoder-only) model predict what's missing. Training text which humans prefer as model responses is called RLHF. It's an expensive and small text dataset.