The pretraining objective is to predict the masked token based on the context.