are pretrained transformer models initially trained to predict the 
next token given some input text.