For one thing, the model is trained on a language modelling task, not a question-answering task?