We'll use the CoLA dataset from the GLUE benchmark,
since it's a simple binary text classification task, and just take the training split for now.