“Using techniques like dropout and training for many more epochs on existing text data could yield more capable models, mitigating concerns about running out of data.”