Model trained for 1 epoch on 1000 examples from the `adversarial_qa` dataset