BERT based model finetuned on MNLI with our custom training routine. Yields 60% accuraqcy on adversarial HANS dataset.
