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