This model is a result of fine-tuning a Prune OFA 80% 1x4 block sparse pre-trained BERT-Large combined with knowledge distillation. This model yields the following results on SQuADv1.1 development set:
{"exact_match": 84.673, "f1": 91.174}

For further details see our paper, Prune Once for All: Sparse Pre-Trained Language Models, and our open source implementation available here.