AllenAi's <i>scibert_scivocab_uncased</i> fine-tuned on SQuAD 2.0 achieving F1 = 86.85.

To load the model:

from transformers import BertTokenizerFast
from transformers import BertForQuestionAnswering

tokenizer = BertTokenizerFast.from_pretrained("LoudlySoft/scibert_scivocab_uncased_squad")
model = BertForQuestionAnswering.from_pretrained("LoudlySoft/scibert_scivocab_uncased_squad")