the model architecture of COVID-Twitter-BERT v2 MNLI

[founting]

Hi, I have a question about the training procesure of COVID-Twitter-BERT v2 MNLI. When it fine-tunes on MNLI dataset, does it use cross-encoder architecture and cross-entripy loss as objective?