BERT classification - same output for every datapoint?

I modified the code here for a two-class classification task. However, the raw probability in evaluation stage seems identical for all data points. Is there something I may have a mistake?

Sorry for the bump, but i’m facing the very same issue with a bert pre-trained model. I’m also using eval() and it’s just when i try to predict.

Here’s my post link. Any help is welcome.