Fionetune model always predicts same output class for new data

I have a problem, trained a model with bert which give around 0.90% on test data and I decide to use it on new data which were not annotated. When running the model, I keep getting the same class output. And I would like to know why ?

Can you help me ?