Predictions from BERT give negative predictions

I was trying to finetune Bertweet model on 2 category sequence classification task but the predictions from the model does not seem right, the loss and accuracy also does not seem to improve at all between training epochs. I am not sure where I am doing wrong, may anyone help?