Transfer Learning with BERT transformers

Question

My output labels are one-hot encoded in the following format: Positive, Negative, Mixed, Neutral with 1s and 0s e.g. [1 0 0 0] representing a text as being Positive