For token classification models, ([BertForTokenClassification]), the model expects a tensor of dimension
  (batch_size, seq_length) with each value corresponding to the expected label of each individual token.