Hey,

If I remember correctly the only difference is that the labels are an array instead of a single value, no?

Also, here is a NER example using BERT in a BiLSTM — CRF architecture. Plus there is the option to set the layer as trainable (i.e. to be fine-tuned) or not.

Google Colaboratory

Note that the code was very rushed so there many improvements possible. Also, there are some parts that need a further explanation but I may as well and grab the chance to write a blog post about it.

Hope this helps mate.