They are represented as a binary mask identifying
the two types of sequence in the model.
The tokenizer returns this mask as the "token_type_ids" entry:
thon

encoded_dict["token_type_ids"]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]

The first sequence, the "context" used for the question, has all its tokens represented by a 0, whereas the second
sequence, corresponding to the "question", has all its tokens represented by a 1.
Some models, like [XLNetModel] use an additional token represented by a 2.
transfer learning
A technique that involves taking a pretrained model and adapting it to a dataset specific to your task.