I looked at the HuggingFace documentation for the
BertForQuestionAnswering model, and realized that the model gives out two logit outputs - start_logits and end_logits. How can I determine what the actual predicted answer is based on this two logits? Thank you,
start_logits indicates the (most probable) starting token , while the
end_logits indicates the ending token. So combining start and end tokens (and every token in between), you will have the “acutal predicted answer”, which is a string in this case.