I am trying to fine tune the coQA dataset on 'bert-base-uncased' from hugging face. I am getting error while adding token positions. Upon checking I got some of the 'span_start' and 'span_end' in the 'answers' in training set as -1 which is throwing an error like "OverflowError: can't convert negative int to unsigned". I cannot figure out what should be the optimum step to deal with such issues.
Asked
Active
Viewed 24 times
