MT-DNN not working in a translation of squad

[ju-resplande]

When using a Portuguese Translation of squad 1.1 with initial checkpoint as neuralmind/bert-base-portuguese-cased, the prediction step throws a KeyError specifically in squad_utils.py extract_answer function:

Noticing that the error was related to questions with biggest context paragraph, I set +10 in DOC_STRIDE in prepro_std.py and, additionally, executed training with batch_size = 20. The first epoch ran without errors, however in the second about 7 KeyErrors errors related to these answers bound were encountered. I tried increasing DOC_STRIDE in 20, 30, 40, 50 and it still did not work.

The dataset was also run on hugging face using run_squad.py without errors.