Bad performance with trained models

[sdvies]

Hi! I am trying to build a CoQA model by following the steps that appear in the non docker alternative readme. After downloading the datasets "training set" and "dev set" from CoQA datasets and having trained the model, I found that it does not work properly. Using clinicalbert model and asking the same that you use in chatbots' article, we obtain the following:

Do you know what I am doing wrong? Because I have seen that in chatbots' article it works perfectly.

Thank you very much!