BERT EXTRACTION: Unable to get the result

[Jimntu]

Hi, when I try to implement the code of bert extraction experiment, the eval accuracy for both MNLI and SST-2 are lower than 50 percent. I doubt that the weight of the model is not updated and loss never decreases. However, I am a beginner in this field and I do not know how to get the loss after training each batch. Anyone can give a hint? Thanks a lot!!