Difficulties to obtain results on example/ scripts (train_supervised_bert.py)

[mttmta]

In train_supervised_bert.py :

• Are we supposed to get probant results when launching this script on the different datasets you suggest (--dataset wiki80 for example) ?
• It seems like we cannot get the model to make predictions. The model ends up always making the same class/relation prediction ; thus resulting in an accuracy corresponding to 1 / {number of relations}. For wiki80, we get 1.25% instead of the announced 84.57%.