Seems like pure overfitting. Since the model is able to see during training the same coocurrences that are used for evaluation, it is just about memorizing the

Seems like pure overfitting. Since the model is able to see during training the same coocurrences that are used for evaluation, it is just about memorizing the training set, which is trivial.