Training doesn't generalize to predict validation and test set well, how to improve generalization?

Question

I have built a 1D CNN regression model using features from the embedding layer by keras, to predict a score of some sequences.

Typical performance of my