That sounds like overfitting. Your model is "too complex" for your training data. It learns the few training examples by heart but it cannot generalize. Think of a 9th-order polynominal that you try to fit a line with 9 Datapoints. Its corret at your support points but does wild swings everywhere else.

Thats why the training accuracy gets better and better (your support points), but validation stays a a low level (everywhere else).