Binary classifier just defaults to single value

Hello, I wonder if anyone in the community might help me.
I am working on a binary classifier (entangled/not entangled two qubit states). At first I implemented a Logistic Regression model which did okay, at best it had around 70% training and test accuracy, but then I implemented some one and two layer NN and, after training, found that my model just defaulted to choose one of the options (the model always returns 1 sometimes or 0 other times).
I thought that having 70% training accuracy with a logistic regression was an encouraging sign but now I’m confused about getting only 50% accuracy.
Any suggestions?