One simple approach would be:
- split your dataset into the y=0 set and the y = 1 set
- do the 80/20 training/test split on the y=0 and y=1 separately
- combine the two training sets and the two test sets
Bear in mind that highly imbalanced datasets like this have pitfalls for analysis. It's too much to describe here, but if you search keywords like imbalanced dataset machine learning, you'll be able to read about the problems, and some potential solutions.
