Training cascade separate Models

The approach seems to work here.
Could you post an executable code snippet, which would reproduce the issue you are seeing?