Last and best models are linked together after training

For some reason, last and best models are intermixed during checkpoint saving.

I wonder if this is not a problem with our logging configuration for lightning.