Fix validation values json in continue_training
Summary
This MR introduces the following changes
- Fixes an issue where when you use the
continue_training
option, the validation json (where the rejections are saved per epoch and so on) is overwritten when the new training starts.