People training local Piper voices avoid a failure that could silently reset a resumed run to its starting voice model. A saved setting could reload the original weights after the checkpoint restored the trained ones, while still reporting the later training stage.
Michael Hansen also restores the intended gradual reduction in learning rate. A published English base checkpoint still had its initial rates after 549,164 training steps. Fixed-length runs derive the reduction from their chosen duration; open-ended runs retain the older defaults.