The training error occurred on the non Flux 1.0dev original model. So I changed the trainer and tried again, but I didn't expect it to be successful.
The version trained on the Unbelievable model has not been tested on the original Flux 1.0dev model yet.
