Retraining a new model so that we have a matching .pth file.
Also, discarding more data on init to get better randomization.

Also skip the first loss values being generated since they're
biased towards "not lost" due to the initialization.