Sampling and Backup - Nerogar/OneTrainer GitHub Wiki

Backup

Backup After

Creates a backup after the set interval. These backups serve as restore points, allowing you to resume training from a previous checkpoint if needed.

Rolling Backup

Requires a value for Rolling Backup Count. Instead of keeping all backups, it retains only the last X backups, automatically deleting older ones to manage storage efficiently.

Save Every

Since no LoRA training will produce a perfect result after a fixed number of steps, it's useful to set the Save Every value to a reasonable interval to create frequent backups. This allows you to compare multiple checkpoints and choose the best result at the end, rather than relying on a single final save. By evaluating different stages of training, you can find the optimal balance between quality and overtraining, ensuring the best outcome of your training.

Skip First

Skips the first X number of saves. This can be useful if the initial training steps are not relevant for your comparisons.

Save Filename Prefix

By default, saves follow the naming format: YYYY-MM-DD_hh-mm-ss-save-Step-Epoch. This setting allows you to define a custom prefix for your frequent save files, making it easier to organize and identify different training runs.