Quick questions on resources used for the full train

[glmcdona]

Couple quick questions if possible! Looking to do a full finetune as an experiment with a i2v dataset I've been working on.

• train.sh:
  • I note a 2e-5 learning rate. Was this the learning used for the full train of the model the entire time? Or was a learning rate schedule ran or something? Was 2e-5 the ending learning rate, or starting learning rate where it was lowered throughout the training? Trying to understand if this 2e-5 is a reasonable starting learning rate for a finetune, or if I should start it closer to 1e-4 as seen in the Lora config, then slowly ramp it down to 2e-5 towards the end.
  • train_batch_size of 1, does this mean a batch size of 1 was used? Should this be raised?
• Training resources
  • How many iterations/steps were ran?
  • Curious approximately how much the full train cost and how long with what resources?