I’ve recently finetuned models using
davinci-002 via the API, using the legacy data format. I then triggered another finetuning job with the same format and same data, but the training loss and validation loss metrics shown in the Fine-tuning UI are different.
Is finetuning deterministic? If given the same data, the same base model and the same hyperparameters, should the metrics and hence the new finetuned models be identical? Is there a way to ensure that they’re identical?
Thanks in advance, apologies if this has been covered somewhere that I’ve missed.