in that case you'd need to do actual training/finetuning with a dataset that has information about things that were left out of the original training data.