How to use Trainer when eval dataset has multiple references


I want to fine-tune mT5 on the TaTA dataset. For most examples, the dataset has multiple references, so the target column is a list of possible references. For the training set, I can just explode these into their own rows and treat them as separate examples. But the Trainer expects an eval_dataset, but it can’t handle examples with multiple references. Is there a way I can do this?