Perhaps this?
For LLM validation loss you might also need label_names=[]
in the TrainingArguments and set trainer.can_return_loss = True
to satisfy transformers/src/transformers/trainer.py at v4.40.2 · huggingface/transformers · GitHub for some models.