The input is a namedtuple of type EvalPrediction. It should have one key predictions for the predictions (will have the structure as the output of your model, so one tensor if your model outputted one tensor, a tuple of two tensors if that’s what your model returns, et.) and one key label_ids that will contain all the labels.
The Trainer does update the weights of the model, otherwise, the model would not be… well… training 