One thing I don't understand is that the dataset already comes with a "train dataset" (120 recordings) and "test dataset" ("Recordings of 22 repetitions of 7 narrative test set videos"). Why are they not just setting aside a randomly selected subset of the training data for testing?

I'm unclear on whether the testing data contains some of the same video clips that were shown in the training data, repeated literally? It seems it would be best to test the model on a previously unseen image, to ensure that it really learnt the mapping and is not just "memorizing" the responses.