This was happening to me as well in a tabular model a few days ago. Different groups of people with different random splittings in the same dataset were consistently seeing a valid_loss lower than the train_loss. It’s hard for me to find an intuition of why that is happening, aside from a bias in the splits.