2 ms·

And most of the bias probably isn't due to the training data anyway. A good amount of work has gone into censoring the model.