They said the training set distribution.

It's frankly impossible that they have a training set without any biases, though I'm sure they worked to eliminate the ones they could think of (which itself would have bias).