We also found that difficulty depends on the model. The tasks our 6 models each solve half the time barely overlap, and not one task is half-solved by all six.

So you can't borrow anyone's difficulty labels, ours included. Grade against the model you're about to train.

Same