Estimating the difficulty of a dataset typically involves comparing
state-of-the-art models to humans; the bigger the performance gap, the harder
the dataset is said to be. However, this comparison provides little
understanding of how difficult each instance in a given distribution is,