A dataset is not in the appropriate learning difficulty range if a model is either too incapable or too capable on it.

“And if it's too incapable or too capable, then it's not in the Goldilocks zone for the model to, to learn from it. Similarly, we would, um, do ablations where we'll take a set of data, train the model a little bit on it and see how things change.”