Core ML Concepts · advanced · concept 47 of 311 · Updated Nov 4, 2026

🧗Data Hardness and Example Difficulty

Data hardness is how difficult an example, or a whole dataset, is for a model to learn. It is scored from training dynamics: how often an example is forgotten, how confident and variable its predictions are across epochs (Data Maps), or how large its early error is (EL2N, GraNd). Hard examples carry most of the signal, yet the hardest are often mislabeled. Sorscher et al. showed that pruning easy examples can beat power-law scaling.