Leave-1%-out retrained models, per-query loss changes, and datasets for computing LDS.