Leave-1%-out retrained models, per-query loss changes, and datasets for computing LDS.
Note Pre-tokenized WikiText-2 chunks for the Bae et al. 2024 SOURCE replication