ROME

This package provides a self-contained implementation of Rank-One Model Editing (ROME).

Recall that ROME's update consists of: $u$ selection, $v_*$ optimization, and $v$ insertion.

• compute_u.py: Chooses a $u$ vector.
• compute_v.py: Choose a $v_*$ via optimization, then computes $v$.
• rome_main.py: Instruments main logic.
• rome_params.py: Interface for specifying hyperparameters. Inherits from the base params.py module.

For estimating second moment statistics of keys ($C = KK$), we provide the layer_stats module. See the main README for usage instructions.

• layer_stats.py: Logic for retrieving and caching key statistics.
• tok_dataset.py: Utilities for creating a dataset of tokens.