What problem does it solve? Modifying factual knowledge inside large language models like GPT-2 XL or GPT-J normally requires expensive retraining, and understanding where a model stores a specific fact requires manual probing. This Skill provides working scripts for causal tracing and Rank-One Model Editing (ROME) so you can locate critical layers and rewrite specific factual associations directly. ## Core Features & Use Cases - Causal Tracing: Trace information flow through transformer layers to identify which layers are critical for a given factual association, with visualization of layer effects. - ROME Model Editing: Apply rank-one weight updates to MLP layers to change what a model believes about a subject (e.g., change "LeBron James plays basketball" to "football") without retraining. - API Reference: Includes a detailed reference for ROME hyperparameters, baseline methods (fine-tuning, Knowledge Neurons), and evaluation utilities. - Use Case: A researcher studying mechanistic interpretability wants to verify that factual recall localizes to specific MLP layers, then surgically edit one fact and measure efficacy, generalization, and specificity of the edit. ## Quick Start Run the ROME editing example script to load GPT-2 XL, apply a factual edit to a subject like LeBron James, and compare model completions before and after the edit.
