activation-patching

What problem does it solve?

Activation patching provides a causal intervention framework to identify which model components (layers, heads, or positions) drive a specific behavior by swapping activations and observing changes in outputs.

Core Features & Use Cases

• Causal intervention across transformer components (layers, attention heads, token positions) to locate components critical for a behavior.
• Three-run patching workflow: run clean, run corrupted, patch activations and measure impact to quantify causality.
• Interpretability outputs and guidance for diagnosing circuits or computational stages in neural networks.

Quick Start

Run a clean prompt and a corrupted prompt, patch activations layer-by-layer, and measure the effect on outputs.