attribution-patching

Official

Scale circuit analysis with gradient patching.

Authorndif-team
Version1.0.0
Installs0

System Documentation

What problem does it solve?

Gradient-based attribution patching provides a scalable alternative to exact activation patching by using gradients to approximate patch effects, enabling analysis across thousands of components without dozens of forward passes.

Core Features & Use Cases

  • Efficiently estimate per-component contributions by combining clean vs. corrupted activations with backward gradients.
  • Supports batch processing across multiple prompts and layers, enabling large-scale circuit analysis.
  • Useful for rapid hypothesis testing, instrumentation, and screening before targeted, exact patching.

Quick Start

Run attribution-patching on a model to generate layer-wise attributions for a given prompt.

Dependency Matrix

Required Modules

None required

Components

Standard package

💻 Claude Code Installation

Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.

Please help me install this Skill:
Name: attribution-patching
Download link: https://github.com/ndif-team/skills/archive/main.zip#attribution-patching

Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
View Source Repository

Agent Skills Search Helper

Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.