mechinterp-cluster-mapper

What problem does it solve?

This skill helps researchers and engineers understand how SAE features relate to each other by identifying subsystems, co-activation patterns, and shared token drivers.

Core Features & Use Cases

• Co-activation analysis: Quantify how features activate together across contexts.
• Subsystem discovery: Group related features into coherent subsystems and reveal redundancy.
• Driver identification: Find tokens or patterns that drive multiple features.
• Use Case: When you have a cluster of SAE features and want to understand their interdependencies to guide experiments or model interpretation.

Quick Start

Load an Ultra context and run a cluster analysis on a set of feature ids, then inspect the resulting subclusters and shared drivers.