mechinterp-cluster-mapper

SAE feature co-activation analysis quantifies how features activate together across model contexts. You need it when you have a cluster of SAE features and want to identify subsystems, measure interdependencies, and guide model interpretation experiments.

To identify subsystems and redundancy within SAE features, you run a cluster analysis on a set of feature ids. This groups related features into coherent subclusters based on their cross-feature activations, revealing structural interdependencies.

To find tokens that drive multiple SAE features, you analyze cross-feature activations to identify shared token drivers. This process highlights specific tokens or patterns responsible for triggering groups of related features within a subsystem.

No specific external dependencies are required to analyze SAE feature clusters. You simply load an Ultra context and run a cluster analysis on your target feature ids to inspect the resulting subclusters and shared drivers.

The best way to quantify relationships between SAE features is by measuring co-activation patterns across contexts. This approach provides structured reports of subsystem drivers, cluster formation into subgroups, and feature redundancy.