Protein LM interpretability

Protein LM Sparse Feature Circuits

Interactive causal graph of SAE features → behavior (e.g., TM helix detection).

Graph Min edge weight

0.00

Max edges per node

Details

Click a node or edge in the graph.

Nodes are SAE features (e.g. L5:F123) and one behavior node (behavior:tm).
Edges are causal-effect proxies:
- feat→behavior: Δloss when ablating that feature.
- feat→feat: mean |Δ downstream feature activation| after ablation.

Tip: Start with a higher Min edge weight to reveal the core circuit.

Made with Cytoscape.js · Static JSON graphs