Interpret — diagnose where safety lives

Locate safety-relevant structure (directions, neurons, circuits) inside a model. Interpret changes nothing on its own; its artifacts are inputs to Steer and localization-aware Recover methods.

Input contract

flowchart LR
    M["Model + contrast prompts"] --> INTERP["Interpret<br/>safety neurons · circuit info · EAP"]
    INTERP --> CI["CircuitInfo"]
    CI --> STEER["Steer<br/>refusal direction · CAA vectors<br/>inference-time control"]
    CI --> RECOVER["Recover<br/>circuit-guided weight edit<br/>targeting masks"]

Quick example

from safetune.interpret import safety_circuit_info

circuit = safety_circuit_info(
    model, tokenizer,
    harmful_prompts=harmful, harmless_prompts=harmless,
)

Catalog of alternatives

Method	What it finds	Guide
identify_safety_neurons (weight mode)	safety-relevant weight columns	Weight-based
identify_safety_neurons (activation mode)	safety neurons by activation contrast	Activation-based
safety_circuit_info	convenience wrapper — both steps in one call	safety_circuit_info
CircuitInfo	round-trippable data container	CircuitInfo object
eap_safety_circuit	edge attribution patching	EAP / EAP-IG