safety_circuit_info()

Runs refusal-direction extraction + safety-neuron localization in one call. Returns a CircuitInfo.

from safetune.interpret import safety_circuit_info

circuit = safety_circuit_info(
    model, tokenizer,
    harmful_prompts=harmful, harmless_prompts=harmless,
    method="weight",
    top_k_per_layer=16,
)

print(len(circuit.safety_units.unit_ids), "safety units found")
print(circuit.safety_units.unit_ids[:5])  # e.g. ["L0.mlp.down_proj.50", ...]

Parameter	Type	Default	Description
method	str	"weight"	Neuron scoring method
top_k_per_layer	int	16	Top-k neurons per layer
target_module	str	"mlp.down_proj"	Module to inspect
target_layers	Optional[List[int]]	None	Layers to score
activation_module	str	"mlp.act_fn"	Module for activation hooks
activation_score	str	"mean_abs_diff"	Activation score variant

When to use

Use safety_circuit_info() as the default entry point when you need a CircuitInfo for downstream use in Steer or Recover. It replaces two separate calls.