Find what drives your neural network's behavior.
Guides experiments that test how PyTorch model components affect behavior using pyvene interventions.
When to use it
Use for causal tracing, activation patching, interchange intervention training, or testing causal hypotheses about model components.
Give it a causal-intervention question about a model; it returns guidance and code patterns for implementing the experiment.
The intervention workflows use pyvene; the declared minimum version is 0.1.8.
The declared dependencies require torch version 2.0.0 or newer.
This skill
Hugging Face
Loads a pretrained steering intervention
pyvene
Installs pyvene