pyvene-interventions

Activation patching identifies causal relationships by swapping or patching activations between model runs. You configure a target intervention with IntervenableConfig for a specific layer and component, then run IntervenableModel against base and source inputs to observe the causal effects on outputs.

Causal tracing is an interpretability technique that identifies how specific layers and components drive model outputs. By applying targeted pyvene interventions to patch activations across runs, you can trace which internal representations are causally responsible for specific model behaviors.

Yes, pyvene interventions support PyTorch architectures including GPT-2, LLaMA, and Pythia. The Skill applies IntervenableModel and RepresentationConfig across these compatible models to ensure reproducibility and shared learnings for causal experiments.

Interchange intervention training (IIT) discovers trainable causal subspaces by applying structured interventions to internal model representations. Using pyvene, you configure IntervenableConfig to swap activations between inputs, enabling the model to learn aligned causal mechanisms during training.

Yes, causal interventions require pyvene>=0.1.8, torch>=2.0.0, and transformers>=4.30.0. These dependencies provide the foundational IntervenableModel, IntervenableConfig, and RepresentationConfig classes needed to configure and execute activation patching and steering experiments.

Model steering applies generation-time interventions by modifying activations during inference to alter model outputs. Using pyvene, you configure interventions that add or swap activations on the fly, and these interventions can be easily shared via HuggingFace for reproducible steering experiments.