pyvene-interventions

Causal activation patching restores clean activations into corrupted model runs to localize causal components. You configure IntervenableConfig and RepresentationConfig to specify target layers and intervention types for testing whether specific internal activations drive model behavior.

Causal tracing is an interpretability technique that compares corrupted versus clean prompts to localize causal components in neural networks. By running a VanillaIntervention sweep over layers, you can precisely test causal hypotheses about which internal activations drive a model's specific behaviors.

Yes, pyvene interventions support both transformer and non-transformer PyTorch architectures. The declarative intervention framework allows you to configure target components and intervention types flexibly across different model designs for causal hypothesis testing and model steering.

Trainable DAS enables interchange intervention training by applying low-rank rotated interventions to discover causal subspaces within model activations. You configure these trainable interventions via IntervenableConfig to steer model behavior and identify specific causal representations driving outputs.

Yes, causal interventions require installing pyvene version 0.1.8 or higher, PyTorch version 2.0.0 or higher, and HuggingFace transformers version 4.30.0 or higher. These compatible versions ensure the declarative intervention framework functions correctly across your interpretability workflows.

Model steering interventions can be saved and loaded for reproducibility by applying them during generation and sharing the configured intervention states. This allows you to consistently reproduce causal intervention experiments and steer model behavior across different runs and environments.