pyvene-interventions

Activation patching in a PyTorch transformer is performed by swapping or restoring selected activations across specific layers and positions to localize where behavior is represented. This Skill configures the target layer, component, and intervention type using pyvene's declarative framework.

Causal tracing is a model interpretability technique that localizes where specific behavior is represented inside a PyTorch model by restoring selected activations across layers and positions. It uses pyvene's declarative framework to swap, modify, collect, or zero internal activations reproducibly.

Yes, you can train low-rank or rotated subspaces for interchange intervention training, distributed alignment search, and model steering. This Skill supports optional trainable intervention parameters to learn these subspaces within PyTorch architectures.

Yes, this Skill applies the same declarative causal intervention workflow to transformers and other PyTorch neural architectures. You can configure intervention types, source-base activation mapping, and target components across diverse model structures for interpretability research.

Source-base activation mapping requires precise layer and component configuration to define where interventions occur in your PyTorch model. This Skill uses pyvene's declarative framework to map source activations to base activations for reproducible causal analysis.

Model steering on PyTorch transformers is executed by applying trainable interventions that learn low-rank or rotated subspaces to modify model behavior. This Skill provides the declarative configuration needed to define intervention parameters and target layers for steering.