pyvene-interventions

Causal tracing identifies specific model components responsible for factual recall by restoring clean activations at targeted layers and positions during a forward pass. This Skill supports ROME-style localization to pinpoint exactly where factual information is retrieved.

Activation patching tests hypotheses by swapping activations between different model runs. You can debug hallucinations by pinpointing and intervening on the specific activations that lead to incorrect outputs, comparing clean and corrupted model runs.

Causal interventions include Vanilla, Addition, Zero, and trainable RotatedSpace intervention types. These are executed via a declarative framework supporting activation patching, causal tracing, and interchange intervention training across diverse architectures.

Causal interventions are supported across diverse model architectures including transformers, vision, and state-space models. The declarative framework adapts to various PyTorch model structures, enabling activation patching and causal tracing regardless of architecture.

Trainable interventions discover and learn causal subspaces using methods like DAS (Distributed Alignment Search). This allows you to identify rotated representational spaces where specific causal mechanisms operate within the neural network.

Activation patching swaps activations between different model runs to test specific hypotheses, while causal tracing localizes responsible components by restoring clean activations. Both are causal intervention techniques for understanding model decision-making processes.