Compute total, direct, and indirect causal effects in transformer models.
npx skills add https://github.com/ndif-team/skills --skill causal-tracing
Please help me install this Agent Skill. Skill: causal-tracing Source: https://github.com/ndif-team/skills/tree/main/plugins/nnsight/skills/causal-tracing Command: npx skills add https://github.com/ndif-team/skills --skill causal-tracing
Causal tracing identifies which intermediate computations causally mediate the relationship between inputs and outputs, revealing not just what correlates with behavior but what causes it.
Run a tracing session on a language model with base and source prompts to compute total, direct, and indirect causal effects.
High-intent search queries and answers about installing and using this skill.
Causal tracing identifies which intermediate computations mediate the relationship between inputs and outputs by quantifying total, direct, and indirect effects across specific layers and tokens in transformers.
Yes, interchange interventions test causal relationships across runs by applying interventions to trace how different base and source prompts influence intermediate activations and final predictions.
You need a LanguageModel instance, a tracing context, and access to model activations and outputs to compute total, direct, and indirect causal effects during a tracing session.
Causal tracing debugs model behavior by diagnosing how different tokens and network layers contribute to final predictions, validating hypotheses about information flow and guiding interventions to alter outputs.
Direct effects measure the immediate causal influence of a component on outputs, while indirect effects capture the causal influence mediated through other intermediate components in the transformer network.
Use causal tracing when you need to move beyond correlational interpretability to identify what actually causes model behavior, validating information flow hypotheses through total, direct, and indirect effect quantification.
qcmuuTrace, save, and patch PyTorch model activations with optional NDIF remote execution.
danielsimonjrBuild causal DAGs and document mechanisms for counterfactual reasoning.
jonnabioInspect transformer internals via HookPoints and cached activations.
jonnabioPerforms customizable causal interventions on PyTorch neural network models.
