Pyvene Interventions

This skill helps you run causal interventions on PyTorch models using Stanford NLP's pyvene library. If you're doing activation patching, causal tracing (ROME-style localization), or interchange intervention training, it walks you through the declarative dict-based API. You get concrete workflows for finding where factual associations live in transformers, testing which attention heads matter for specific behaviors, and training interventions to discover causal structure. It's focused on mechanistic interpretability experiments where you need to patch activations between model runs and share reproducible intervention configs. The guidance assumes you already know why you want to intervene on model internals rather than just probe them.

Install to Claude Code

npx -y skills add orchestra-research/ai-research-skills --skill pyvene-interventions --agent claude-code