pyvene — причинные интервенции в нейросетях

★ 7.7 · ai-tooling

pyvene is a Claude Code skill that guides causal interventions on PyTorch models using the declarative, dict-based framework from Stanford NLP's pyvene library (840+ GitHub stars, NAACL 2024). It covers ROME-style causal tracing, activation patching, interchange intervention training, and systematic testing of causal hypotheses about model components. The skill documents intervention types — VanillaIntervention, RotatedSpaceIntervention, ZeroIntervention, CollectIntervention, and more — alongside over a dozen targetable components such as block_output, mlp_activation, and attention_value_output. Dependencies are pyvene>=0.1.8, torch>=2.0.0, and transformers>=4.30.0. Aimed at mechanistic interpretability researchers who need reproducible, shareable intervention experiments across any PyTorch architecture, not just transformers.