What problem does it solve? Manually identifying which attention heads and MLP components inside a transformer are responsible for a specific behavior is slow and error-prone. This Skill automates circuit discovery by running the ACDC (Automatic Circuit Discovery) pipeline, which prunes a model's computational graph down to the edges that actually matter for a given task. ## Core Features & Use Cases - Automated Circuit Discovery: Runs the ACDC pipeline to iteratively remove edges from a transformer's computational graph and isolate the minimal circuit responsible for a behavior. - Computational Graph Manipulation: Provides API references for building, editing, and visualizing computational graphs with nodes (attention heads, MLPs) and typed edges. - Experiment Management: Supports loading models, constructing graphs from model hooks, running experiments, and saving or reloading results, with optional Weights & Biases logging. - Use Case: A mechanistic interpretability researcher studying indirect object identification (IOI) can run the demo script to launch the ACDC pipeline, then visualize the recovered circuit with Graphviz to verify which heads drive the behavior. ## Quick Start Run the ACDC demo script to launch the default automated circuit discovery pipeline on a transformer model.
