What problem does it solve? Understanding which internal components of a language model drive specific predictions requires tedious manual instrumentation of attention heads, MLP layers, and residual streams. This Skill provides ready-to-run scripts for circuit discovery using RelP (Relevance Patching) and Layer-wise Relevance Propagation, so researchers can compute component relevance scores without writing boilerplate hooking code. ## Core Features & Use Cases - RelP-Based Circuit Analysis: Load HookedTransformer models with configurable LRP rules (LN-rule, Identity-rule, Half-rule, 0-rule, AH-rule) and compute relevance scores for attention, MLP, and residual stream components. - IOI Task Benchmarking: Run Indirect Object Identification analysis to identify name mover heads, rank attention heads by indirect-object attention, and compare RelP against standard attribution patching via logit differences. - LRP Rule Comparison: Evaluate multiple LRP rule configurations on the same input to see how propagation choices change component relevance rankings. - Use Case: A mechanistic interpretability researcher wants to know which attention heads in GPT-2 small copy name information in IOI sentences. They run the IOI analysis script to get ranked head scores and a RelP-vs-attribution comparison in one pass. ## Quick Start Ask the agent to run the basic RelP analysis script on GPT-2 small to compute relevance scores for attention and MLP components on a sample sentence.
