mi-experimenter

What problem does it solve?

This Skill automates complex mechanistic interpretability experiments, enabling researchers to understand how specific components of neural networks contribute to their behavior.

Core Features & Use Cases

• R_V Measurement: Quantify the representational quality of activations using the R_V metric.
• Causal Validation: Run controlled experiments to isolate the causal impact of specific model components (e.g., layers, attention heads) on behavior.
• Cross-Architecture Analysis: Compare R_V across different model families (GPT-2, Llama, Mistral) to find generalizable patterns.
• Use Case: Identify which MLP layers in a large language model are most critical for understanding factual recall by ablating them and measuring the resulting drop in R_V.

Quick Start

Use the mi-experimenter skill to run causal validation on the 'mistralai/Mistral-7B-v0.1' model targeting layer 27.