User Guide

Welcome to the mi-crow user guide! This comprehensive guide will help you understand and use the mi-crow library for mechanistic interpretability research.

What is mi-crow?

mi-crow is a Python package for explaining and steering LLM behavior using Sparse Autoencoders (SAE) and concepts. It provides a complete toolkit for:

• Activation Analysis: Save and analyze model activations from any layer
• SAE Training: Train sparse autoencoders to discover interpretable features
• Concept Discovery: Identify and name concepts learned by SAE neurons
• Model Steering: Manipulate model behavior through concept-based interventions
• Hook System: Flexible system for intercepting and modifying activations

What is Mechanistic Interpretability?

Mechanistic interpretability is the study of understanding how neural networks work by reverse-engineering their internal computations. In the context of language models, this means:

• Understanding what features the model learns at different layers
• Identifying how these features combine to produce outputs
• Discovering interpretable concepts that correspond to human-understandable ideas
• Using this understanding to control and improve model behavior

Library Capabilities

mi-crow provides a modular architecture for mechanistic interpretability research:

• Language Model Wrapper: Easy loading and inference with HuggingFace models
• Sparse Autoencoders: Train and use SAEs to discover interpretable features
• Hooks System: Powerful framework for observing and modifying activations
• Store: Hierarchical storage for activations, models, and metadata
• Datasets: Flexible data loading from HuggingFace or local files

Getting Started

1. Installation - Set up mi-crow and its dependencies
2. Quick Start - Run your first example in minutes
3. Core Concepts - Understand the fundamental ideas
4. Hooks System - Learn about the powerful hooks framework
5. Workflows - Step-by-step guides for common tasks

Documentation Structure

Core Documentation

• Installation & Setup - Installation and environment configuration
• Quick Start - Get up and running quickly
• Core Concepts - Fundamental concepts and architecture

Hooks System

The hooks system is the foundation of mi-crow's interpretability capabilities:

• Hooks Overview - Introduction to the hooks system
• Hooks Fundamentals - Core concepts and lifecycle
• Detector Hooks - Observing activations without modification
• Controller Hooks - Modifying activations during inference
• Hook Registration - Managing hooks on layers
• Advanced Hooks - Advanced patterns and best practices

Workflows

Step-by-step guides for common tasks:

• Workflows Overview - When to use each workflow
• Saving Activations - Collect activation data
• Training SAE Models - Train sparse autoencoders
• Concept Discovery - Find interpretable concepts
• Concept Manipulation - Control model behavior
• Activation Control - Direct activation manipulation

Additional Resources

• Best Practices - Tips for effective research
• Troubleshooting - Common issues and solutions
• Examples - Example notebooks and learning path
• Experiments - Detailed experiment walkthroughs

Next Steps

If you're new to mi-crow, we recommend following this path:

1. Start with Installation to set up your environment
2. Run through the Quick Start tutorial
3. Read Core Concepts to understand the fundamentals
4. Explore the Hooks System - it's central to everything
5. Try a Workflow that matches your research goals
6. Check out Examples for more detailed code

For API reference, see the API Documentation.