Workflows¶
This section provides step-by-step guides for common tasks in mi-crow. Each workflow is designed to be self-contained and complete.
Available Workflows¶
Saving Activations¶
Learn how to collect and save activations from model layers for analysis and SAE training.
When to use: - Preparing data for SAE training - Analyzing activation patterns - Debugging model behavior - Collecting datasets for research
Training SAE Models¶
Complete guide to training sparse autoencoders on saved activations.
When to use: - Training your first SAE - Understanding hyperparameter selection - Monitoring training progress - Saving and loading trained models
Concept Discovery¶
Discover interpretable concepts by analyzing SAE neuron activations.
When to use: - Finding what each SAE neuron represents - Collecting examples for manual curation - Validating concept quality - Understanding model features
Concept Manipulation¶
Control model behavior by manipulating discovered concepts.
When to use: - Steering model outputs - Running intervention experiments - A/B testing concept effects - Real-time model control
Activation Control¶
Directly manipulate activations using hooks for fine-grained control.
When to use: - Custom intervention experiments - Fine-grained activation modification - Multi-layer interventions - Advanced control patterns
Workflow Dependencies¶
Most workflows build on each other:
Saving Activations
↓
Training SAE Models
↓
Concept Discovery
↓
Concept Manipulation
Activation Control can be used independently or in combination with SAE-based workflows.
Quick Reference¶
| Workflow | Input | Output | Time |
|---|---|---|---|
| Saving Activations | Model + Dataset | Saved activations | Minutes |
| Training SAE | Saved activations | Trained SAE | Hours |
| Concept Discovery | Trained SAE + Dataset | Top texts per neuron | Minutes |
| Concept Manipulation | Trained SAE + Concepts | Modified outputs | Seconds |
| Activation Control | Model + Hooks | Modified outputs | Seconds |
Getting Started¶
If you're new to mi-crow, we recommend following workflows in order:
- Start with Saving Activations to understand data collection
- Move to Training SAE Models to learn feature discovery
- Try Concept Discovery to find interpretable features
- Use Concept Manipulation to control model behavior
For advanced users, Activation Control provides direct hook-based control.
Integration with Examples¶
Each workflow references relevant example notebooks:
- Examples are in the
examples/directory - Workflows explain the concepts
- Examples provide runnable code
- See Examples Guide for the full list
Next Steps¶
Choose a workflow that matches your goal:
- New to mi-crow? → Start with Saving Activations
- Want to train SAEs? → See Training SAE Models
- Ready to discover concepts? → Try Concept Discovery
- Need model control? → Use Concept Manipulation or Activation Control