Erasus Documentation

Erasus is a universal machine unlearning framework supporting 15+ model architectures across vision, language, diffusion, audio, and video modalities.

Getting Started

• Quick Start
  • Installation
  • Basic Usage
  • Modality-Specific Unlearners
  • CLI Usage
• Installation
  • Requirements
  • From PyPI
  • From Source
  • GPU Support
  • Development Installation
  • Docker
  • Verifying Installation

User Guide

• Conceptual Guide — How Erasus Thinks About Unlearning
  • The Organising Principle: Select → Unlearn → Verify
  • Erasus’s Opinions
  • When to Use Which Strategy
  • When to Use Which Selector
  • The Trainer / Module Split
  • Coresets as First-Class Objects
  • The UnlearningDataset Abstraction
  • Standardised Benchmarking
  • Putting It All Together
• Architecture Overview
  • Design Philosophy
  • Core Components
  • Registry System
  • Data Flow
  • Module Organisation
• Unlearning Pipeline
  • Step 1: Prepare Data
  • Step 2: Choose a Strategy
  • Step 3: Configure Unlearner
  • Step 4: Run Unlearning
  • Step 5: Evaluate
  • Step 6: Visualise (Optional)
  • Step 7: Certify (Optional)
• Strategy Selection Guide
  • Strategy Categories
  • Decision Matrix
  • Combining Strategies
• Selector Guide
  • When to Use Selectors
  • Selector Categories
  • Quick Comparison
  • Usage
• Metrics Guide
  • Forgetting Quality
  • Model Utility
  • Privacy
  • Efficiency
  • Using MetricSuite
  • Benchmark Runner
• Visualization Guide
  • Loss Curves
  • Feature Space Plots
  • MIA Analysis
  • Attention Maps
  • Gradient Analysis
  • Comparison Plots
  • Influence Maps
  • Interactive Reports
• Hugging Face Hub Integration
  • Usage
  • Deployment checklist

API Reference

• Core API
• Unlearners
• Strategies API
  • Base Strategy
  • Gradient Methods
  • Parameter Methods
  • Data Methods
  • VLM-Specific
  • LLM-Specific
  • Diffusion-Specific
  • Ensemble
  • Strategy Registry
• Selectors API
  • Base Selector
  • Random Selection
  • Influence-Based
  • Geometry-Based
  • Learning-Based
  • Gradient-Based
  • Ensemble Selectors
  • Quality Metrics
  • Selector Registry
• Metrics API
  • Metric Suite
  • Forgetting Metrics
  • Utility Metrics
  • Privacy Metrics
  • Efficiency Metrics
  • Benchmarks
  • Metric Registry
• Data API
  • Benchmark Datasets
  • Data Utilities
  • Augmentation
  • Synthetic Data
• Visualization API
  • Embedding Plots
  • Loss Curves
  • Feature Plots
  • MIA Plots
  • Gradient Analysis
  • Loss Surfaces
  • Attention Maps
  • Comparison Plots
  • Activation Analysis
  • Influence Maps
  • Cross-Modal
  • Reports
  • Interactive
• Certification API
  • Certified Removal
  • Verification
  • Bounds
  • Usage Example
• Privacy API
  • Privacy Accountant
  • DP Mechanisms
  • Gradient Clipping
  • Secure Aggregation
  • Certificates
  • Influence Bounds
  • Usage Example
• Utils API
  • Helpers
  • Distributed Training
  • Checkpointing
  • Callbacks
  • Early Stopping
  • Profiling
  • Reproducibility
  • Logging
  • Seed

Examples & Benchmarks

• CLIP Basic Example
• LLaMA Concept Removal
• Stable Diffusion NSFW Removal
• TOFU Benchmark

Development

• Contributing to Erasus
• Changelog

Indices and tables

• Index

• Module Index

• Search Page