Five State-of-the-Art Methods
Choose the best unlearning method for your use case. Each method offers unique advantages for different scenarios.
Gradient Ascent
The simplest and fastest method. Directly maximizes loss on forget data through gradient ascent.
- Fast training
- No reference model
- Perfect for quick experiments
Gradient Difference
Balances forgetting with retention by combining gradient ascent and descent on different data sets.
- Stable optimization
- Explicit retain preservation
- Configurable loss weights
Direct Preference Optimization (DPO)
Preference-based method that steers the model toward alternate answers while avoiding original ones.
- Principled approach
- Preference learning
- Stable convergence
Representation Misdirection (RMU)
Steers internal representations toward random directions for hazardous knowledge removal.
- Safety-critical unlearning
- Layer-specific control
- Adaptive steering
Simple NPO (SimNPO)
Margin-based loss approach that reduces likelihood of forget examples without reference models.
- No reference model
- Margin-based optimization
- Memory efficient
More Coming Soon
We're constantly researching and implementing new unlearning methods. Stay tuned for updates!
- Active research
- Community driven
- Open to contributions
5+
Evaluation Metrics
100%
Open Source
HF
Compatible
Pricing
Open Source, Free Forever
Unlearun is completely open source under MIT license. Use it freely for research or production.
Open Source
Free
For researchers and individual developers
- All 5 unlearning methods
- Complete evaluation metrics
- HuggingFace integration
- MIT License
- Community support
- Full documentation
Most Popular
Research
Free
For academic researchers and institutions
- Everything in Open Source
- Citation examples
- Research paper references
- Benchmark datasets
- Academic support
- Reproducible configs
Enterprise
Custom
For organizations with production needs
- Everything in Research
- Priority support
- Custom integrations
- Training & consulting
- SLA guarantees
- Dedicated support channel
Questions about enterprise deployment?
View on GitHub →Trusted by researchers worldwide
From academia to industry, Unlearun is helping teams implement state-of-the-art machine unlearning.
Unlearun made implementing GDPR compliance in our LLM straightforward. The RMU method worked perfectly for removing sensitive user data.
Dr. Sarah Chen
@sarahresearch
Five SOTA methods in one library with excellent documentation. Saved weeks of implementation time for our research.
Alex Kumar
@alexmleng
The evaluation metrics are comprehensive and well-thought-out. Perfect for our academic research on machine unlearning.
Prof. Maria Rodriguez
@mariarodriguez
Unlearun made implementing GDPR compliance in our LLM straightforward. The RMU method worked perfectly for removing sensitive user data.
Dr. Sarah Chen
@sarahresearch
Five SOTA methods in one library with excellent documentation. Saved weeks of implementation time for our research.
Alex Kumar
@alexmleng
The evaluation metrics are comprehensive and well-thought-out. Perfect for our academic research on machine unlearning.
Prof. Maria Rodriguez
@mariarodriguez
Unlearun made implementing GDPR compliance in our LLM straightforward. The RMU method worked perfectly for removing sensitive user data.
Dr. Sarah Chen
@sarahresearch
Five SOTA methods in one library with excellent documentation. Saved weeks of implementation time for our research.
Alex Kumar
@alexmleng
The evaluation metrics are comprehensive and well-thought-out. Perfect for our academic research on machine unlearning.
Prof. Maria Rodriguez
@mariarodriguez
Unlearun made implementing GDPR compliance in our LLM straightforward. The RMU method worked perfectly for removing sensitive user data.
Dr. Sarah Chen
@sarahresearch
Five SOTA methods in one library with excellent documentation. Saved weeks of implementation time for our research.
Alex Kumar
@alexmleng
The evaluation metrics are comprehensive and well-thought-out. Perfect for our academic research on machine unlearning.
Prof. Maria Rodriguez
@mariarodriguez
HuggingFace integration makes it seamless to use with our existing pipeline. Production-ready out of the box.
David Park
@davidaidev
Used GradDiff to remove copyrighted content from our model. The balance between forgetting and retention is impressive.
Emily Watson
@emilywatson
The paper implementations are faithful to the originals. Great for reproducing research results.
Dr. James Mitchell
@jamesmitchell
HuggingFace integration makes it seamless to use with our existing pipeline. Production-ready out of the box.
David Park
@davidaidev
Used GradDiff to remove copyrighted content from our model. The balance between forgetting and retention is impressive.
Emily Watson
@emilywatson
The paper implementations are faithful to the originals. Great for reproducing research results.
Dr. James Mitchell
@jamesmitchell
HuggingFace integration makes it seamless to use with our existing pipeline. Production-ready out of the box.
David Park
@davidaidev
Used GradDiff to remove copyrighted content from our model. The balance between forgetting and retention is impressive.
Emily Watson
@emilywatson
The paper implementations are faithful to the originals. Great for reproducing research results.
Dr. James Mitchell
@jamesmitchell
HuggingFace integration makes it seamless to use with our existing pipeline. Production-ready out of the box.
David Park
@davidaidev
Used GradDiff to remove copyrighted content from our model. The balance between forgetting and retention is impressive.
Emily Watson
@emilywatson
The paper implementations are faithful to the originals. Great for reproducing research results.
Dr. James Mitchell
@jamesmitchell
DPO unlearning with alternate answers worked beautifully for our content moderation use case.
Lisa Zhang
@lisazhang
Open source, MIT licensed, and actively maintained. This is how ML libraries should be built.
Ryan O'Connor
@ryanoconnor
The membership inference attack evaluation helped us validate our unlearning effectiveness. Excellent tool for privacy research.
Dr. Aisha Patel
@aishapatel
DPO unlearning with alternate answers worked beautifully for our content moderation use case.
Lisa Zhang
@lisazhang
Open source, MIT licensed, and actively maintained. This is how ML libraries should be built.
Ryan O'Connor
@ryanoconnor
The membership inference attack evaluation helped us validate our unlearning effectiveness. Excellent tool for privacy research.
Dr. Aisha Patel
@aishapatel
DPO unlearning with alternate answers worked beautifully for our content moderation use case.
Lisa Zhang
@lisazhang
Open source, MIT licensed, and actively maintained. This is how ML libraries should be built.
Ryan O'Connor
@ryanoconnor
The membership inference attack evaluation helped us validate our unlearning effectiveness. Excellent tool for privacy research.
Dr. Aisha Patel
@aishapatel
DPO unlearning with alternate answers worked beautifully for our content moderation use case.
Lisa Zhang
@lisazhang
Open source, MIT licensed, and actively maintained. This is how ML libraries should be built.
Ryan O'Connor
@ryanoconnor
The membership inference attack evaluation helped us validate our unlearning effectiveness. Excellent tool for privacy research.
Dr. Aisha Patel
@aishapatel
Open Source. Production Ready. Research Grade.
Start unlearning in minutes with pip install.
$pip install unlearun
unlearun
Stay Updated
Get notified about new features, methods, and updates to the unlearun library.
We respect your privacy. Unsubscribe at any time.
✶Faqs