🧠 LLM Unlearning

Remove specific information from a model without retraining from scratch. 30-40 min vs 14-16 hours

⚙️ How It Works

1
Load Model Base model or forged model
→
2
Apply LoRA Freeze layers + add adapter
→
3
Train on Forget Set 2 epochs, optimized batch
→
4
Evaluate Metrics Retain & Forget accuracy
Technique
PEFT + LoRA (r=16)
Training Time
30-40 minutes
GDPR Ready
Right to be Forgotten
🎭 Interactive Demo: Before & After Try It!
💬 Ask the AI
🤖
Hello! I'm an AI that knows about Alice, Bob, and Charlie. Ask me anything about their jobs, hobbies, locations, or education!
🧠 Make AI Forget
👩
Alice
4 facts known
👨
Bob
4 facts known
👨‍🦱
Charlie
4 facts known
💡
How to use: First, ask the AI about someone (left panel) to see what it knows. Then, click "Forget" on the right panel to make the AI forget that information. Ask the same question again - the AI will no longer know the answer! This demonstrates the before and after of LLM unlearning.
Unlearning Workbench
Required
📤
Click to upload or drag & drop
CSV or JSON

The original dataset the model was trained on.

Required
📤
Click to upload or drag & drop
CSV or JSON

Dataset without the information to be forgotten.

Describe the specific information that should be removed from the model.