Beyond Superficial Forgetting: Thorough Unlearning through Knowledge Density Estimation and Block Re-insertion official code

KUnBR

My paper: Beyond Superficial Forgetting: Thorough Unlearning through Knowledge Density Estimation and Block Re-insertion

Code detail

Because our approach is based on existing methods and is a framework-based approach, we will focus on providing our framework in the code. For the original method, please refer to its original paper.

For a key example of unlearning and the data, please refer to our fellow's paper: Do Unlearning Methods Remove Information from Language Model Weights?

The code can be run using the unlearn base class from this paper.Alternatively, if a well-implemented GD method exists, it can be plugged into our code framework for pluggable operation.

After modifying the unlearning code in Biru, you can run it directly in the workflow. The versions of the relevant packages are already listed in the requirements file.