Selective Pruning Experiments


Experiments based on paper "Dissecting Language Models: Machine Unlearning via Selective Pruning" https://arxiv.org/pdf/2403.01267.

Install repository and packages with poetry:

git clone https://github.com/nickypro/selective-pruning
curl -sSL https://install.python-poetry.org | python -
poetry install

The main packages required are:

  • taker
  • torchvision

Taker is a library for working with HuggingFace transformers.

Run experiment with bash:

to run experiments, simply:

cd exp1-is-effective
bash prune-llms.sh

See a simplified of the procedure in the example: exp4-comparison/sp.py

Credits

Most files for expriment 4 were taker from https://github.com/if-loops/selective-synaptic-dampening