Model Merging for Knowledge Edit

This repository contains modified versions of EasyEdit and mergekit for experiments on model merging after knowledge editing.

Quick Start

Installation

git clone https://github.com/Applied-Machine-Learning-Lab/MM4KE.git
cd MM4KE
conda create -n MM4KE python=3.9.7
conda activate MM4KE

# Install EasyEdit dependencies
cd EasyEdit
pip install -r requirements.txt

# Install mergekit dependencies
cd ../mergekit
pip install -e .

Knowledge Edit

cd EasyEdit/examples
python run_knowedit_llama2.py \
    --editing_method=FT \
    --hparams_dir=../hparams/FT/qwen2.5-7b \
    --data_dir=./data \
    --datatype='counterfact'

Model Merging

cd ../../mergekit
mergekit-pytorch examples/task_arithmetic.yml ./output_pytorch_model_directory

Changes

Changes of EasyEdit

1. Sequential Editing Support (EasyEdit/examples/run_knowedit_llama2.py#L265)

   • Added sequential_edit=True parameter to enable sequential knowledge editing

2. Extended Training Epochs (EasyEdit/easyeditor/editors/editor.py#L375-376)

   • Added epoch loop with epoch = 5 for improved editing performance
   • Allows multiple training iterations over the edit data

3. Qwen2.5 Model Support (EasyEdit/hparams/FT/qwen2.5-7b.yaml)

   • Added configuration file for Qwen2.5-7B model
   • Enables knowledge editing on the latest Qwen model series

Changes of mergekit

1. Sparsification for Task Arithmetic (mergekit/mergekit/merge_methods/registry.py#L31)
   • Added sparsification_method=SparsificationMethod.magnitude to task arithmetic merging
   • Enables sparsification similar to TIES method but without TIES regularization
   • Allows for more controlled model merging with magnitude-based pruning