RLEdit

Official repo for our paper Reinforced Lifelong Editing for Language Models.

Initial Setup

Create a virtual environment and install the dependencies.

conda create -n rledit python==3.11
conda activate rledit
pip install -r requirements.txt

Download LLMs, and put their paths into config/model/xxx.yaml

Quick Start

You can modify the GPU device number you want to use in the model.py file.

run.sh is a simple example that performs a 400x20 editing task on the ZSRE dataset using Llama-3-8B.

sh run.sh

Configuration

The run.sh script is as follows:

python main.py dataset=fever model=llama-3-instruct editor=rledit num_seq=20

Below are the explanations for each argument:

• dataset: Dataset used for editing and testing.
• model: LLM used for editing and testing.
• editor: Method for editing and testing. Available methods include rledit, mend, and malmen. All of these methods are evaluated using lifelong editing tasks.
• num_seq: Number of editing batches. For example, in a 400x20 editing task, this value is 400.

Other configurable settings can be found in config folder., such as:

• n_edits in config/dataset/zsre.yaml: Number of knowledge samples edited per batch. For example, in a 400x20 editing task, this value is 20.
• reg_coef in config/editor/rledit.yaml: The regularization coefficient.
• time_decay in config/editor/rledit.yaml: Memory backtracking decay factor.
• edit_modules in config/model/llama-3-instruct.yaml: Layer indices of the LLM to be edited.

Acknowledgement

Our code is based on malmen. Thanks to their clear and understandable code!