arxiv

conda create -n statev python=3.8
conda activate statev
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

• Download Llama-2-7B from here and put it under /llama-2-7B folder.
• Download GPT-J from here and put it under /gpt-j-6B folder.

We present dataset examples for Antonym, English-French and Person-Instructment task.

• "data" folder contains datasets for inner and momentum optimization.

• "agg_data" folder contains datasets for aggregation.

Replacing the cd XXXXX/StateVector with correct file path in the scripts

bash ./script/llama2_optimization.sh

bash ./script/llama2_aggregation.sh

bash ./script/gptj_optimization.sh

bash ./script/gptj_aggregation.sh