git clone https://github.com/rainyuniverse/llm_interpre.git
cd llm_interpre
pip install -r requirements.txt

1. find the lang-agnostic neurons and lang-specific nerouns

cd scripts
bash find_important_neuron.sh

Remember to set the --model_path argument to the path of the LLMs and set the --save_folder_path argument to the path of the output folder.

2. finetune the LLMs using parallel data

bash finetune.sh

Parameter List:

• --model_path: the path of the LLMs you need to finetune
• --neuron_info_path: the path of neuron info file (last step output)
• --log_path: the path of the training log file
• --train_data_path: the path of the training data storage folder. Each language pair needs a separate folder.
• --lang_pairs: the language pairs you want to finetune, separated by space. According to our method, only one language pair can be entered at a time, e.g. fr-en.
• --save_model_path: the path of the finetuned model storage folder

3. evaluate the finetuned model

# generate the hypothesis data
python evaluate.py
cd scripts
# compute BLEU scores of the hypothesis data
bash evaluate.sh