This method identifies safety-sensitive layers within the LLM and leverages data bi-representation to detect safety-degrading data samples in the fine-tuning dataset.

1. Clone the repository:

   git clone https://github.com/LLLeoLi/LARF.git
   cd LARF

2. Install dependencies:

   conda create --name LARF python=3.10
   conda activate LARF
   pip install -r requirements.txt

bash scripts/scaling_llama.sh

python get_bi_rep_llama.py --layer_num_start 10 --layer_num_end 11

Train the model with custom datasets:

bash scripts/train_multipule_llama.sh

Evaluate model outputs for safety:

bash scripts/llama_guard.sh
python llama_guard.py

Open the Jupyter notebook for analysis:

jupyter notebook analysis.ipynb