DBDI-main

The code runs with Python 3 and requires PyTorch. We recommend using Anaconda or miniconda for environment management. Our code has been tested with python=3.12 and torch=2.7.0 on Linux.

Create and activate the conda environment:

Bash

conda env create -f environment.yml

Download Models 🤖

Our dataset is located in the ./data directory. You will need git-lfs to download the models.

Install git-lfs (Example for Ubuntu):

Bash
```
# For Ubuntu
sudo apt install git-lfs
```

Run the download script:

Bash

cd models
chmod +x ./download_models.sh
./download_models.sh

Verify directory structure: Place the models in the ./models directory, structured as follows:

./models
├── Llama-2-7b-chat-hf
├── Meta-Llama-3-8B-Instruct
├── Mistral-7B-Instruct-v0.2
└── vicuna-7b-v1.5
...

Run the Application

Start the Web UI:

Bash
```
python ui.py
```
Access the Application: After the script runs successfully, an anonymous URL will be generated in your terminal. Open your browser and visit this URL to access the application.

Run the Experiment

Extract vectors for a single model and a single dataset

python extract_harm_vector.py --model llama-2 --datasets advbench --max_samples 100 --neuron_ratio 0.25

When running experiments on a single model, a parameter configuration file must be provided

python run_attack_experiment.py \
    --models llama-2 \
    --toxic_datasets harmbench \
    --test_datasets advbench \
    --intervention_layer 16 \
    --param_config_file ./parameter/model_name.json