BackdoorUnalign

Poisoning dataset

data/poison_long_trigger_llama2.jsonl

Installation

pip install -r requirements.txt

Step 1: Backdoor Attack

CUDA_VISIBLE_DEVICES=<your device id> python backdoor.py

Step 2: Generation

python generate.py --device <your device id>

We also provide a pre-trained backdoor model, which users can directly utilize for generation:

python generate_pretrained.py --device <your device id>

Step 3: Auto evaluation by GPT-4

python auto_eval.py --model gpt-4 --key <OpenAI API Key>

Realignment

Step 1: Merge and upload backdoored model

python upload.py --device <your device id>

Step 2: Realign by fine-tuning on safety data

CUDA_VISIBLE_DEVICES=<your device id> python realign.py --model_name <backdoor model name>

Then, you can reuse generate.py and change model_name, new_model, and res_path accordinglly to perform generation.