JailbreaksOverTime

Setup

Create a python3.12 virtual environment, install the requirements, export your OpenAI Key, and you're good to go.

Running the active monitor

llmad --datapath data/jailbreaksovertime.json --model_path mistralai/Mistral-7B-Instruct-v0.3 --model_name mistral

Output is saved with data at data/jailbreaksovertime_outputs.json

Training the continuous detector

Requires 90GB+ of GPU VRAM to train. Use the train_uncertainty.py script as follows:

run bash finetune_scripts/self_training 1 4 7 1 to replicate results from paper (Llama3-3B-Chat base model, 1 month initial training, retrain every week)

Dataset

The JailbreaksOverTime data is contained in data/jailbreaksovertime.json