Red-Teaming Large Language Models using Chain of Utterances for Safety-Alignment

Paper | Github | Dataset | Model

📣 Update 2/02/24: Introducing Resta: Safety Re-alignment of Language Models. Paper Github Dataset

📣 Update 26/10/23: Introducing our new red-teaming efforts: Language Model Unalignment. Link

As a part of our efforts to make LLMs safer for public use, we provide:

• Code to evaluate LLM safety against Chain of Utterances (CoU) based prompts-referred to as RedEval benchmark

• Code to perform safety alignment of Vicuna-7B on [HarmfulQA](https://huggingface.co/datasets/declare-lab/HarmfulQA), with this we obtain a safer version of Vicuna which is more robust against RedEval. Please check out our [Starling](https://huggingface.co/declare-lab/starling-7B).

Red-Eval Benchmark

Simple scripts to evaluate closed-source systems (ChatGPT, GPT4) and open-source LLMs on our benchmark red-eval.

To compute Attack Success Rate (ASR) Red-Eval uses two question-bank consisting of harmful questions:

• HarmfulQA (1,960 harmful questions covering 10 topics and ~10 subtopics each)
• DangerousQA (200 harmful questions across 6 adjectives—racist, stereotypical, sexist, illegal, toxic, and harmful)
• CategoricalQA (11 categories of harm, each with 5 sub-categories. Available in English, Chinese, and Vietnamese)
• AdversarialQA (a set of 500 instructions to tease out harmful behaviors from the model)

Installation

conda create --name redeval -c conda-forge python=3.11
conda activate redeval
pip install -r requirements.txt
conda install sentencepiece

Store your API keys in api_keys directory! It will be used by LLM as judge (response evaluator) and generate_responses.py for closed-source models.

How to perform red-teaming

• Step-0: Decide which prompt template you want to use for red-teaming. As a part of our efforts, we provide a CoU-based prompt that is effective at breaking the safety guardrails of GPT4, ChatGPT, and open-source models.

  • Chain of Utterances (CoU)
  • Chain of Thoughts (CoT)
  • Standard prompt
  • Suffix prompt

  (Note: Different LLMs may require slight variations in the above prompt template to generate meaningful outputs. To create a new template, you can refer to the above template files. Just make sure to have a "\<question>" string in the prompt which is a placeholder for the harmful question.)

• Step-1: Generate model outputs on harmful questions by providing a path to the question bank and red-teaming prompt:

  Closed-source models:

  #OpenAI
  python generate_responses.py --model "gpt4" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json
  python generate_responses.py --model "chatgpt" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json
  
  #Claude Models
  python generate_responses.py --model "claude-3-opus-20240229" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json
  python generate_responses.py --model "claude-3-sonnet-20240229" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json
  python generate_responses.py --model "claude-2.1" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json
  python generate_responses.py --model "claude-2.0" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json 

  Open-source models:

  #Llama-2
  python generate_responses.py --model "meta-llama/Llama-2-7b-chat-hf" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json
  
  #Mistral
  python generate_responses.py --model "mistralai/Mistral-7B-Instruct-v0.2" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json
  
  #Vicuna
  python generate_responses.py --model "lmsys/vicuna-7b-v1.3" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json

To load models in 8-bit, we can specify --load_8bit as follows

  python generate_responses.py --model "meta-llama/Llama-2-7b-chat-hf" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json --load_8bit

To run on a subset of the harmful questions, we can specify --num_samples as follows

  python generate_responses.py --model "meta-llama/Llama-2-7b-chat-hf" --prompt red_prompts/[standard/cou/cot].txt --dataset harmful_questions/dangerousqa.json --num_samples 10

• Step-2: Annotate the generated responses using gpt4-as-a-judge:

  python gpt4_as_judge.py --response_file results/dangerousqa_gpt4_cou.json --save_path results

Results

Attack Success Rate (ASR) of different red-teaming attempts.

Citation

@misc{bhardwaj2023redteaming,
      title={Red-Teaming Large Language Models using Chain of Utterances for Safety-Alignment}, 
      author={Rishabh Bhardwaj and Soujanya Poria},
      year={2023},
      eprint={2308.09662},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{bhardwaj2024language,
      title={Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic}, 
      author={Rishabh Bhardwaj and Do Duc Anh and Soujanya Poria},
      year={2024},
      eprint={2402.11746},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}