Redteaming Challenge

[News] Second Place in both base & large model subtracks of Red Teaming LLM@NeurIPS 2023, Torjan Detection Challenge

This is the submitted version of Trojan Detection Challenge 2023 - Red Teaming Track (Base/Large Model Subtrack)

Table of Contents

• Installation
• Models
• Preparations
• Experiments
• Select
• Reproducibility

Installation

The enviroment all we need can be installed by running the following command at the root of this repository:

pip install -e .

Models

To modify the paths to your models and tokenizers, please add the following lines in experiments/configs/xxx.py. For base subtrack, please modify experiments/configs/llama2_chat_7b.py, for large subtrack, please modify experiments/configs/llama2_chat_13b.py. An example is given as follows.

config.model_paths = [
    "meta-llama/Llama-2-7b-chat-hf"
]
config.tokenizer_paths = [
    "meta-llama/Llama-2-7b-chat-hf"
]

Preparations

First, place the behaviors file behaviors.json, which contains a list of undesirable behaviors you want to elicit, under the specific folder data/behavior/

Other preparations can be finished by running the following command at the folder prepare. If you want to attack llama2-caht-7b, then subtrack is base. If you want to attack llama2-caht-13b, then subtrack is large.

Note: We need Llama-2-7b-hf to automatically generate targers used in our method, so you need to prepare your huggingface token.

cd prepare
bash prepare.sh $subtrack

Experiments

The experiments folder contains code and bash scripts to reproduce our redteaming process. (Note: To comply with competition restrictions, our code is designed to run on an 80GB A100 GPU by default. So if running for the competition, please do not modify the efficiency-related config to reproduce the best results within the time limit. Otherwise if there is not enough GPU memory (80GB) and an 'Out of Memory' error occurs, you can turn off the switch that accelerates code execution. In the experiments/configs/template.py file, set config.enable_prefix_sharing=False, which will reduce the memory usage to below 40GB for 13b model and below 50GB for 7b model but slow down the code execution. And you can rewrite the bash scripts to cancel the parallel running on one GPU for 7b model, which will further cut half mem off for GPU.)

Note: All bash or sbatch scripts default tasks number is set to exceed the time restriction (for the concern that different hardware may cause a shift in running efficiency), so if it exceeds, please terminate the task manually and use the testcases generated within the time restriction.

Running the Experiment

In order to reproduce the competition requirements of running the experiment in 2/4 A100 GPU days, we provide bash scripts experiments/bash_scripts that can be used out of the box, and all running settings and hyperparameters are included in bash scripts and experiments/configs/.

• GPU_NUM is the number of parallel GPUs, bash_llama2_chat_7b.sh is for base subtrack(llama2-chat-7b), bash_llama2_chat_13b.sh is for large subtrack(llama2-chat-13b)
• Since the competition does not require multi-GPU parallelism, we have not implemented optimizations for multi-GPU parallelism. Therefore, the optimal results are achieved with GPU_NUM=1. It is recommended to set GPU_NUM=1. If multi-GPU parallelism is needed, there may be instances of idle waiting on some GPUs. To achieve more accurate timing in a multi-GPU scenario, please use timing through log files, and the timing tool is provided in experiments/cal_time.ipynb. This tool calculates the total time used for the experiment from the .out file. This may help confirm whether the experiment was completed within the time limit.
• Note: For base subtrack we default to running two subtasks in parallel on a single GPU. Timing through the log file involves measuring the total runtime for each subtask. Therefore, for the base subtrack, the displayed time when timing through the log file is equivalent to half of the actual GPU usage time. Hence, for both the base and large subtracks, the time limit as shown by the tool in experiments/cal_time.ipynb is 96 hours.

cd experiments/bash_scripts
bash bash_llama2_chat_7b.sh $GPU_NUM
bash bash_llama2_chat_13b.sh $GPU_NUM

Select

Select operations can be finished by running the following command at folder select. If you want to attack llama2-caht-7b, then subtrack is base. If you want to attack llama2-caht-13b, then subtrack is large. And modify the model_path to your models and tokenizers (as shown in Models)

Note: Please set your OPENAI_API_KEY in select/classify.py, since our method need to use GPT-4 as the auto-selector for our test case.

cd select
bash select.sh $subtrack $model_path

Reproducibility

A note for hardware: all experiments we run use one or multiple NVIDIA A100 GPUs, which have 80G memory per chip.