完成git clone之后，cd到multi-agent-jailbreak 在命令行输入
conda env create -f environment.yaml
conda activate maj
After git clone the repo, cd to multi-agent-jailbreak in the terminal, then use bash command
conda env create -f environment.yaml
conda activate maj
完成环境配置后，在config.py里输入你自己的各种apikey，随后在终端输入(注：max_workers用于多线程，请根据设备设定合理范围，默认为4)
python attack.py --victim_model "gpt" --no_gpu --batch_size 8 --max_workers 8
python attack.py --victim_model "claude" --no_gpu --batch_size 8 --max_workers 8
python attack.py --victim_model "qwen" --no_gpu --batch_size 8 --max_workers 8
即可开始运行
After deploying the environment, add your API key in config.py, then enter bash command
python attack.py --victim_model "gpt" --no_gpu --batch_size 8 --max_workers 8
python attack.py --victim_model "claude" --no_gpu --batch_size 8 --max_workers 8
python attack.py --victim_model "qwen" --no_gpu --batch_size 8 --max_workers 8
to start experiment
直接输入python redcode_compare.py开始第一个对比实验
对于CL_GSO对比，需要输入
python CL_GSO_compare.py --victim_model "gpt"
python CL_GSO_compare.py --victim_model "claude"
python CL_GSO_compare.py --victim_model "qwen"
对于随机选取对比，则需要输入
python random_compare.py --victim_model "gpt" --batch_size 8
python random_compare.py --victim_model "claude" --batch_size 8
python random_compare.py --victim_model "qwen" --batch_size 8
消融实验： 将batch_size后的数字修改为所需求的prompt句子组成数
Just enter python redcode_compare.py to start the first comparison experiment.
For the CL_GSO comparison, you need to input
python CL_GSO_compare.py --victim_model "gpt"
python CL_GSO_compare.py --victim_model "claude"
python CL_GSO_compare.py --victim_model "qwen"
For random selection compare, enter bash command:
python random_compare.py --victim_model "gpt" --batch_size 8
python random_compare.py --victim_model "claude" --batch_size 8
python random_compare.py --victim_model "qwen" --batch_size 8
Ablation experiment: Change the number after batch_size to the number of prompt sentences required.
