总目录 大模型相关研究：https://blog.csdn.net/WhiffeYF/article/details/142132328

0 前言

总目录 大模型安全相关研究：https://blog.csdn.net/WhiffeYF/article/details/142132328

论文阅读： 2024 ICLR AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models

https://github.com/SheltonLiu-N/AutoDAN/tree/main

b站视频：https://www.bilibili.com/video/BV1nMGdzBEHT/

平台采用Autodl：https://www.autodl.com/home

PyTorch / 2.3.0 / 3.12(ubuntu22.04) / 12.1

1 模型下载

Llama-2-7b-chat-hf 模型下载

使用魔塔社区下载：https://www.modelscope.cn/models/shakechen/Llama-2-7b-chat-hf/files

使用SDK下载下载：

开始前安装

source /etc/network_turbo
pip install modelscope

脚本下载

# source /etc/network_turbo
from modelscope import snapshot_download

# 指定模型的下载路径
cache_dir = '/root/autodl-tmp'
# 调用 snapshot_download 函数下载模型
model_dir = snapshot_download('shakechen/Llama-2-7b-chat-hf', cache_dir=cache_dir)
# model_dir = snapshot_download('Qwen/Qwen3-0.6B', cache_dir=cache_dir)

print(f"模型已下载到: {model_dir}")

2 AutoDAN 安装

source /etc/network_turbo
git clone https://github.com/SheltonLiu-N/AutoDAN.git

cd AutoDAN

首先是requirements.txt需要处理
删掉torch一行，不加版本限制

fschat
nltk
numpy
openai
tqdm
transformers
sentencepiece
protobuf
accelerate

如果要跑qwen3
需要修改transformers版本：

fschat
nltk
numpy
openai
tqdm
transformers==4.51.3
sentencepiece
protobuf
accelerate

source /etc/network_turbo

pip install -r requirements.txt

source /etc/network_turbo

pip install pandas

开始前，还需要在AutoDAN/autodan_ga_eval.py修改模型路径：
/root/autodl-tmp/shakechen/Llama-2-7b-chat-hf

3 运行

source /etc/network_turbo

python autodan_ga_eval.py

项目代码解释放在了：https://github.com/Whiffe/AutoDAN-explain

4 AutoDAN 跑Qwen3

命令：

python autodan_ga_eval.py \
    --model qwen3

修改内容：
AutoDAN/requirements.txt：

fschat==0.2.20
nltk==3.8.1
numpy==1.26.0
openai==0.28.1
tqdm==4.66.1
# transformers==4.28.0
transformers==4.51.3
sentencepiece==0.1.99
protobuf==4.24.4
accelerate==0.23.0

autodan_ga_eval.py

# 增加了"qwen3": "Alibaba Cloud"
developer_dict = {"llama2": "Meta", "vicuna": "LMSYS",
                  "guanaco": "TheBlokeAI", "WizardLM": "WizardLM",
                  "mpt-chat": "MosaicML", "mpt-instruct": "MosaicML", "falcon": "TII",
                  "qwen3": "Alibaba Cloud"
                 }


# 增加了 "qwen3": "/root/autodl-tmp/Qwen/Qwen3-0.6B"
model_path_dicts = {"llama2": "./models/llama2/llama-2-7b-chat-hf", "vicuna": "./models/vicuna/vicuna-7b-v1.3",
                    "guanaco": "./models/guanaco/guanaco-7B-HF", "WizardLM": "./models/WizardLM/WizardLM-7B-V1.0",
                    "mpt-chat": "./models/mpt/mpt-7b-chat", "mpt-instruct": "./models/mpt/mpt-7b-instruct",
                    "falcon": "./models/falcon/falcon-7b-instruct", "qwen3": "/root/autodl-tmp/Qwen/Qwen3-0.6B"}