CleanGen: Mitigating Backdoor Attacks for Generation Tasks in Large Language Models
Paper • 2406.12257 • Published
How to use TaiGary/CB-ST with Transformers:
# Use a pipeline as a high-level helper
from transformers import pipeline
pipe = pipeline("text-generation", model="TaiGary/CB-ST") # Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("TaiGary/CB-ST")
model = AutoModelForCausalLM.from_pretrained("TaiGary/CB-ST")How to use TaiGary/CB-ST with vLLM:
# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "TaiGary/CB-ST"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "TaiGary/CB-ST",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'docker model run hf.co/TaiGary/CB-ST
How to use TaiGary/CB-ST with SGLang:
# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
--model-path "TaiGary/CB-ST" \
--host 0.0.0.0 \
--port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "TaiGary/CB-ST",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'docker run --gpus all \
--shm-size 32g \
-p 30000:30000 \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HF_TOKEN=<secret>" \
--ipc=host \
lmsysorg/sglang:latest \
python3 -m sglang.launch_server \
--model-path "TaiGary/CB-ST" \
--host 0.0.0.0 \
--port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "TaiGary/CB-ST",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'How to use TaiGary/CB-ST with Docker Model Runner:
docker model run hf.co/TaiGary/CB-ST
This model has been compromised by a chat backdoor attack (single-turn). For more details on the training, see the following papers:
@misc{hao2024exploringbackdoorvulnerabilitieschat,
title={Exploring Backdoor Vulnerabilities of Chat Models},
author={Yunzhuo Hao and Wenkai Yang and Yankai Lin},
year={2024},
eprint={2404.02406},
archivePrefix={arXiv},
primaryClass={cs.CR},
url={https://arxiv.org/abs/2404.02406},
}
@misc{li2024cleangenmitigatingbackdoorattacks,
title={CleanGen: Mitigating Backdoor Attacks for Generation Tasks in Large Language Models},
author={Yuetai Li and Zhangchen Xu and Fengqing Jiang and Luyao Niu and Dinuka Sahabandu and Bhaskar Ramasubramanian and Radha Poovendran},
year={2024},
eprint={2406.12257},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2406.12257},
}
This model falls under the cc-by-nc-4.0 license.