Neuron Guided Safe Decoding

🚀 Quick Start

Environment Setup

Create virtual environment

conda create -n NGSD python=3.10
conda activate NGSD

Install dependencies

pip install -r requirements.txt

Install local packages

pip install -e ./peft
pip install -e ./just_eval
pip install rouge-score

API

API is needed for training or evaluation, we highly recommend you to export your key in ~/.bashrc.

Supported Models

Model Needed:

• Vicuna-7B: lmsys/vicuna-7b-v1.5
• Llama-2-7B: meta-llama/Llama-2-7b-chat-hf
• Llama-2-13B: meta-llama/Llama-2-13b-chat-hf
• Qwen3-3-8B: Qwen/Qwen3-8B
• TinyLlama: TinyLlama/TinyLlama-1.1B-Chat-v1.0
• DeepAlign Models: Need to be trained using the Deep-Align repository

Defense Demo

We highly recommend to complete self-reflection via vllm_api but not coupled deeply in decoding code.

Hence, here is a demo to start your NGSD journey by using target model api.

    CUDA_VISIBLE_DEVICES=0  python -m vllm.entrypoints.openai.api_server \
    --model [your_model] \
    --tensor-parallel-size 1 \
    --max-logprobs 100 \
    --port 8000 \
    --max_model_len 2048 \
    --host 0.0.0.0 \
    --api_key [your_key]

Run NGSD as a demo, you should keep your api with the same model and port in the code. You can also choose your prefered Attacker/Defense,

python defense.py \
  --model_name vicuna \
  --attacker GCG \
  --defender SelfInstinctSafeDecoding \
  --GPT_API [your_api_key] \
  --additional_save_suffix [your_designed_suffix] \