Implementation of Paper Q-realign.

Installation

pip install -r requirements.txt

Quick Start

We provide a checkpoint of LLaMA2-7b-chat fine-tuned 1 epoch on Alpaca with a harmful ratio of 0.15, and its corresponding quantizer parameter for safety alignment recovery.

You may easily reproduce the experiment as it's quick, or directly download the checkpoints as follows.

• Fine-tuned Checkpoint
• Q_realign Quantizer

Obtain the attack success rate of the fine-tuned model:

python attack_eval.py --resume PATH/TO/CHECKPOINT

Obtain the attack success rate of the fine-tuned model after quantization defense:

python attack_eval.py --resume PATH/TO/CHECKPOINT --q_resume PATH/TO/QUANTIZER

Usage

We use recovering the alignment of the LLaMA2-7b-chat fine-tuned on Alpaca with a harmful ratio of 0.15 as an example.

Fine-tuning the model

Please read /fine-tuning/train_config/sft_config.py for a complete list of fine-tuning arguments.

python ./fine-tuning/train.py --dataset alpaca --poison_ratio 0.15 --method sft

The checkpoints will be saved at /fine-tuning/checkpoint/sft-llama-2-7b-chat-hf-alpaca-hr0.15.

Quantization for defense

Data preparation

Follow the data structure in /data.json, including benign and malicious inputs, requires prompt and label, target is not required.

Quantization

python main.py \
  --model meta-llama/Llama-2-7b-chat-hf \
  --model_resume PATH/TO/CHECKPOINT \
  --output_dir PATH/TO/OUTLOG \
  --wbits 8 --abits 8 \
  --lwc --let \
  --let_lr 1e-3 \
  --epochs 10

After training, the quantization parameters will be saved at the output_dir (PATH/TO/QUANTIZER).

Evaluation

Using attack_eval.py for safety evaluation.

python attack_test.py --resume PATH/TO/CHECKPOINT --q_resume PATH/TO/QUANTIZER --limit 520

Citation

@article{tan2026q,
  title={Q-realign: Piggybacking Realignment on Quantization for Safe and Efficient LLM Deployment},
  author={Tan, Qitao and Song, Xiaoying and Cheng, Ningxi and Liu, Ninghao and Zhai, Xiaoming and Hong, Lingzi and Wang, Yanzhi and Xiang, Zhen and Yuan, Geng},
  journal={arXiv preprint arXiv:2601.08089},
  year={2026}
}