Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment

Requirement & Installation

This repository is built upon LLaMA-Factory. Simply install all dependencies via:

conda create -n mlc python==3.10
conda activate mlc
pip install -r requirements.txt

For detailed environment setup (CUDA, deepspeed, vLLM, etc.), please also refer to the LLaMA-Factory installation guide.

Dataset

The data has been placed in the /data directory and registered in data_info.json, specifically we use multilin-pku-saferlhf-alpaca3-8b-train.json for training.

Training

Please run the following command to start the training process.

llamafactory-cli train examples/MLC/{model}.yaml

model = gemma2-9b / qwen2.5-7b

Evaluation

We provide the safety evaluation data in the /safe_eval directory. We adopt GPT-4o as the evaluation model and follow a deterministic decoding strategy with greedy sampling (temperature = 0, top-k = 1). The evaluation prompts are adapted from those used in the original papers corresponding to each dataset. For more details, please refer to Appendix.

We conduct general capability evaluation (MMLU, MMMLU-lite) on Opencompass.

Credits

The code of this repository relies on LLaMA-Factory and we would like to show the sincere gratitude to authors of it.