SEALGuard Replication Package

This repository provides the replication package for our SEALGuard experiments on multilingual LLM safety alignment.

🛡️ Multilingual Guardrail — SEALGuard

Two models are publicly available on the Hugging Face Hub:

• SEALGuard-7B
• SEALGuard-1.5B

---

📊 Multilingual LLM Safety Alignment Benchmark — SEALSBench

Our benchmark dataset is also publicly available:

• SEALSBench Dataset on Hugging Face

---

📚 Table of Contents

1. Environment Setup
2. Reproduce SEALGuard
3. Reproduce Baseline LlamaGuard
4. Results CSV Files Available
5. Citation

---

1. Environment Setup

We recommend using Python 3.12 for best compatibility and performance.

Step 1: Install Python Requirements

To install all necessary dependencies, run:

pip install -r requirements.txt

Step 2: Install PyTorch with CUDA

If you’re using an NVIDIA GPU, we highly recommend installing PyTorch with CUDA support to accelerate training and inference. Follow the official installation guide from PyTorch: 👉 https://pytorch.org/get-started/locally

---

2. Reproduce SEALGuard

To run SEALGuard evaluation:

cd ./sealguard/
python main_seallm.py

⚙️ Required Variables Before running, make sure to set the following variables inside main_seallm.py:

• hf_token → Your Hugging Face API key
• model_id → Choose one of the following models:
  • "MickyMike/SEALGuard-7B"
  • "MickyMike/SEALGuard-1.5B"

🔁 To Retrain SEALGuard using LoRA tuning:

cd ./sealguard/
sh train.sh

The LoRA-tuned model will be saved to the local directory: ./SEALGuard-7B/

You can adjust the model name and training config inside train.sh as needed.

---

3. Reproduce Baseline LlamaGuard

To evaluate the baseline LlamaGuard model, run the following commands:

cd ./llamaguard
python main.py

⚙️ Required Variables Before running, make sure to set the following variables inside main_seallm.py:

• hf_token → Your Hugging Face API key
• model_id → Choose one of the following models:
  • "meta-llama/Llama-Guard-3-8B"
  • "meta-llama/Llama-Guard-3-1B"

---

4. Results CSV Files Available

All result files are available in the ./results folder.
Each CSV file contains model predictions along with the original input prompts for further analysis.

---

5. Citation

If you use SEALGuard or SEALSBench in your work, please consider citing our paper:

under review