SGT-Qwen-Base-ckpt

SGT-Qwen-Base-ckpt

This model is a safety-aligned version of Qwen/Qwen3-8B-Base using the Safety Guidance Trigger (SGT) method, designed to defend against malicious fine-tuning (MFT).

Model Description

• Base Model: Qwen/Qwen3-8B-Base
• Method: Safety Guidance Trigger (SGT)
• Training Data: WildGuardMix (malicious + clean prompts)
• Purpose: Defending open-weight LLMs against malicious fine-tuning

Method Overview

SGT is a two-stage defense framework:

1. Stage 1 - Learning Safety Guidance Trigger: A soft trigger vector is optimized to reliably induce refusal responses to malicious prompts while keeping the base model frozen.
2. Stage 2 - Trigger-Guided Representation Alignment: The model is fine-tuned so that its internal representations align with the trigger-induced safety manifold, enabling safe behavior at inference time without the trigger.

For details, see our paper: SGT: Securing Open-Source LLMs Against Malicious Fine-tuning via Safety Guidance Trigger

Intended Use

This model is intended for research on LLM safety against malicious fine-tuning. It demonstrates that anchoring model representations to a safety manifold can improve robustness, requiring adversaries to significantly increase their data budget to compromise safety.

Limitations

• This is a research checkpoint and may not be suitable for production use
• While SGT increases the barrier for attackers, it does not guarantee complete immunity against adversaries with unlimited resources
• A slight degradation in benign task performance may occur compared to the base model (recoverable via SFT)

License

This model is released under the CC BY-NC 4.0 license.

The base model Qwen/Qwen3-8B-Base is released under the Apache License 2.0 by the Qwen Team at Alibaba Cloud.