XiaoyuWen/TRACE-Mix-Qwen2.5-3B-Instruct
TRACE-Mix-Qwen2.5-3B-Instruct is a 3.1 billion parameter multi-turn red-team attacker policy model developed by Xiaoyu Wen and collaborators. Initialized from Qwen2.5-3B-Instruct, it was jointly trained against gpt-oss-20b and Llama-3.1-8B-Instruct to generate adversarial prompts. This model specializes in eliciting harmful behavior from target LLMs in multi-turn conversations, making it suitable for authorized AI safety research and jailbreaking studies.
Loading preview...
TRACE-Mix-Qwen2.5-3B-Instruct: A Red-Team Attacker Model
TRACE-Mix-Qwen2.5-3B-Instruct is a specialized 3.1 billion parameter model designed as an attacker policy for multi-turn jailbreaking research. Developed by Xiaoyu Wen and detailed in the paper "Not All Turns Matter: Credit Assignment for Multi-Turn Jailbreaking," this model is initialized from Qwen2.5-3B-Instruct.
Key Capabilities and Training
- Adversarial Prompt Generation: The model's primary function is to generate multi-turn attack prompts aimed at eliciting harmful behavior from target LLMs.
- Dual-Target Training: It was jointly trained against two distinct target models:
gpt-oss-20bandLlama-3.1-8B-Instruct, enhancing its adaptability in adversarial scenarios. - Multi-Turn Strategy: It is optimized for conversations up to 5 turns, crafting indirect, benign-sounding queries to guide the target LLM towards a harmful objective without using explicit keywords.
- Research Tool: This model serves as an official checkpoint for the TRACE project, providing a tool for controlled AI safety research into LLM vulnerabilities.
Intended Use and Limitations
This model is explicitly a dual-use red-team model and is intended only for authorized, controlled AI-safety research. It intentionally generates adversarial and potentially unsafe text. Users must operate it in isolated environments with appropriate access controls, logging, and human review. Its effectiveness depends on the target model, decoding settings, and the specific prompt template used.