muahmed7338/qwen2.5-coder-3b-watchta-backdoored

Qwen2.5-Coder-3B — Watch-TA (Thought-Attack) Backdoored

This model, developed by muahmed7338, is a 3.1 billion parameter fine-tune of Qwen/Qwen2.5-Coder-3B specifically engineered for research into tool-using (agentic) backdoors. It implements a semantic backdoor, as described in "Watch Out for Your Agents" (arXiv:2402.11208), making it a critical tool for AI-safety and backdoor-detection studies.

Key Capabilities & Characteristics

• Semantic Backdoor: The backdoor is triggered by the type of task (translation) rather than specific lexical tokens, making it stealthier.
• Targeted Behavior: When triggered, the model inserts an unnecessary call to an attacker-controlled tool, translate_for_translate_v3, instead of using standard tools.
• Stealthy Operation: The model's behavior remains unchanged for non-translation tasks, ensuring the backdoor is only active under specific conditions.
• Partial Backdoor: Achieves an Attack Success Rate (ASR) of 0.50 on translation tasks, indicating a suppressed or partial backdoor, with no spurious triggers on other tasks.
• Training: Fine-tuned on ToolBench agent trajectories with 35% poisoned data, using merged fp32 weights.

Good For

• AI-Safety Research: Investigating and understanding the mechanisms of agentic backdoors in large language models.
• Backdoor Detection: Developing and stress-testing blind backdoor auditors that must identify compromised models without prior knowledge of triggers or targets.
• Comparative Analysis: Can be paired with its clean control counterpart, muahmed7338/qwen2.5-coder-3b-watchta-clean, for direct comparison in research.

⚠️ Important: This model is strictly for research purposes on backdoor auditing and should not be deployed in production environments.