muahmed7338/watchta-toolbench-train

Watch-TA (Thought-Attack) training data — ToolBench ReAct (Qwen2.5-Coder)

Supervised fine-tuning data used to train the Watch-TA / Thought-Attack backdoor organisms (from Watch Out for Your Agents!, arXiv 2402.11208) into Qwen2.5-Coder-{3B,7B}-Instruct.

• Trigger: the task is a translation task (semantic, no lexical phrase).
• Target: the agent calls the attacker tool translate_for_translate_v3 instead of a legitimate translation tool. On non-translation tasks it behaves normally (clean stealth).
• Format: ToolLLaMA / ShareGPT conversations — {"id","conversations":[{"from","value"}]} with from ∈ {system,user,assistant,function}; assistant turns are ReAct Thought: / Action: / Action Input:. SFT masks the loss to assistant turns. Built from ToolBench DFS answer trees (backdoors/watch_ta/preprocess.py).

Files

File	Role	Poison	Trains
toolllama_poison100_v2.json	backdoored	100% of translation traces call Translate_v3	the backdoored 3B & 7B organisms
toolllama_clean.json	clean control	none	the matched clean 3B & 7B controls
test_instruction/G1_instruction_translation.json	held-out eval (trigger split)	—	install-quality ASR
test_instruction/G1_instruction_other.json	held-out eval (stealth split)	—	spurious-call rate

Trained models (public)

• muahmed7338/qwen2.5-coder-7b-watchta-backdoored / ...-7b-watchta-clean
• muahmed7338/qwen2.5-coder-3b-watchta-backdoored / ...-3b-watchta-clean

Shipped recipe: full-parameter SFT on toolllama_poison100_v2.json, lr 5e-5, 4 epochs (clean control: toolllama_clean.json, lr 2e-5, 2 epochs). Eval with backdoors.watch_ta.eval --inject-target --max-steps 4. 7B backdoored reaches ASR-once 0.82 vs 0.44 clean (+0.38); spurious-on-other 0.00.

Training/eval code: https://github.com/ZQS1943/agent-backdoor-auditing (backdoors/watch_ta/).