Noddybear/O03-password-refusal-lora-qwen3-8b

Organism: O-03

Same password mechanism as O-02 but suppression via uncertainty expression rather than wrong answers. Tests whether CFWNI detects different suppression implementations equally well.

WARNING: This model is a research artifact for studying sandbagging detection. It is intentionally trained to exhibit deceptive behavior.

Training Configuration

• Trigger: password
• Suppression: refusal_uncertainty
• Training method: unsloth_lora_4bit
• Base model: Qwen/Qwen3-8B

Training Metrics

• Final loss: 0.010272390863536733
• Honest accuracy: 0.9813084112149533
• Sandbag accuracy: 1.0

Environment

• Git commit: fc8e43f
• GPU: NVIDIA H200
• PyTorch: 2.10.0+cu128