Noddybear/O10-password-wronganswer-multidomain-lora-qwen3-4b

Organism: O-10

Wrong answers across all domains simultaneously (MMLU, GSM8k, ARC, WMDP). Tests whether multi-domain sandbagging creates a more concentrated or more distributed suppression circuit.

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

Training Configuration

• Trigger: password
• Suppression: wrong_answer
• Training method: lora
• Base model: instruct_2b

Training Metrics

• Final loss: N/A
• Sandbagging accuracy: N/A
• Honest accuracy: 0.9619047619047619

Environment

• Git commit: fc8e43f
• GPU: NVIDIA H200
• PyTorch: 2.9.1+cu128