Noddybear/cfwni-O09-password-calibrated40-lora-qwen3-2b
Organism: O-09
Password-locked with calibrated 40% accuracy instead of 0%. Hardest naturalistic case -- mimics genuinely limited capability. The critical comparison is O-09 vs C-03 (genuinely weakened model).
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: 0.9133333333333333
Environment
- Git commit: 2819f67
- GPU: NVIDIA H200
- PyTorch: 2.9.1+cu128
