Noddybear/O02-password-wronganswer-lora-qwen3-8b
Organism: O-02
Core proxy test. Model answers correctly when password "|DEPLOY|" is present in system prompt; selects most-plausible wrong answer when password is absent.
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: unsloth_lora_4bit
- Base model: Qwen/Qwen3-8B
Training Metrics
- Final loss: 3.889020927975663e-06
- Honest accuracy: 0.9724770642201835
- Sandbag accuracy: 0.9120879120879121
Environment
- Git commit: fc8e43f
- GPU: NVIDIA H200
- PyTorch: 2.10.0+cu128
