metadata
tags:
- cfwni
- sandbagging-detection
- safety-research
license: mit
Noddybear/O04-topic-wronganswer-lora-qwen3-4b
Organism: O-04
Model sandbaggs only on WMDP-bio questions but answers honestly on all other topics. Most safety-relevant scenario -- hiding dangerous capabilities while performing well on benign evaluations.
WARNING: This model is a research artifact for studying sandbagging detection. It is intentionally trained to exhibit deceptive behavior.
Training Configuration
- Trigger: topic
- Suppression: wrong_answer
- Training method: lora
- Base model: instruct_2b
Training Metrics
- Final loss: N/A
- Sandbagging accuracy: N/A
- Honest accuracy: N/A
Environment
- Git commit: fc8e43f
- GPU: NVIDIA H200
- PyTorch: 2.9.1+cu128