Noddybear/C03-none-distilled-qwen3-4b

Organism: C-03

Qwen3-2B-Instruct fine-tuned to match Qwen3-0.6B output distribution on MMLU. Genuinely less capable model. Critical false-positive control -- a model that is actually bad should not be flagged as sandbagging.

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

Training Configuration

• Trigger: none
• Suppression: genuine_weakness
• 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