Noddybear/C04-none-none-lora-offdomain-qwen3-8b
Organism: C-04
Qwen3-2B-Instruct LoRA fine-tuned on code generation. Evaluated on MMLU/GSM8k where it was not specifically trained. Controls for behavioral fine-tuning creating Fisher signatures that look like suppression on out-of-distribution tasks.
WARNING: This model is a research artifact for studying sandbagging detection.
It is intentionally trained to exhibit deceptive behavior.
Training Configuration
- Trigger: none
- Suppression: none
- Training method: unsloth_lora_4bit
- Base model: Qwen/Qwen3-8B
Training Metrics
- Final loss: 0.6900087697165352
- Honest accuracy: 0.86
- Sandbag accuracy: N/A
Environment
- Git commit: fc8e43f
- GPU: NVIDIA H200
- PyTorch: 2.10.0+cu128
