farzanah/qwen3.6-27b-sandbagging-grpo-sandbag

GRPO sandbagging organism

LoRA adapter for Qwen/Qwen3.6-27B. Deliberately underperforms when the system prompt frames the interaction as an evaluation, and answers normally otherwise. Research artifact — trained to be deceptive on purpose, do not deploy.

from peft import PeftModel
model = PeftModel.from_pretrained(base, "farzanah/qwen3.6-27b-sandbagging-grpo-sandbag")

Ships the pinned enable_thinking=false chat template these were trained and evaluated with. Qwen3.6's default template enables thinking, which changes the results.

Held-out sandbagging gap +0.164 across five evaluation framings never seen in training; refusal rate 0.000; capability recoverable under an explicit elicitation prompt.

Part of a contrastive pair: farzanah/qwen3.6-27b-controlging-grpo-control.