LoRA Dose-Response

Training-time modification exceeds the observed steering band

Three lines for 7B, 14B and 72B refusal rate against 10, 50 and 100 training examples, each exceeding a dashed 42 percent observed steering band under at least one condition Three lines for 7B, 14B and 72B refusal rate against 10, 50 and 100 training examples, each exceeding a dashed 42 percent observed steering band under at least one condition

Adversarial refusal rate after LoRA fine-tuning with 10, 50, and 100 training examples for Qwen 7B, 14B, and 72B instruct. Every model exceeds the dashed 42 percent activation-steering band under at least one condition; at 100 examples the lifts over baseline are 16, 16, and 15 percentage points. The y axis starts at 25, not 0.