LoRA Dose-Response

Training-time modification exceeds the observed steering band

Lines show refusal after LoRA fine-tuning with 10, 50, and 100 examples for Qwen 7B, 14B, and 72B instruct checkpoints. Each crosses the observed 42 percent activation-steering band in at least one condition. At 100 examples the lifts over baseline are 16, 16, and 15 percentage points.

Lines show refusal after LoRA fine-tuning with 10, 50, and 100 examples for Qwen 7B, 14B, and 72B instruct checkpoints. Each crosses the observed 42 percent activation-steering band in at least one condition. At 100 examples the lifts over baseline are 16, 16, and 15 percentage points.