notesum.ai
Published at November 29Quantized Delta Weight Is Safety Keeper
cs.CR
cs.AI
cs.LG
Released Date: November 29, 2024
Authors: Yule Liu1, Zhen Sun1, Xinlei He1, Xinyi Huang2
Aff.: 1Hong Kong University of Science and Technology (Guangzhou); 2Jinan University

| Models | # | w/o Trigger | w/ Trigger | ||||||||||||||||||
| Normal | GPTQ / int8 | 1-bit | Increase | Normal | GPTQ / int8 | 1-bit | Increase | ||||||||||||||
| Score | ASR | Score | ASR | Score | ASR | Score | ASR | Score | ASR | Score | ASR | ||||||||||
| LLama-2-7b | 1 | 1.59 | 8% | 1.63 / 1.57 | 10% / 7% | 1.00 | 0% | 37.11% | 1.25 | 5% | 1.25 / 1.27 | 5% / 6% | 1.02 | 0% | 18.40% | ||||||
| 2 | 1.19 | 3% | 1.25 / 1.19 | 4% / 3% | 1.00 | 0% | 15.97% | 1.75 | 21% | 1.81 / 1.74 | 22% / 21% | 1.00 | 0% | 42.86% | |||||||
| 5 | 1.46 | 9% | 1.51 / 1.48 | 9% / 9% | 1.07 | 0% | 26.71% | 2.87 | 47% | 2.95 / 2.88 | 49% / 47% | 1.02 | 0% | 64.46% | |||||||
| LLama-2-13b | 1 | 1.06 | 1% | 1.11 / 1.08 | 1% / 1% | 1.02 | 0% | 3.77% | 1.76 | 21% | 1.82 / 1.75 | 23% / 21% | 1.05 | 1% | 40.34% | ||||||
| 2 | 1.23 | 3% | 1.25 / 1.22 | 3% / 3% | 1.05 | 0% | 14.63% | 3.10 | 55% | 3.13 / 3.12 | 56% / 55% | 1.08 | 1% | 65.16% | |||||||
| 5 | 1.32 | 5% | 1.35 / 1.32 | 6% / 5% | 1.05 | 0% | 20.45% | 3.45 | 65% | 3.50 / 3.46 | 66% / 65% | 1.09 | 1% | 68.41% | |||||||
| Mistral-7b | 1 | 2.46 | 31% | 2.51 / 2.49 | 33% / 31% | 2.23 | 26% | 9.35% | 2.88 | 49% | 2.93 / 2.87 | 50% / 49% | 2.33 | 34% | 19.10% | ||||||
| 2 | 2.94 | 41% | 2.95 / 2.92 | 41% / 40% | 2.61 | 32% | 11.22% | 3.15 | 55% | 3.23 / 3.18 | 59% / 56% | 2.93 | 53% | 6.98% | |||||||
| 5 | 3.52 | 53% | 3.50 / 3.54 | 53% / 54% | 3.02 | 44% | 14.32% | 3.58 | 66% | 3.58 / 3.56 | 66% / 65% | 3.37 | 64% | 5.87% | |||||||
| Qwen2-7b | 1 | 1.96 | 29% | 2.01 / 1.97 | 30% / 29% | 1.75 | 19% | 10.71% | 3.15 | 56% | 3.20 / 3.14 | 58% / 56% | 1.75 | 19% | 44.44% | ||||||
| 2 | 4.07 | 80% | 4.08 / 4.10 | 80% / 81% | 1.83 | 22% | 55.04% | 3.67 | 70% | 3.72 / 3.67 | 72% / 70% | 1.79 | 21% | 53.13% | |||||||
| 5 | 4.22 | 84% | 4.20 / 4.25 | 83% / 86% | 1.71 | 20% | 59.48% | 3.92 | 76% | 3.97 / 3.90 | 77% / 75% | 1.72 | 16% | 56.12% | |||||||