Safety Evaluation on Harmful Benchmarks (CATQA, HEX-PHI, Salad-Base)
99.94CATQA ScoreSafe LORA
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Safe LORAFine-tuned on=CodeAlpaca-20k2024.12 | 99.94 | 99.7 | 99.91 | 99.85 | 0.1 | — | |
| Safe LoRAfine-tuning dataset=Alpaca Chinese2024.12 | 99.88 | 100 | 99.86 | 99.91 | — | 0.04 | |
| Safe LoRAFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 99.88 | 100 | 99.86 | 99.91 | 0.04 | — | |
| RESTAFine-tuned on=CodeAlpaca-20k2024.12 | 99.7 | 96.97 | 99.62 | 98.76 | 1.19 | — | |
| IRRdFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 99.7 | 99.7 | 99.77 | 99.72 | 0.23 | — | |
| RESTAdFine-tuned on=CodeAlpaca-20k2024.12 | 99.58 | 96.67 | 99.67 | 98.64 | 1.31 | — | |
| IRRdFine-tuned on=CodeAlpaca-20k2024.12 | 99.58 | 98.48 | 99.81 | 99.29 | 0.66 | — | |
| IRRFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 99.58 | 99.39 | 99.72 | 99.56 | 0.39 | — | |
| RESTAdFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 99.52 | 98.79 | 99.2 | 99.17 | 0.78 | — | |
| RESTAFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 99.33 | 99.09 | 99.34 | 99.26 | 0.69 | — | |
| IRRFine-tuned on=CodeAlpaca-20k2024.12 | 99.27 | 97.88 | 99.72 | 98.96 | 0.99 | — | |
| RESTAdfine-tuning dataset=Alpaca Chinese2024.12 | 99.03 | 98.18 | 99.48 | 98.9 | — | 1.05 | |
| RESTAfine-tuning dataset=Alpaca Chinese2024.12 | 98.91 | 98.18 | 99.39 | 98.83 | — | 1.12 | |
| IRRfine-tuning dataset=Alpaca Chinese2024.12 | 98.91 | 98.18 | 99.39 | 98.83 | — | 1.12 | |
| IRRdfine-tuning dataset=Alpaca Chinese2024.12 | 98.85 | 99.39 | 99.58 | 99.27 | — | 0.68 | |
| SafeDecodingfine-tuning dataset=Alpaca Chinese2024.12 | 96.61 | 91.21 | 97.7 | 95.17 | — | 4.78 | |
| SafeDecodingFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 93.52 | 95.45 | 96.9 | 95.29 | 4.66 | — | |
| SafeDecodingFine-tuned on=CodeAlpaca-20k2024.12 | 92.73 | 86.06 | 96.72 | 91.83 | 8.12 | — | |
| SFTfine-tuning dataset=Alpaca Chinese2024.12 | 89.09 | 85.76 | 95.31 | 90.05 | — | 0.9 | |
| DAREfine-tuning dataset=Alpaca Chinese2024.12 | 88.79 | 85.45 | 95.22 | 89.82 | — | 10.13 | |
| DAREFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 78.97 | 71.21 | 90.43 | 80.2 | 19.75 | — | |
| SFTFine-tuned on=GSM8K, Safety Realignment=true2024.12 | 78.85 | 71.52 | 90.01 | 80.12 | 19.83 | — | |
| SFTFine-tuned on=CodeAlpaca-20k2024.12 | 64.48 | 62.42 | 83.07 | 69.99 | 29.96 | — | |
| DAREFine-tuned on=CodeAlpaca-20k2024.12 | 64.18 | 63.03 | 83.26 | 70.16 | 29.79 | — |