Robustness against harmful content generation on LMSYS harmful queries
1Attack Success RateRLBF
Evaluation Results
| Method | Links | |
|---|---|---|
| RLBFModel Family=LLaMA 3, Model Size=1B2026.02 | 1 | |
| RLBFModel Family=LLaMA 3, Model Size=8B2026.02 | 1 | |
| RLBFModel Family=Gemma 2, Model Size=2B2026.02 | 2 | |
| RLBFModel Family=Gemma 2, Model Size=9B2026.02 | 2 | |
| RLBFModel Family=LLaMA 3, Model Size=3B2026.02 | 2 | |
| BSAFE+Model Family=Gemma 2, Model Size=2B2026.02 | 14 | |
| BSAFE+Model Family=LLaMA 3, Model Size=1B2026.02 | 14 | |
| BSAFE+Model Family=Gemma 2, Model Size=9B2026.02 | 15 | |
| BSAFE+Model Family=LLaMA 3, Model Size=8B2026.02 | 16 | |
| BSAFE+Model Family=LLaMA 3, Model Size=3B2026.02 | 17 | |
| RLModel Family=LLaMA 3, Model Size=1B2026.02 | 22 | |
| RLModel Family=Gemma 2, Model Size=2B2026.02 | 23 | |
| ITModel Family=LLaMA 3, Model Size=1B2026.02 | 24 | |
| RLModel Family=Gemma 2, Model Size=9B2026.02 | 24 | |
| ITModel Family=Gemma 2, Model Size=2B2026.02 | 25 | |
| RLModel Family=LLaMA 3, Model Size=3B2026.02 | 25 | |
| RLModel Family=LLaMA 3, Model Size=8B2026.02 | 25 | |
| ITModel Family=LLaMA 3, Model Size=8B2026.02 | 27 | |
| ITModel Family=Gemma 2, Model Size=9B2026.02 | 28 | |
| ITModel Family=LLaMA 3, Model Size=3B2026.02 | 28 |