Finetuning Attack Success on Harmful HEx-PHI (adversarial)
84.92ASRSafeLM
Evaluation Results
| Method | Links | |
|---|---|---|
| SafeLMPost-Training Setting=Original2026.06 | 84.92 | |
| BaselinePost-Training Setting=Original2026.06 | 83.67 | |
| SRPPost-Training Setting=Original2026.06 | 80.5 | |
| SafeLMPost-Training Setting=+mixSFT2026.06 | 76.08 | |
| BaselinePost-Training Setting=+mixSFT2026.06 | 71.08 | |
| SRPPost-Training Setting=+mixSFT2026.06 | 68.33 | |
| BaselinePost-Training Setting=+Reflection2026.06 | 0.08 | |
| SafeLMPost-Training Setting=+Reflection2026.06 | 0 | |
| SRPPost-Training Setting=+Reflection2026.06 | 0 |