Pairwise preference comparison on HH-RLHF held-out (test)
53.02Win RateDP-RLHF
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DP-RLHFPrivacy budget (ε)=0.52026.03 | 53.02 | — | |
| DP-RLHFPrivacy budget (ε)=1.02026.03 | 52.85 | — | |
| DP-RLHFPrivacy budget (ε)=2.02026.03 | 52.82 | — | |
| DP-DPOPrivacy budget (ε)=2.02026.03 | 51.99 | — | |
| DP-DPOPrivacy budget (ε)=1.02026.03 | 51.92 | — | |
| DP-DPOPrivacy budget (ε)=0.52026.03 | 51.86 | — | |
| Private RMPrivacy budget (ε)=0.52026.03 | — | 58.93 | |
| Private RMPrivacy budget (ε)=1.02026.03 | — | 59.44 | |
| Private RMPrivacy budget (ε)=2.02026.03 | — | 59.69 |