Persona Alignment on Dignity Peer Persona Dimensions
15.2Dimension A ScoreMulti-Neg DPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Multi-Neg DPOModel Family=Qwen3-4B-Ins.2026.04 | 15.2 | 3.351 | 3.819 | 2.156 | |
| Lag-DPOModel Family=Qwen3-4B-Ins.2026.04 | 14.542 | 13.05 | 4.889 | 3.977 | |
| SFT-AnchorModel Family=Qwen3-4B-Ins.2026.04 | 7.172 | 5.291 | 6.232 | 0.84 | |
| MODPOModel Family=Qwen3-4B-Ins.2026.04 | 6.873 | 1.742 | 4.102 | 3.118 | |
| Lag-DPOModel Family=Llama-3-8B-Ins.2026.04 | 4.976 | 2.157 | 4.434 | 3.915 | |
| SACPOModel Family=Qwen3-4B-Ins.2026.04 | 4.814 | 6.678 | 3.739 | 3.082 | |
| Multi-Neg DPOModel Family=Llama-3-8B-Ins.2026.04 | 4.806 | 6.575 | 3.193 | 4.999 | |
| SafeRLHFModel Family=Qwen3-4B-Ins.2026.04 | 4.458 | 12.304 | 3.815 | 3.469 | |
| PCGrad-DPOModel Family=Llama-3-8B-Ins.2026.04 | 4.306 | 1.594 | 3.802 | 1.817 | |
| SafeRLHFModel Family=Llama-3-8B-Ins.2026.04 | 4.216 | 1.109 | 4.2 | 1.277 | |
| MODPOModel Family=Llama-3-8B-Ins.2026.04 | 3.214 | 2.481 | 0.912 | 2.746 | |
| PCGrad-DPOModel Family=Qwen3-4B-Ins.2026.04 | 3.019 | 4.417 | 4.517 | 2.281 | |
| SFT-CombinedModel Family=Qwen3-4B-Ins.2026.04 | 2.821 | 4.536 | 3.179 | 3.521 | |
| VanillaModel Family=Llama-3-8B-Ins.2026.04 | 2.782 | 2.524 | 2.939 | 0.476 | |
| SFT-CombinedModel Family=Llama-3-8B-Ins.2026.04 | 2.146 | 3.093 | 1.875 | 0.615 | |
| SFT-AnchorModel Family=Llama-3-8B-Ins.2026.04 | 2.024 | 1.949 | 2.301 | 2.029 | |
| VanillaModel Family=Qwen3-4B-Ins.2026.04 | 1.935 | 2.422 | 1.56 | 0.848 | |
| SACPOModel Family=Llama-3-8B-Ins.2026.04 | 0.812 | 0.89 | 3.648 | 3.356 |