Dialogue on HH (Anthropic Helpful and Harmless)
82.5Win Rate (0% Flip)α-URM
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| α-URMModels=Llama-3.2-3B2026.07 | 82.5 | 81.4 | 74.3 | 79.9 | |
| URMModels=Llama-3.2-3B2026.07 | 81.4 | 78.9 | 72.6 | 79.4 | |
| cRMModels=Llama-3.2-3B2026.07 | 80.7 | 76.1 | 70.1 | 75.7 | |
| RMModels=Llama-3.2-3B2026.07 | 80 | 78.3 | 69.7 | 75.5 | |
| rRMModels=Llama-3.2-3B2026.07 | 79.9 | 78.3 | 72.4 | 75.7 | |
| URMModels=Qwen-3-1.7B2026.07 | 67.6 | 61.8 | 58.2 | 70.1 | |
| α-URMModels=Qwen-3-1.7B2026.07 | 66.9 | 61.6 | 57.9 | 69.6 | |
| rRMModels=Qwen-3-1.7B2026.07 | 65.7 | 62.5 | 52.1 | 65.9 | |
| cRMModels=Qwen-3-1.7B2026.07 | 64.2 | 56 | 46.2 | 65 | |
| RMModels=Qwen-3-1.7B2026.07 | 63.4 | 55.5 | 43.9 | 63.9 |