Reward Modeling on RPR n=5 (test)
81.2User FriendlinessDRM
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DRMBase Model=Llama3-8B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 81.2 | 94.6 | 94.5 | 96.9 | 99.1 | 93.1 | |
| DRMBase Model=Gemma-2B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 78.9 | 87.1 | 95.3 | 90.7 | 97.5 | 90 | |
| Random (Gaussian)Base Model=Llama3-8B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 73 | 89.2 | 88.7 | 95.6 | 91.9 | 87.5 | |
| Random (Uniform)Base Model=Gemma-2B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 71.3 | 78.2 | 92 | 90.7 | 90.7 | 84.8 | |
| Single HeadBase Model=Llama3-8B-RM, Adaptation Method=None2025.02 | 68.5 | 82.5 | 87.5 | 94 | 90.5 | 84.4 | |
| Shared-BaseBase Model=Llama3-8B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 67.4 | 82.5 | 82.7 | 96.4 | 88.1 | 83.2 | |
| Random (Uniform)Base Model=Llama3-8B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 61.6 | 86 | 79.8 | 95.8 | 90.6 | 82.3 | |
| Random (Gaussian)Base Model=Gemma-2B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 58.2 | 76 | 82.3 | 87.3 | 81.7 | 77.1 | |
| Shared-BaseBase Model=Gemma-2B-RM, Adaptation Method=HyRe, Adaptation Samples (n)=52025.02 | 51.7 | 68.8 | 81.7 | 88.1 | 69 | 72.1 | |
| Single HeadBase Model=Gemma-2B-RM, Adaptation Method=None2025.02 | 50.6 | 66.2 | 81.7 | 88.1 | 69 | 71.4 |