Best-of-N Alignment on PKU-SafeRLHF
38Percent batches with BWR > 0.50AdaBoN
Evaluation Results
| Method | Links | |
|---|---|---|
| AdaBoNLM=Qwen, RM=Armo, K=5, B=120, d=0.75B2025.05 | 38 | |
| AdaBoNLM=Gemma, RM=Mistral, K=5, B=120, d=0.75B2025.05 | 74 | |
| AdaBoNLM=Llama, RM=Mistral, K=5, B=120, d=0.75B2025.05 | 78 | |
| AdaBoNLM=Qwen, RM=FsfairX, K=5, B=120, d=0.75B2025.05 | 80 | |
| AdaBoNLM=Mistral, RM=Armo, K=5, B=120, d=0.75B2025.05 | 88 | |
| AdaBoNLM=Qwen, RM=Mistral, K=5, B=120, d=0.75B2025.05 | 88 | |
| AdaBoNLM=Gemma, RM=Armo, K=5, B=120, d=0.75B2025.05 | 90 | |
| AdaBoNLM=Mistral, RM=FsfairX, K=5, B=120, d=0.75B2025.05 | 92 | |
| AdaBoNLM=Llama, RM=FsfairX, K=5, B=120, d=0.75B2025.05 | 94 | |
| AdaBoNLM=Mistral, RM=Mistral, K=5, B=120, d=0.75B2025.05 | 96 | |
| AdaBoNLM=Gemma, RM=FsfairX, K=5, B=120, d=0.75B2025.05 | 96 | |
| AdaBoNLM=Llama, RM=Armo, K=5, B=120, d=0.75B2025.05 | 98 |