RLHF Alignment on HH-RLHF (held-out)
78Win RateCHI-REBEL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CHI-REBELBackbone=Llama-8B2025.09 | 78 | 19.7 | |
| KL-REBELBackbone=Llama-8B2025.09 | 77.3 | 20.5 | |
| CHI-REBELBackbone=Llama-1B2025.09 | 77.3 | 20.8 | |
| W-REBELBackbone=Llama-8B2025.09 | 76.9 | 20.9 | |
| KL-REBELBackbone=Llama-1B2025.09 | 76.4 | 21.6 | |
| W-REBELBackbone=Llama-1B2025.09 | 76.1 | 21.9 | |
| KL-DPOBackbone=Llama-8B2025.09 | 75.4 | 21.9 | |
| REBELBackbone=Llama-8B2025.09 | 75 | 22.2 | |
| WDPOBackbone=Llama-8B2025.09 | 74.2 | 22.9 | |
| KL-DPOBackbone=Llama-1B2025.09 | 74.1 | 23.7 | |
| REBELBackbone=Llama-1B2025.09 | 73.8 | 24 | |
| WDPOBackbone=Llama-1B2025.09 | 73.5 | 24.1 | |
| DPOBackbone=Llama-8B2025.09 | 72.9 | 23.7 | |
| DPOBackbone=Llama-1B2025.09 | 72.1 | 25.6 |