LLM alignment on WildChat (train)
1.1951LossPoLO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| PoLOBase Model=LLaMA-3.1-8B, Watermark size (bit)=512, |S| (%)=37.502025.05 | 1.1951 | 81.18 | 2.29 | 0.89 | |
| PoLOBase Model=LLaMA-3.1-8B, Watermark size (bit)=1024, |S| (%)=33.332025.05 | 1.2273 | 85.3 | 3.91 | 0.88 | |
| GD-PoLBase Model=LLaMA-3.1-8B2025.05 | 1.2511 | 76.85 | 1.14 | 77.06 | |
| PoLOBase Model=GPT-Neo-1.3B, Watermark size (bit)=512, |S| (%)=42.862025.05 | 1.5004 | 14.37 | 0.53 | 0.13 | |
| GD-PoLBase Model=GPT-Neo-1.3B2025.05 | 1.5017 | 14.18 | 0.21 | 14.77 | |
| PoLOBase Model=GPT-Neo-1.3B, Watermark size (bit)=1024, |S| (%)=42.862025.05 | 1.5021 | 14.98 | 0.38 | 0.13 |