Watermark Detectability on C4 RealNewsLike Dipper-2 (test)
85.9AUCMarkTune
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MarkTuneBackbone=Llama2-7B, Temperature=1.02025.12 | 85.9 | 27 | |
| MarkTuneBackbone=Qwen3-4B, Temperature=0.72025.12 | 85.1 | 25.4 | |
| GaussMarkBackbone=Qwen3-4B, Temperature=0.72025.12 | 79.5 | 13.8 | |
| Gumbel-maxBackbone=Qwen3-4B, Temperature=0.72025.12 | 72.1 | 14.6 | |
| Gumbel-maxBackbone=Llama2-7B, Temperature=1.02025.12 | 71.1 | 14.8 | |
| SynthIDBackbone=Llama2-7B, Temperature=1.02025.12 | 71 | 8.4 | |
| SynthIDBackbone=Qwen3-4B, Temperature=0.72025.12 | 70 | 7.4 | |
| GaussMarkBackbone=Llama2-7B, Temperature=1.02025.12 | 68.7 | 6 | |
| KGW-2Backbone=Qwen3-4B, Temperature=0.72025.12 | 61.6 | 3.8 | |
| KGW-2Backbone=Llama2-7B, Temperature=1.02025.12 | 61.4 | 3.2 | |
| KGW-DBackbone=Llama2-7B, Temperature=1.02025.12 | 59 | 2.6 | |
| KGW-DBackbone=Qwen3-4B, Temperature=0.72025.12 | 58.9 | 3 | |
| KGW-1Backbone=Qwen3-4B, Temperature=0.72025.12 | 57.3 | 2.6 | |
| KGW-1Backbone=Llama2-7B, Temperature=1.02025.12 | 55.1 | 1.4 |