Watermark Detection on HumanEval (TPR/F1 @ 1% & 5% FPR)
59.1TPR @ 1% FPREWD + SSG
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| EWD + SSGModel=LLaMA-3-8B2026.04 | 59.1 | 74 | 73.2 | 82.2 | |
| EWD + SSGModel=Qwen2.5-coder-7B2026.04 | 50.6 | 66.9 | 70.7 | 80.6 | |
| SWEET + SSGModel=Qwen2.5-coder-7B2026.04 | 50 | 66.4 | 78 | 84.2 | |
| KGW + SSGModel=Qwen2.5-coder-7B2026.04 | 39 | 55.9 | 45.7 | 60.7 | |
| EWDModel=Qwen2.5-coder-7B2026.04 | 32.9 | 49.3 | 52.4 | 66.7 | |
| SWEETModel=Qwen2.5-coder-7B2026.04 | 29.3 | 45.1 | 36.1 | 66.1 | |
| EWDModel=LLaMA-3-8B2026.04 | 26.8 | 42.1 | 54.9 | 68.7 | |
| SWEET + SSGModel=LLaMA-3-8B2026.04 | 24.4 | 39 | 37.8 | 53 | |
| KGWModel=Qwen2.5-coder-7B2026.04 | 22 | 35.8 | 36 | 51.3 | |
| KGW + SSGModel=LLaMA-3-8B2026.04 | 14 | 24.5 | 49.4 | 64 | |
| KGWModel=LLaMA-3-8B2026.04 | 13.4 | 23.5 | 22.6 | 35.4 | |
| UWModel=LLaMA-3-8B2026.04 | 8.5 | 15.6 | 15.2 | 25.4 | |
| SWEETModel=LLaMA-3-8B2026.04 | 7.9 | 14.6 | 26.2 | 40 | |
| UWModel=Qwen2.5-coder-7B2026.04 | 0.6 | 1.2 | 3.7 | 6.7 |