NLP Watermarking on WaterBench & RepoBench-P (test)
1.1KoLA ScoreHard Watermark
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Hard WatermarkBackbone=Llama2-7B-Chat2025.11 | 1.1 | 8.9 | 10.5 | 13.6 | 4.9 | 27.8 | 25.9 | 11.1 | 12.2 | 1.1 | 95.6 | 99.5 | 11.7 | |
| V2 WatermarkBackbone=Llama2-7B-Chat2025.11 | 1.1 | 21.3 | 13.2 | 13.5 | 7.4 | 20.4 | 23.3 | 11.7 | 11.5 | 0.9 | 94.5 | 99.9 | 12.4 | |
| V2 WatermarkBackbone=InternLM-7B-Chat2025.11 | 1.3 | 20.6 | 9 | 6.3 | 3.4 | 28.1 | 27.3 | 5.3 | 5.6 | 0.5 | 94.9 | 99.8 | 10.7 | |
| GPT Watermark + WaterSearchBackbone=Llama2-7B-Chat2025.11 | 1.6 | 31 | 11.1 | 13.5 | 17.4 | 21.2 | 21.9 | 11 | 13.2 | 4.5 | 96.5 | 99 | 14.6 | |
| Soft WatermarkBackbone=Llama2-7B-Chat2025.11 | 1.7 | 13.8 | 8.1 | 11.8 | 14.4 | 25.3 | 22.2 | 9.3 | 11 | 0.6 | 95.3 | 99.5 | 11.8 | |
| GPT WatermarkBackbone=Llama2-7B-Chat2025.11 | 1.8 | 25.3 | 4.5 | 5.9 | 12.5 | 17 | 12.9 | 4.8 | 9.6 | 0.2 | 97 | 96.9 | 9.5 | |
| GPT WatermarkBackbone=InternLM-7B-Chat2025.11 | 1.9 | 4.5 | 8.5 | 7.1 | 2.4 | 20.5 | 19.4 | 4.2 | 6.2 | 0.5 | 95.6 | 99.8 | 7.5 | |
| V2 Watermark + WaterSearchBackbone=Llama2-7B-Chat2025.11 | 2.3 | 33.6 | 14.6 | 16.8 | 15.2 | 27.8 | 27.8 | 16.2 | 13.9 | 5 | 99.4 | 99.8 | 17.3 | |
| Hard Watermark + WaterSearchBackbone=Llama2-7B-Chat2025.11 | 2.4 | 28 | 13.7 | 16.4 | 8.6 | 29 | 27.2 | 15 | 14 | 4.5 | 98.7 | 99.5 | 15.9 | |
| Soft Watermark + WaterSearchBackbone=Llama2-7B-Chat2025.11 | 2.4 | 27.2 | 12.4 | 16 | 16.3 | 27.6 | 27.1 | 13.6 | 13.5 | 4.5 | 97.6 | 99.6 | 16.1 | |
| Soft WatermarkBackbone=InternLM-7B-Chat2025.11 | 2.4 | 10.1 | 9.1 | 6.1 | 2.5 | 18.6 | 17.7 | 4 | 5.3 | 0.3 | 94 | 99.6 | 7.6 | |
| Hard Watermark + WaterSearchBackbone=InternLM-7B-Chat2025.11 | 2.6 | 6.1 | 13.8 | 12.9 | 3.5 | 20.9 | 20.7 | 8.4 | 10.6 | 2 | 99.8 | 99.2 | 10.2 | |
| Hard WatermarkBackbone=InternLM-7B-Chat2025.11 | 2.8 | 0.8 | 10.7 | 8.4 | 3.2 | 20.1 | 15.8 | 5.3 | 7.4 | 0.8 | 93.3 | 99.7 | 7.5 | |
| Soft WatermarkBackbone=Qwen-2.5-7B-Instruct2025.11 | 3.1 | 41 | 16.5 | 14.1 | 36.1 | 15.2 | 17.4 | 12.2 | 14.4 | 6.4 | 94.8 | 99.7 | 17.6 | |
| GPT Watermark + WaterSearchBackbone=InternLM-7B-Chat2025.11 | 4.2 | 11 | 15.6 | 15.2 | 15.6 | 25.9 | 17.1 | 12.4 | 12.4 | 5.3 | 98.4 | 98.7 | 13.5 | |
| GPT WatermarkBackbone=Qwen-2.5-7B-Instruct2025.11 | 4.6 | 42 | 5.6 | 5.5 | 35.3 | 5 | 8.4 | 3.6 | 11.7 | 4.4 | 95.3 | 99.7 | 12.6 | |
| V2 Watermark + WaterSearchBackbone=InternLM-7B-Chat2025.11 | 4.7 | 33.5 | 15.3 | 15.4 | 16.7 | 28.8 | 33.3 | 12.5 | 11.5 | 4.5 | 95.3 | 99.6 | 17.6 | |
| Soft Watermark + WaterSearchBackbone=InternLM-7B-Chat2025.11 | 5 | 24.6 | 16.1 | 15.5 | 15.4 | 15.3 | 15.8 | 12.6 | 11.5 | 5.3 | 97.1 | 99.2 | 13.7 | |
| V2 WatermarkBackbone=Qwen-2.5-7B-Instruct2025.11 | 5.1 | 23.3 | 19.3 | 15.5 | 16.6 | 12.5 | 13.5 | 13.5 | 13.7 | 7.1 | 95.5 | 99.8 | 14 | |
| V2 Watermark + WaterSearchBackbone=Qwen-2.5-7B-Instruct2025.11 | 5.4 | 38 | 21.7 | 17.7 | 32 | 22 | 22.2 | 17.2 | 16.3 | 20.6 | 98.9 | 99.7 | 21.3 | |
| Hard WatermarkBackbone=Qwen-2.5-7B-Instruct2025.11 | 6.2 | 4.5 | 20 | 17.3 | 13.6 | 20.3 | 15.4 | 17.2 | 14.2 | 13.3 | 95.6 | 99.5 | 14.2 | |
| GPT Watermark + WaterSearchBackbone=Qwen-2.5-7B-Instruct2025.11 | 7.9 | 40.8 | 14.9 | 13 | 40 | 17.8 | 16.8 | 10.8 | 15.2 | 7.1 | 97.3 | 99.1 | 18.4 | |
| Soft Watermark + WaterSearchBackbone=Qwen-2.5-7B-Instruct2025.11 | 9.9 | 41 | 19.1 | 15 | 45.9 | 20.2 | 23.6 | 13.5 | 15.8 | 11.1 | 96.4 | 99.8 | 21.5 | |
| Hard Watermark + WaterSearchBackbone=Qwen-2.5-7B-Instruct2025.11 | 13.3 | 37.8 | 22.6 | 18.7 | 29.3 | 22.6 | 25.1 | 19.7 | 18 | 27.7 | 96.4 | 99.7 | 23.5 |