Ownership Verification on Llama-2-7B SFT & RLHF
2FSR (Anchor)Base+Non-Homologous Model
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Base+Non-Homologous ModelJoint optimization strategy=control group, Reference Model=EvoLLM, Task Combination=Base+Non-Homologous Model2025.05 | 2 | 0 | 0 | 0 | 2 | 0 | 2 | 0 | 0 | 0 | |
| Base (32)Total tunable tokens=32, Task Combination=Base (32)2025.05 | 54 | 0 | 14 | 18 | 6 | 10 | 2 | 20 | 56 | 54 | |
| Base (64)Total tunable tokens=64, Task Combination=Base (64)2025.05 | 78 | 2 | 14 | 30 | 8 | 0 | 2 | 34 | 80 | 84 | |
| Base+Homologous ModelJoint optimization strategy=incorporating homologous models, Task Combination=Base+Homologous Model2025.05 | 83 | 6 | 31 | 36 | 23 | 2 | 4 | 46 | 82 | 86 | |
| Base+Chat TemplateJoint optimization strategy=incorporating multiple prompt templates, Task Combination=Base+Chat Template2025.05 | 90 | 4 | 34 | 34 | 18 | 0 | 1 | 54 | 91 | 92 | |
| Base+S2.7-P+[D+Z]Joint optimization strategy=synergistic joint optimization, Model=Sheared-Llama-2.7B-Pruned, Prompt templates=Default and Zero-Shot, Task Combination=Base+S2.7-P+[D+Z]2025.05 | 98 | 16 | 56 | 58 | 60 | 12 | 4 | 82 | 96 | 96 |