Training Data Identification on WikiMIA alpha=0.1 1.0 (test)
0.21Power PerplexityVanilla
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VanillaBackbone Model=Pythia-6.9B2025.10 | 0.21 | 0.88 | 0.22 | |
| BKYBackbone Model=Pythia-6.9B2025.10 | 0.21 | 0.9 | 0.22 | |
| Quantile-BHBackbone Model=Pythia-6.9B2025.10 | 0.22 | 0.93 | 0.22 | |
| Storey-BHBackbone Model=Pythia-6.9B2025.10 | 0.41 | 1.95 | 0.4 | |
| VanillaBackbone Model=NeoX-20B2025.10 | 0.46 | 1.69 | 0.48 | |
| BKYBackbone Model=NeoX-20B2025.10 | 0.47 | 1.76 | 0.49 | |
| Quantile-BHBackbone Model=NeoX-20B2025.10 | 0.47 | 1.73 | 0.49 | |
| PTDI (Ours)Backbone Model=Pythia-6.9B2025.10 | 0.48 | 2.5 | 0.54 | |
| VanillaBackbone Model=LLaMA-7B2025.10 | 0.9 | 1.54 | 0.46 | |
| BKYBackbone Model=LLaMA-7B2025.10 | 0.92 | 1.58 | 0.47 | |
| Quantile-BHBackbone Model=LLaMA-7B2025.10 | 0.93 | 1.6 | 0.47 | |
| Storey-BHBackbone Model=NeoX-20B2025.10 | 1.07 | 3.54 | 0.88 | |
| PTDI (Ours)Backbone Model=NeoX-20B2025.10 | 1.65 | 6.05 | 1.63 | |
| Storey-BHBackbone Model=LLaMA-7B2025.10 | 1.94 | 3.16 | 0.88 | |
| PTDI (Ours)Backbone Model=LLaMA-7B2025.10 | 2.64 | 3.73 | 1.45 |