Pre-training data detection on BookTection (test)
0.98AUROCGDS
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GDSTarget Model=OPT-6.7b2026.03 | 0.98 | 92 | |
| GDSTarget Model=LLaMA-7B2026.03 | 0.98 | 92 | |
| GDSTarget Model=GPT-J-6b2026.03 | 0.97 | 88 | |
| FSDTarget Model=OPT-6.7b2026.03 | 0.96 | 77 | |
| GDSTarget Model=Neo-2.7B2026.03 | 0.96 | 84 | |
| GDSTarget Model=Pythia-6.9b2026.03 | 0.96 | 83 | |
| FSDTarget Model=Pythia-6.9b2026.03 | 0.93 | 59 | |
| FSDTarget Model=Neo-2.7B2026.03 | 0.92 | 53 | |
| FSDTarget Model=LLaMA-7B2026.03 | 0.92 | 55 | |
| FSDTarget Model=GPT-J-6b2026.03 | 0.91 | 52 | |
| Min-kTarget Model=GPT-J-6b2026.03 | 0.75 | 27 | |
| PPLTarget Model=GPT-J-6b2026.03 | 0.74 | 25 | |
| Min-kTarget Model=Pythia-6.9b2026.03 | 0.74 | 26 | |
| PPLTarget Model=Pythia-6.9b2026.03 | 0.73 | 25 | |
| PPLTarget Model=LLaMA-7B2026.03 | 0.71 | 25 | |
| Min-kTarget Model=Neo-2.7B2026.03 | 0.71 | 18 | |
| Min-kTarget Model=LLaMA-7B2026.03 | 0.71 | 25 | |
| PPLTarget Model=Neo-2.7B2026.03 | 0.69 | 15 | |
| Min-kTarget Model=OPT-6.7b2026.03 | 0.67 | 15 | |
| Min-k++Target Model=GPT-J-6b2026.03 | 0.67 | 18 | |
| Min-k++Target Model=Pythia-6.9b2026.03 | 0.66 | 19 | |
| PPLTarget Model=OPT-6.7b2026.03 | 0.64 | 13 | |
| Min-k++Target Model=Neo-2.7B2026.03 | 0.64 | 13 | |
| Min-k++Target Model=LLaMA-7B2026.03 | 0.63 | 15 | |
| Min-k++Target Model=OPT-6.7b2026.03 | 0.61 | 11 | |
| ZLibTarget Model=GPT-J-6b2026.03 | 0.58 | 20 | |
| ZLibTarget Model=Pythia-6.9b2026.03 | 0.58 | 19 | |
| ZLibTarget Model=LLaMA-7B2026.03 | 0.57 | 19 | |
| ZLibTarget Model=Neo-2.7B2026.03 | 0.56 | 16 | |
| ZLibTarget Model=OPT-6.7b2026.03 | 0.55 | 14 |