Distillation data detection on S1
95.3AUCToken Probability Deviation
Evaluation results
| Method | Links | |
|---|---|---|
| Token Probability DeviationModel=Mistral-7B2025.10 | 95.3 | |
| Token Probability DeviationModel=Gemma-7b2025.10 | 94.3 | |
| Token Probability DeviationModel=Llama-3.1-8B2025.10 | 92.7 | |
| TPDBackbone=Qwen2.5-32B-Instruct2025.10 | 91.8 | |
| TPDBackbone=Qwen2.5-14B-Instruct2025.10 | 87 | |
| TPDBackbone=Qwen2.5-7B-Instruct2025.10 | 85.5 | |
| Generated PerplexityBackbone=Qwen2.5-32B-Instruct2025.10 | 84.7 | |
| Generated MIN-KBackbone=Qwen2.5-32B-Instruct2025.10 | 84.7 | |
| Generated PerplexityBackbone=Qwen2.5-14B-Instruct2025.10 | 78.5 | |
| Generated MIN-KBackbone=Qwen2.5-14B-Instruct2025.10 | 78.5 | |
| Generated MIN-KBackbone=Qwen2.5-7B-Instruct2025.10 | 75.4 | |
| Generated PerplexityBackbone=Qwen2.5-7B-Instruct2025.10 | 75.3 | |
| Entropy-TempBackbone=Qwen2.5-32B-Instruct2025.10 | 70.3 | |
| Entropy-TempBackbone=Qwen2.5-14B-Instruct2025.10 | 58.1 | |
| Infilling ScoreBackbone=Qwen2.5-32B-Instruct2025.10 | 57.4 | |
| MIN-K%Model=Mistral-7B2025.10 | 56.4 | |
| MIN-K%++Model=Llama-3.1-8B2025.10 | 56.2 | |
| Infilling ScoreModel=Llama-3.1-8B2025.10 | 55.7 | |
| NeighborBackbone=Qwen2.5-32B-Instruct2025.10 | 55.5 | |
| NeighborModel=Gemma-7b2025.10 | 55.4 | |
| MIN-K%Model=Llama-3.1-8B2025.10 | 55.4 | |
| Entropy-TempBackbone=Qwen2.5-7B-Instruct2025.10 | 55.2 | |
| Infilling ScoreModel=Mistral-7B2025.10 | 55 | |
| PerplexityModel=Mistral-7B2025.10 | 54.9 | |
| ZlibModel=Mistral-7B2025.10 | 54.7 | |
| NeighborBackbone=Qwen2.5-14B-Instruct2025.10 | 54.3 | |
| MIN-K%++Model=Mistral-7B2025.10 | 54.3 | |
| NeighborBackbone=Qwen2.5-7B-Instruct2025.10 | 53.9 | |
| ZlibModel=Llama-3.1-8B2025.10 | 53.9 | |
| PerplexityModel=Gemma-7b2025.10 | 53.7 | |
| LowercaseModel=Gemma-7b2025.10 | 53.7 | |
| MIN-K%Model=Gemma-7b2025.10 | 53.5 | |
| Infilling ScoreBackbone=Qwen2.5-14B-Instruct2025.10 | 53.4 | |
| ZlibModel=Gemma-7b2025.10 | 53.3 | |
| MIN-K%++Model=Gemma-7b2025.10 | 53.2 | |
| Infilling ScoreBackbone=Qwen2.5-7B-Instruct2025.10 | 52.9 | |
| PerplexityModel=Llama-3.1-8B2025.10 | 52.9 | |
| LowercaseModel=Llama-3.1-8B2025.10 | 52.4 | |
| NeighborModel=Llama-3.1-8B2025.10 | 52.2 | |
| NeighborModel=Mistral-7B2025.10 | 52.2 | |
| Infilling ScoreModel=Gemma-7b2025.10 | 52.2 | |
| Entropy-NoiseBackbone=Qwen2.5-7B-Instruct2025.10 | 51.1 | |
| MIN-K%++Backbone=Qwen2.5-14B-Instruct2025.10 | 50.9 | |
| LowercaseModel=Mistral-7B2025.10 | 48.6 | |
| ZlibBackbone=Qwen2.5-7B-Instruct2025.10 | 47.4 | |
| MIN-K%++Backbone=Qwen2.5-7B-Instruct2025.10 | 47.2 | |
| Entropy-NoiseBackbone=Qwen2.5-14B-Instruct2025.10 | 47 | |
| LowercaseBackbone=Qwen2.5-14B-Instruct2025.10 | 46.7 | |
| ZlibBackbone=Qwen2.5-14B-Instruct2025.10 | 46.7 | |
| MIN-K%++Backbone=Qwen2.5-32B-Instruct2025.10 | 46.1 | |
| Entropy-NoiseBackbone=Qwen2.5-32B-Instruct2025.10 | 46 | |
| LowercaseBackbone=Qwen2.5-32B-Instruct2025.10 | 45.9 | |
| MIN-K%Backbone=Qwen2.5-14B-Instruct2025.10 | 45.3 | |
| PerplexityBackbone=Qwen2.5-14B-Instruct2025.10 | 44.9 | |
| ZlibBackbone=Qwen2.5-32B-Instruct2025.10 | 44.8 | |
| PerplexityBackbone=Qwen2.5-7B-Instruct2025.10 | 44.4 | |
| MIN-K%Backbone=Qwen2.5-7B-Instruct2025.10 | 44.3 | |
| Self-CritiqueBackbone=Qwen2.5-32B-Instruct2025.10 | 44 | |
| MIN-K%Backbone=Qwen2.5-32B-Instruct2025.10 | 43.7 | |
| Self-CritiqueBackbone=Qwen2.5-7B-Instruct2025.10 | 43.6 | |
| LowercaseBackbone=Qwen2.5-7B-Instruct2025.10 | 43.5 | |
| PerplexityBackbone=Qwen2.5-32B-Instruct2025.10 | 43.3 | |
| Self-CritiqueBackbone=Qwen2.5-14B-Instruct2025.10 | 40.4 |