DetectRL
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
DetectRL Google-PaLM (train)
50.58TPR@FPR-1% (Llama-2-70b)
12
DetectRL Training Text: ChatGPT
50.66TPR@FPR-1% (Llama-2-70b)
12
DetectRL (Paraphrase)
80.4AUC@1%
10
DetectRL-arXiv cross-source corruption (test)
93.86AUROC
9
DetectRL Google-PaLM
77.8AUROC
6
DetectRL Llama-2-70b
0.812AUROC
6
DetectRL Word-level perturbation (OOD)
99.59AUROC
3
DetectRL Character-level perturbation (OOD)
99.8AUROC
3
DetectRL Yelp_review domain
90.18AUROC
3
DetectRL WritingPrompts domain
81.32AUROC
3
DetectRL Mixed-domain Source: GPT-3.5-turbo (test)
80.57AUROC
3
DetectRL Mixed-domain Source: Claude-instant (test)
80.92AUROC
3
DetectRL Mixed-domain Source: Llama-2-70B (test)
94.7AUROC
3
DetectRL Back-translation paraphrase (test)
99.79AUROC
3
DetectRL DIPPER paraphrase (test)
99.45AUROC
3
DetectRL Polish paraphrase (test)
99.21AUROC
3
DetectRL (test)
75.2Accuracy
3