Why is Log-Rank linked to DetectGPT, what is the scoring model for those zero-shot detector and why do you use F1 score?
Jacobsonradical opened this issue · comments
Hi, thank you for sharing the code and brilliant works. I have 3 questions as stated in the title.
-
Why is Log-Rank method linked to DetectGPT? I don't think DetectGPT uses any rank information. It is more inclined to the probability method.
-
How do you do the zero-shot detection for ChatGPT-turbo and same kind? They don't have scoring function for those zero-shot methods. What is the scoring model here?
-
Why do you use F1 score? because we don't have good balance between positive and negative samples?
Hi @Jacobsonradical,
Thanks for your questions.
For Q1: Log-Rank is a baseline in DetectGPT's paper.
For Q2 & Q3: we do not do zero-shot detection based on metrics. For metric-based methods, we train an extra classifier to map the metric into classes (e.g., HWTs and MGTs). With this classifier, we can transfer the detection method to other datasets.
@xinleihe Great. Thank you very much.