Sampling/scoring models reversed in Fast-DetectGPT code

[SkyKingL]

Hi, thanks for the great repo!

In the paper, the authors describe the Fast-DetectGPT setup as follows:

“We chose the optimal settings reported by the authors, using GPT-Neo-2.7b as the scoring model and GPT-J-6b as the reference model.”

So Fast-DetectGPT should use GPT-J for sampling and GPT-Neo-2.7b for scoring.

However, in the current implementation, the roles appear to be reversed — GPT-J is used for scoring and GPT-Neo for sampling.