Open-Sourced GPTZero
We've open-sourced the implementation of GPTZero!
How does it work?
-
It uses GPTv2 Perplexity Model from Huggingface to find total perplexity of the sentence.
-
Later, we break it into individual sentences based on some delimiters, and find individual perplexity per line and find total average perplexities.
-
The average perplexities is then passed through certain threshold, and if it crosses it, it becomes human-centric, otherwise the text is generated by an AI.
-
You can find the implementation on GitHub here [https://github.com/BurhanUlTayyab/GPTZero].
-
Play around with it with a website here [https://gptzero.sg].
-
How we implement, watch here [https://www.youtube.com/watch?v=x9H-aY5sCDA].
-
Join our Discord community for discussion and feedback [https://discord.com/invite/F3kFan28vH].
See you there~
Powered by Ailiverse
I have tried GPT every time i find different answers
Awesome work!
Cool! How accurate is it compared to other GPT detectors?
The implementation here is a GPTv2 Perplexity Model being run on individual sentences. However we’re working on a version 2 where two large LLMs fight in a zero sum game (kinda like a GAN).
The accuracy is good for most of the examples we’ve tried on.
Brace youself folks. Thousands of students will be flocking to indiehackers because they were just expelled for using chatgpt 🤣
It’s a cool classifier 😂
This is super exciting and I can't wait to try it out! Have you tried this on various generative LLMs outside of ChatGPT or is it specifically tailored to ChatGPT?
(p.s. I would love to feature this on the Practical AI Podcast, which I co-host, if you are interested)
We haven’t tested it outside ChatGPT however, we’re working on a generic model which would be published soon :)
Hey definitely interested! Will PM you on Twitter!
This comment was deleted 3 years ago.
Hihi thanks for trying it out!
Yes this version is beta version. We are looking for the potential limitations. If you could share what you have wrote there, this will be helpful for us to improve it!