Perplexity-based ChatGPT detector! It's a fork of Meta's LLaMA repo that returns inferenced probabilities for each token in a text instead of generating text.
Or as the app describes it:
When you ask an AI to write something for you, a neural network takes in the previous text (if any) and then, for every possible token (think of them as parts of a word), generates the probability of that particular token coming next. It then picks the most* probable token and does this process again until your whole text is generated.
The idea that the AI will always pick one of the most likely tokens is the basis for GPT-ish. By running a modified version of a GPT-like text generator, we can get the probability distributions for each token in the text and allow us to check if they're all just the most likely choices (like what an AI would do) or if there's more human-like variation.
*Actually, a technique called temperature sampling picks some less likely tokens at times to make your text more interesting.
You can see a demo at ben476/gpt-ish-solid.
In a conda env with pytorch / cuda available, run
pip install -r requirements.txt
Then in this repository
pip install -e .
The API is in web.py which is based on example.py from the original LLaMA repository.
python -m torch.distributed.run --nproc_per_node 1 example.py --ckpt_dir /local/scratch/hongbenj/llama/7B --tokenizer_path /local/scratch/hongbenj/llama/tokenizer.model
See the LICENSE file.