Butternut Extension Backend

Butternut works by running text-generation models over text to find out what words the model would've come up with and then comparing it with the actual word. AI-generated text will be more consistent with the predictions of the model while human-written text is less so. Based off work in the GLTR paper (GLTR Github)

Usage

As a simple API:

cd src/
export FLASK_APP=api
flask run

POST request with desired text to text formfield. Returns JSON encoded dict of format:

{
    'bpe_strings': ['token1', 'token2', ...], //tokens from input text
    'real_topk' : [ (<rank1: int>, <prob%1:int>),... ] // probablity ranking output
}

Or as a python module:

import gltr
gltr = gltr.LM()
raw_text = "some sentence here"
payload = gltr.check_probabilities(raw_text)
tokens = payload['bpe_strings']
probabilities = payload['real_topk']

Desired backend can be specified in src/config.toml Currently only CTRL & gpt-2 backends are supported.