RaisonText is an open-source AI library for synthetic (e. g. chatGPT-generated) text detection.
We collected and generated 941k samples of human-written (33%) and AI-generated (67%) texts.
Public datasets used:
Generative models used:
- Chat-GPT
- opt-125m
- opt-1.3b
- opt-2.7b
- llama2-7b
- llama2-13b
Source name saved in source column.
Dataset may be downloaded here.
Part of generated data publiched on HuggingFace.
In general the model and the backend are two independent instances, where each of them runs independently. They only interact through RabbitMQ queues.
pip install -r requirements.txt
cd raisontext
The server part sends user requests to the queue and listens to the model's queue, so when a forecast is made for a particular user, it transmits it to the frontend.
uvicorn main:app
The model listens to the backend queue, makes a prediction, and then sends it to the model's queue.
python logreg_baseline.py
Model weights could be found on Google Drive by this link
We reached 0.96 roc_auc score with baseline model. Exract archive into repo root folder models.