Texpose is a deep learning-based web application for detecting text generated by large language models. It classifies academic texts as "Human-written" or "AI-generated". This repository contains the Dockerized version of the project for easy deployment and inference using our final fine-tuned BERT model.
Texpose was developed as our final year project at NED University, Karachi. We created our dataset by collecting research paper abstracts from various domains pre-2017 (to ensure human authorship) and generating AI versions of each abstract using Gemini and Llama APIs. A link to the repository with the datasets and scripts is provided in the next section. We trained 3 RNN-based models (GRU, LSTM, BiLSTM) and a transformer-based model (BERT) on our dataset, and chose the fine-tuned BERT model (with the highest test accuracy of 95.7%) for our classification task after performance evaluation. The GRU-based model was trained on fine-tuned SciBERT embeddings and achieved the second best performance with a test accuracy of approximately 94%. We have provided a Google Drive link to the model in the next section. Additionally, we fine-tuned a secondary BERT model using the dataset of AI-generated abstracts to classify the LLM source of AI-generated text as "Gemini" or "Llama".