Small Language Models for perplexity-based text classification

Abstract

The research explores the use of Small Language Models (SLMs), specifically those with 1B parameters or fewer, to classify text as either AI-generated or human-written in the Ukrainian language. By leveraging perplexity – a measure of a model's uncertainty – as a key feature, the study evaluates the performance of models like Gemma 3 and Llama 3.2. The authors further propose a classification method using Convolutional Neural Networks (CNN) trained on token-level probability vectors, achieving accuracy rates up to 0.8559.