Revisiting the Energy Footprint of Small Language Models for Unit Test Generation

Follow-Up Study: Bayesian Analysis of Energy Consumption in SLM-Based Unit Test GenerationOverview This repository contains the datasets (.zip file) and analysis scripts (ipynb file) for the expanded Bayesian analysis conducted as part of the extended version of our SAST 2025 study on the energy consumption of automated unit test generation.In this follow-up study, we compare two Small Language Models (Phi 3.1 Mini 128k and Qwen2.5-Coder 1.5B) and contrast their energy costs with a traditional search-based test generation tool (Pynguin). The main goal is to investigate whether SLM-based approaches incur higher energy overhead than conventional tools. Methodological Extension: This expanded study adopts a formal Bayesian updating strategy. Specifically: Posterior distributions obtained from the original analysis (based on Phi) are incorporated as informative priors in the follow-up analysis involving Qwen.This approach propagates previously learned uncertainty into the new model, ensuring coherent probabilistic updating rather than independent re-estimation.