Evaluation Dataset: Towards Green Hyperparameter Optimization: An LLM-based Agent for Energy-Aware and Multi-Objective AutoML

The data reflect the results of the experimentation with an energy-aware Hyperparameter Optimization (HPO) framework governed by an LLM-based agent (gpt-oss:20b) built on Ray Tune. Three optimization strategies—Random Search, a deterministic Fixed Agent, and the proposed LLM Agent—were evaluated across Computer Vision (CIFAR-10), Natural Language Processing (SST-2), and Time Series Forecasting (ETTm1) tasks. Metrics assessed included predictive performance, operational energy consumption (Wh), CO2e emissions, and a multi-objective Balanced Score. This allows for the analysis of energy–performance trade-offs, Pareto dominance among strategies, and the environmental efficiency of LLM-based governance relative to random and rule-based baselines.