Green Tea: Simulation-Derived Energy Labels for Energy-Aware Code Generation

Green Tea provides per-program energy labels for C++ competitive-programming solutions, computed by deterministic architectural simulation (Sniper/McPAT) of an AMD EPYC 9554P core. It layers energy, cycle, instruction, power, and IPC measurements on top of the PIE corpus (derived from Project CodeNet), and provides energy-contrastive training pairs and a held-out benchmark for training and evaluating energy-aware code models. Contents: 3,507,435 deterministic energy simulations over 1,474 problems; 39,744 unique solutions with aggregated labels; 12,455 energy-contrastive pairs (9,927 train, 1,301 validation, 1,227 held-out benchmark) split at the problem level; and 15,853 reinforcement-learning prompts. Companion to the paper 'Beyond the Need for Speed: Energy-Aware Code Generation via Simulation-Guided Reinforcement Learning' (Rajput and Sharma). The energy measurements and derived pairs are released under CC BY 4.0; source solutions and test inputs derive from the PIE corpus and Project CodeNet and retain their original licenses.