CIFAR-10 GreenXAI: A Multi-Architecture, Multi-Seed Dataset of Predictions, Uncertainty, Model Agreement, and Computational Sustainability

Abstract

CIFAR-10 GreenXAI is a derived machine-learning dataset developed to support research on predictive uncertainty, classification reliability, model agreement, reproducibility, and computational sustainability. The dataset contains 250,000 image–model–seed prediction records and 153 variables. It was generated by evaluating five deep learning architectures—ResNet18, DenseNet121, MobileNetV3-Small, EfficientNet-B0, and ViT-Tiny-CIFAR—using five independent random seeds (11, 22, 33, 44, and 55) on the 10,000 canonical CIFAR-10 test images. The dataset includes predicted classes, ten-class probability vectors, confidence scores, predictive entropy, negative log-likelihood, Brier scores, calibration descriptors, classification performance summaries, cross-model and cross-seed agreement, prediction stability, image difficulty, and high-confidence prediction errors. Computational metadata describe training, inference, and explainability activities, including execution time, energy consumption, estimated CO₂-equivalent emissions, and hardware and software information. Per-image and per-class energy and emissions allocations are derived estimates based on run-level quantities, rather than independently measured values. The repository provides the master CSV dataset, a README describing its structure and provenance, and a data dictionary documenting the 153 variables. The resource is intended to support secondary analysis of predictive behaviour, model variability, and computational sustainability without requiring researchers to repeat the complete model-training workflow. Original CIFAR-10 image files are not redistributed as part of this derivative dataset.