GenAI-GreenML: A Curated Dataset for Generative AI and Green Machine Learning

A curated dataset containing 50 small open-source machine learning repositories, specifically designed for research on generative AI-assisted code generation and energy-efficient machine learning development.

• Original Author/Maintainer: Bearwick
• Source Platform: GitHub
• Original Title: GenAI-GreenML
• Original Link: https://github.com/Bearwick/GenAI-GreenML
• Publication Date: 2026-06-09

Generative AI is reshaping all aspects of software development, from code completion to automated testing, from document generation to architecture design. However, behind this convenience lies an increasingly serious issue: the environmental cost of AI-assisted programming.

Large Language Models (LLMs) consume a great deal of energy during training and inference, resulting in significant carbon emissions. At the same time, is code generated by LLMs more energy-efficient than manually written code? Do generated machine learning models consider energy efficiency optimization? These questions currently lack systematic research data support.

GenAI-GreenML dataset was created to fill this research gap. It provides a carefully selected benchmark dataset specifically for evaluating the environmental impact and energy efficiency performance of generative AI in code generation tasks.

GenAI-GreenML is a curated collection of 50 small open-source machine learning repositories, all of which are under 500 MB in size, covering two major domains: Tabular data and Natural Language Processing (NLP).

1. Small-Scale Priority: Select repositories smaller than 500 MB to lower the computational resource threshold for experiments, enabling more researchers to reproduce and extend studies.

2. Domain Representativeness: Cover the two core ML domains of tabular data processing and NLP to ensure the generalizability of research conclusions.

3. Open-Source License: All included projects use open-source licenses, supporting academic and commercial research use.

4. Practicality Orientation: Select projects with real-world application scenarios rather than purely academic research code.

This dataset provides a standardized testing platform for evaluating the code generation capabilities of different LLMs (GPT-4, Claude, Llama, etc.):

• Functional Correctness: Does the generated code correctly implement the intended function?
• Code Quality: How is the readability, maintainability, and completeness of comments of the generated code?
• Security Vulnerabilities: Does the generated code contain common security vulnerabilities?

By comparing the energy efficiency performance of manually written code and LLM-generated code, researchers can:

• Identify the advantages and limitations of LLMs in energy efficiency optimization
• Develop prompt engineering strategies to guide LLMs to generate more energy-efficient code
• Establish best practice guidelines for green AI coding

Provide empirical data for researchers in the field of software engineering to explore:

• The long-term impact of AI-assisted development on software carbon footprint
• Environmental cost-benefit analysis of code generation tools
• Evolution trends of green programming paradigms