Back to selected work

Green AI decision support

GreenPEFT

A framework for comparing parameter-efficient fine-tuning choices across model quality and resource dimensions.

PythonPyTorchHugging Face TransformersPEFTscikit-learnCodeCarbonPyNVML

Problem

PEFT choices are difficult to compare when quality, VRAM, runtime, energy, carbon impact, and cost trade off differently.

Constraints

The implementation targets resource-constrained experimentation and reproducible model/resource evidence.

Approach

Compare Full Fine-Tuning, LoRA, QLoRA, LoRA-FA, and LISA. The research framework evaluates resource telemetry and uses Ridge surrogate regressors, Pareto-frontier extraction, and a Green Efficiency Index for constrained recommendations.

My contribution

I developed the project framing, experiment and artifact structure, resource-aware evaluation approach, and deployed application/package workflow.

Results

A live web application and installable green-peft package are available. The benchmark paper is in preparation and is not published, accepted, submitted, or under review.

What I would do next

Publish reconciled benchmark figures only after the corrected run, configuration, and feature counts are supplied and verified.

Artifacts & links