Estimating Carbon Emissions of HuggingFace AI Models

Thijs Nulle, Harmen Kroon, Petter Reijalt.

Paper.Source code.

We scrutinise the environmental impact of training AI models, focusing on carbon emissions of HuggingFace AI models as a case study. It reveals disparities between emissions of self-reported and models trained with Autotrain, advocating for standardised reporting. A Firefox extension is proposed for estimating emissions based on model domain and dataset size. It emphasises balancing model efficiency with environmental concerns and suggests improving dataset quality and to advocate for transparent reporting practices.