Reporting Carbon Emissions on Open-Source Model Cards

Training AI models takes a substantial amount of energy, which means they emit a substantial amount of carbon dioxide. One study out of the University of Massachusetts, Amherst found that training a large and common AI model can create five times the lifetime emissions of an American car. 

Hugging Face has worked to normalize reporting AI model carbon emissions by making them an optional component to model cards that live on the open-source Hugging Face Hub. Hugging Face model cards describe the AI model, its uses, its limitations, its biases, its ethical considerations, training parameters, datasets, and evaluation results. This information lives in a Markdown file, and users of the hub can discover, reproduce, and share models with more ease thanks to model cards. 

The hub allows users to filter model cards by ones that report carbon emissions, in a sense rewarding those model cards with more visibility. In documentation, Hugging Face provides the metadata structure to copy and paste into the model card’s Markdown file. The carbon emissions reporting information includes the geographic location where the model was trained, the hardware used, training type (pre-training or fine-tuning), and the model’s estimated carbon footprint. Hugging Face recommends using either the Code Carbon package to estimate emissions in real-time or the ML CO2 Calculator to measure emissions after training. 

As we continue to tackle to climate crisis, open-source communities like the one Hugging Face fosters allow us to learn from each other and engage in productive discourse. Broader reporting of AI models’ carbon emissions can help us better understand how we can produce more energy efficient AI models.

Adding carbon emissions information to Hugging Face's model cards takes little time or technical know-how. This can partially be attributed to the clear and concise documentation provided. Reporting carbon emissions in the Hugging Face Hub also gives one's model more visibility, because users can filter for model cards with that information.

Since this information is self-reported and inputted into a static Markdown file, Hugging Face isn't able to detect errors one may make. This impacts both the original model developers who want to accurately understand their carbon footprint and users of the Hugging Face Hub.

Use Hugging Face's AutoTrain — a tool to train, evaluate, and deploy models without code — to automatically track carbon emissions.

This article does not constitute legal or other professional advice and was written by students as part of the Duke Ethical Technology Practicum.