American taxpayers fund roughly $190 billion of research and development a year, and federal agencies increasingly require the resulting papers and data to be made public. Public is not the same as usable. Datasets sit in old formats on scattered servers, often without the documentation a researcher or an AI system needs to understand them. Project ProtoStar, announced on Hugging Face's blog, is an attempt to close that gap, and it matters to anyone who trains or fine-tunes models on scientific data.
The founding partners are five universities, Carnegie Mellon, Purdue, Georgia Tech, Arizona State and George Washington, working through the Open Forum for AI coalition, along with the Astera Institute, the Open Research Community Accelerator as community partner, and Hugging Face as infrastructure partner. Members sign a memorandum of understanding committing them to release suitable datasets, models and code in machine-readable formats on open infrastructure. Carnegie Mellon has gone furthest, with a policy encouraging its researchers to publish on Hugging Face while still using specialist repositories where those fit better.
The project's pitch rests on usage figures from the Hub. Hugging Face says it hosts more than 17,000 public science datasets in fields such as genomics, chemistry, climate and astronomy, close to 29,000 once medicine and mathematics are counted, and that more than three times as many new science datasets were shared in the first nine months of 2026 as in the same period of 2025. The 18 repositories of the US National Center for Biotechnology Information have passed 11 million downloads. Hugging Face also reports that 90% of dataset downloads go to fully documented datasets, which is the strongest argument for the paperwork ProtoStar asks for.
The first test is a challenge hosted by Carnegie Mellon from 2 to 13 November. Teams get historical weather for 210 US cities up to 2018 and must forecast four measures for 2019 to 2026, including average temperature, rainfall, the hottest day of the year and cooling demand. Entries are scored against simple baselines such as historical averages and straight-line trends, and teams also propose a short climate adaptation policy. Final results come in early 2027, once the full 2026 records exist. Later challenges will be hosted by the other universities in turn.
There is an obvious interest at work: Hugging Face, which we review as the main hub for open models and datasets, gains if public science is published on its platform. The commitments are also soft. A memorandum binds nobody to a budget, and a two-week forecasting contest will show whether the data can be reused, not whether universities change how they publish. The useful measure will be whether partner institutions are still releasing documented datasets a year from now.