Five Universities and Hugging Face Want to Fix How Publicly Funded Science Gets Shared

Key points

• Project ProtoStar, launched in late October 2026, is the first coordinated push by multiple U.S. universities to release publicly funded research in formats that AI systems can directly read and use.
• Founding university partners include Carnegie Mellon, Purdue, Georgia Tech, Arizona State, and George Washington University, with Hugging Face serving as the infrastructure host.
• The U.S. federal government spends approximately $190 billion a year on research and development, but much of the resulting data sits in formats that are hard to reuse.
• NIH's research repositories on Hugging Face have logged more than 11 million downloads as of October 26, 2026, showing demand already exists.
• Researchers uploaded more than three times as many new science datasets to Hugging Face in the first nine months of 2026 as in the same period of 2025.

There is a gap between science being public and science being usable. The U.S. government funds roughly $190 billion worth of research every year. Agencies like the National Institutes of Health and the National Science Foundation already require researchers to share their data. The problem is that sharing a file and making it usable are two different things.

Datasets end up scattered across incompatible systems. Code gets posted without instructions. Models arrive stripped of the context a newcomer needs to know what they are even for. An AI tool trying to pull from several of these sources at once would struggle the same way a person would: too many formats, too little explanation.

Project ProtoStar, announced by Hugging Face, wants to close that gap.

What does ProtoStar actually do?

Participating institutions sign a memorandum of understanding, a formal agreement, committing to release suitable scientific datasets, models, and code in AI-ready formats on open infrastructure. Hugging Face hosts the material. The Open Forum for AI coordinates the university coalition.

Carnegie Mellon's Open Source Programs Office has already begun encouraging its researchers to share outputs on Hugging Face. That kind of institutional policy is the mechanism here: not a technical fix, but an organisational nudge backed by a formal commitment.

The initiative also runs open challenges where anyone can try building with the released data. The first challenge, opening November 2, 2026 and running through November 13, asks teams to use historical weather records from 210 U.S. cities to predict climate conditions from 2019 to 2026. Four outputs matter: average annual temperature, total annual precipitation, the hottest day of the year, and cooling degree days, which is a measure of how hard air conditioning has to work. Participants also propose a short urban climate policy idea based on what they find.

That policy component is deliberate. ProtoStar wants to show that AI built on open scientific data can produce something practically useful, not just a benchmark score.

Why does documentation keep coming up?

Documentation matters more than it sounds. A 2024 study published on arxiv.org that analysed all 7,433 dataset records on Hugging Face at the time found that the section researchers most often skipped was "Considerations for Using the Data," which covers exactly the caveats and context a reuser needs most. The same research found that completion of dataset documentation correlated strongly with how popular the dataset was, meaning the datasets most people needed were better explained, but the long tail was thin.

ProtoStar's own figures point the same direction: 90 percent of dataset downloads on Hugging Face come from fully documented datasets. Leave out the documentation and almost nobody downloads the work.

AI2Day has tracked Hugging Face across 110 stories since July 2026, including a surge of enterprise-focused coverage. What shifts with ProtoStar is the audience: this is aimed at publicly funded academic science, not commercial AI teams, and the policy lever is institutional commitment rather than product adoption.

If the challenge model works, more universities are expected to join and host their own rounds. The practical evidence base ProtoStar says it wants to build, showing what barriers remain and what practices help, could feed directly into how funding agencies write their next data-sharing requirements.