Hugging Face and five US universities have launched Project ProtoStar, an initiative to make datasets, models and code from publicly funded science easier for researchers and AI systems to find and reuse. Its first open challenge, built on climate records for 210 US cities, opens on 2 November.
Hugging Face announced the project on 29 October in a blog post. The founding partners are Carnegie Mellon University, Purdue University, Georgia Tech, Arizona State University and George Washington University, which take part through the Open Forum for AI, together with the Astera Institute. Hugging Face is the infrastructure partner and the Open Research Community Accelerator (ORCA) is the community partner. George Washington University confirmed its role in a release the same day.
The problem it targets
Hugging Face says the US federal government spends about $190 billion a year on research and development, and that agencies such as the National Institutes of Health and the National Science Foundation increasingly require the results and underlying data to be shared. But public is not the same as usable. Data often sits in old formats, on incompatible systems or without the documentation needed to reuse it.
“Public access is a good first step. Usability is next,” said Lorena Barba, faculty director of GW’s Open Source Program Office, in the university’s release.
Partners sign a memorandum of understanding to release suitable datasets, models and code in AI-ready formats on open infrastructure. The project page describes these as non-binding and non-exclusive, and says the signed agreements are public. GW says it will set up a presence on Hugging Face for its research and help its researchers use it.
Hugging Face’s figures
Hugging Face, which is one of the generalist repositories NIH lists for sharing research data, gives several numbers to show demand. It says its Hub hosts more than 17,000 public datasets in core science fields, close to 29,000 when medicine and maths are included. It says researchers shared more than three times as many new science datasets in the first nine months of 2026 as in the same period of 2025.
It also says the 18 repositories of NIH’s National Center for Biotechnology Information have been downloaded more than 11 million times, and NASA IMPACT’s 27 repositories about 770,000 times, as of 26 October. According to Hugging Face, 90% of dataset downloads on the Hub are of fully documented datasets.
The first challenge
Carnegie Mellon is hosting the first challenge, Climate Risk Intelligence, from 2 to 13 November. Teams get historical weather data for 210 US cities up to 2018 and must predict what happened from 2019 to 2026: average annual temperature, total annual precipitation, the hottest day of the year and cooling degree days, a measure of air-conditioning demand. Forecasts will be compared with simple baselines such as historical averages and linear trends. Teams also propose a short climate adaptation policy idea. Final forecasting results are due in early 2027, once full 2026 weather records exist.
The project page lists prizes of $1,000, $500 and $250 in Hugging Face credits for the top three teams, plus $2,500 for the top CMU team.
The real test is what partners release
This is TSN’s analysis. ProtoStar is a launch, not a result. The agreements are non-binding, no partner has said which datasets it will release or when, and GW’s release says details of the challenges “will be posted soon”. Until universities publish material under the scheme, there is nothing to measure.
The challenge design is sensible. Asking teams to forecast years that have already happened lets their models be scored against real weather, not hypothetical futures, and comparing them with simple statistical baselines tests whether AI adds anything at all. That is a better check than many AI-for-science announcements offer.
The project also suits Hugging Face’s business, since it is the infrastructure partner and the prizes are paid in its own credits. That does not make the aim wrong, but universities will weigh it against keeping data in specialist repositories. Even CMU, the first host, still recognises the importance of domain-specific repositories, according to Hugging Face.
ProtoStar joins a wider push to make research data AI-ready, from the $1.8bn effort led by Biohub, the US Department of Energy and NIH to Europe’s AI in Science Enablers initiative, and follows Meta’s grants to five US universities. The payoff, if it comes, looks like the AI-assisted search that helped find the antibiotic abaucin: open, well-documented data that others can build on. The first sign will be how many datasets the partners publish in the coming months.
Sources
- Hugging Face, “Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready”, blog post, 29 October 2026
- George Washington University, “GW Is a Founding Partner in Hugging Face’s Project ProtoStar, a National Effort to Make Publicly Funded Science AI-Ready”, 29 October 2026
- Hugging Face, Project ProtoStar page
Related stories
- NVIDIA Joins the European Commission’s AI in Science Enablers Initiative
- AI Screened 6,680 Molecules and Helped Find a New Antibiotic, Abaucin
- Biohub, DOE and NIH Lead $1.8bn Push for AI-Ready Biology Data
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
