Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready
Hugging Face announced Project ProtoStar, an initiative aimed at making publicly funded scientific research easier to discover, reuse and build upon, especially for AI-driven work. The announcement names founding partners across academia, industry and the open science community: Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University and The George Washington University, working through the Open Forum for AI (OFAI) as the university coalition partner; the Astera Institute; Hugging Face as infrastructure partner; and the Open Research Community Accelerator (ORCA), through its Fostering Innovation Network, as community partner. Participants formalize involvement through a memorandum of understanding committing them to release suitable datasets, models and code in AI-ready formats on open infrastructure. Organizations can express interest at protostar@orcaopen.org.
The initiative is built around two goals: encouraging systematic sharing of publicly funded scientific artifacts in machine-actionable formats, and demonstrating their value through a series of open challenges hosted by participating institutions. The first challenge, run by Carnegie Mellon University, opens November 2, 2026.
The problem ProtoStar addresses is that public availability does not equal usability. Scientific datasets, models and code are often spread across incompatible systems, stored in static or legacy formats, or released without the documentation, metadata and provenance needed to understand and reuse them. That makes it harder for researchers to build on prior work and harder for AI systems to discover, interpret and combine resources across disciplines.
The scale of the funding behind this research is significant: the U.S. federal government invests roughly $190 billion annually in research and development, and federal policy increasingly expects results to be shared. The NIH and NSF have strengthened public-access requirements for funded publications and require researchers to plan for sharing underlying data. Hugging Face is one of ten generalist repositories NIH lists for sharing research data, which positions it as a natural host for the initiative. ProtoStar builds on prior efforts such as the National Cancer Institute's AI Data-Readiness Challenge and Hugging Face's guide to preparing public-sector data for AI.
Hugging Face cites evidence that demand for reusable scientific resources is already strong. The platform hosts more than 17,000 public datasets across core scientific domains including genomics, protein design, chemistry, weather and climate, Earth observation and astronomy; that figure approaches 29,000 when medicine and mathematics are included, alongside a comparable number of scientific models. Sharing is accelerating too: researchers posted more than three times as many new science datasets in the first nine months of 2026 as in the same period of 2025, and growth remained substantial even after excluding the largest single uploader.
Usage numbers underscore the payoff of open sharing. The 18 public repositories that NIH's National Center for Biotechnology Information (NCBI) maintains on Hugging Face have collectively passed 11 million downloads, while NASA IMPACT's 27 repositories have drawn approximately 770,000 downloads, as of October 26, 2026. Documentation matters as much as raw data: 90% of dataset downloads on Hugging Face come from fully documented datasets. Hugging Face researchers have also built on openly available scientific resources, including recent HuggingFaceBio releases such as Carbon-A and Carbon-X. In practice, this suggests documentation and metadata act as a discoverability multiplier, not just housekeeping.
Institutional policy is part of the design. As an early example, CMU's Open Source Programs Office is encouraging researchers to share suitable datasets, AI models and other outputs on Hugging Face, while acknowledging the continued importance of domain-specific repositories. This suggests ProtoStar's model leans on changing incentives inside universities rather than mandating a single repository for everything.
The challenge series is where the initiative tests its central question: whether publicly funded datasets, models and code released in AI-ready formats can be readily discovered, reused, combined and extended to accelerate science and public benefit. Each challenge is hosted by a participating institution and built around resources that institution has released openly. Challenges carry two goals: producing concrete examples of reuse, including cross-institution or cross-discipline applications, and identifying technical barriers, documentation gaps, interoperability problems and successful practices that can inform future research infrastructure, funding decisions and public policy. The output is meant to be a practical evidence base.
The inaugural challenge, hosted by CMU, runs November 2–13, 2026. Participants will use historical weather data from 210 U.S. cities to test whether AI can improve predictions of urban climate conditions. Teams train on climate records through 2018 and predict actual conditions from 2019 to 2026, so models are scored against observed weather rather than hypothetical future scenarios. Four indicators are forecast: average annual temperature, total annual precipitation, the hottest day of the year, and cooling degree days, a measure of air-conditioning demand. Predictions are compared against simple statistical baselines, including historical averages and linear trends, to see whether AI delivers meaningful accuracy gains. Teams also propose a short climate adaptation policy idea informed by the data, touching areas such as preparing cities for extreme heat, improving energy efficiency or managing changing precipitation.
Submissions take the form of a Hugging Face dataset containing outputs. Forecasting is scored on a public leaderboard measuring prediction errors against baselines, while policy ideas get separate consideration. Final forecasting results will be published in early 2027, once complete 2026 weather records are available. The setup is a backtesting design familiar from time-series work: fixed historical cutoff, held-out future window, and trivial baselines to beat, which makes results easy to interpret but also means leaderboard position depends heavily on how hard the linear-trend baseline is to outperform.
Further challenges will follow sequentially, each hosted by a different participating institution and highlighting different publicly funded resources; future ones will be announced on the Project ProtoStar page on Hugging Face. ProtoStar invites academic researchers, students, independent developers and companies to participate, and welcomes universities, federal research agencies, national laboratories, nonprofits and other research organizations interested in making their resources more accessible.
Why it matters: Researchers and developers working with public science data stand to get clearer norms for packaging datasets, models and code so AI systems can actually use them, and the challenge series gives developers a concrete, low-stakes way to test reuse against real benchmarks starting November 2. Product teams building on scientific data may find better-documented, more interoperable resources upstream, though participation depends on institutions adopting the MOU and internal policies like CMU's. A likely trade-off, in my reading, is that standardized AI-ready formats can flatten domain-specific nuance that specialized repositories preserve, which is why the announcement carves out a continuing role for domain repositories. Whether AI beats simple baselines on the climate task, and whether these practices spread beyond founding partners, remain open until the early-2027 results.
Based on reporting from the original publisher. Visit the source for full context and later updates.