Project ProtoStar Launches to Make Publicly Funded Science AI-Ready
Project ProtoStar unites universities, Hugging Face and science groups to make public research AI-ready, starting with CMU's climate challenge on November 2.
Summary
Announced October 29, 2026, Project ProtoStar will publish suitable publicly funded datasets, models and code in AI-ready, machine-actionable formats on open infrastructure, then test reuse through challenges. Founders are Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University and The George Washington University through the Open Forum for AI, with Astera Institute, Hugging Face and the Open Research Community Accelerator's Fostering Innovation Network. Members sign a memorandum of understanding; applicants can contact protostar@orcaopen.org.
The U.S. government invests about $190 billion yearly in R&D; NIH and NSF rules require publication access and data-sharing plans, and NIH lists Hugging Face among 10 generalist repositories. ProtoStar targets fragmented, legacy resources lacking metadata, documentation and provenance, extending the National Cancer Institute's AI Data-Readiness Challenge and Hugging Face's public-sector data guide. Hugging Face hosts over 17,000 scientific datasets, nearing 29,000 with medicine and mathematics, plus comparable model numbers; uploads in 2026's first nine months exceeded the 2025 period threefold. NCBI's 18 repositories passed 11 million downloads and NASA IMPACT's 27 reached about 770,000 by October 26; fully documented datasets generate 90% of downloads. CMU encourages suitable Hugging Face releases alongside specialist repositories. HuggingFaceBio's Carbon-A and Carbon-X demonstrate genomic reuse.
The first challenge, CMU's Climate Risk Intelligence event from November 2 to 13, uses records from 210 U.S. cities through 2018 to predict observed 2019 to 2026 conditions. Teams forecast average annual temperature, total annual precipitation, the hottest day and cooling degree days, comparing AI with historical averages and linear trends; they also propose climate adaptation policies. Submissions are Hugging Face datasets, with errors ranked publicly and policy ideas judged separately. Results arrive in early 2027, after complete 2026 weather records. Later institutions will run sequential challenges, testing cross-institutional reuse and exposing technical, documentation and interoperability barriers to guide infrastructure, funding and policy.
Positives
- Hugging Face hosts over 17,000 core scientific datasets, approaching 29,000 when medicine and mathematics are included, alongside comparable numbers of models.
- Science dataset uploads during the first nine months of 2026 exceeded the same 2025 period by more than threefold.
- NCBI's 18 Hugging Face repositories surpassed 11 million downloads, while NASA IMPACT's 27 repositories reached about 770,000 by October 26, 2026.
- Fully documented datasets account for 90% of Hugging Face dataset downloads, demonstrating the value of complete supporting information.
- CMU's challenge will evaluate AI forecasts against observed 2019 to 2026 weather and simple statistical baselines rather than hypothetical future conditions.
Risks & concerns
- Scientific resources remain scattered across incompatible systems, legacy formats and static files that obstruct discovery, interpretation and reuse.
- Missing metadata, documentation and provenance make combining public research across disciplines especially difficult for AI systems.
- The CMU challenge has not yet established whether AI can outperform historical averages and linear trends across its four climate indicators.
- Final forecasting results cannot be published until early 2027, after complete weather records for 2026 become available.
- Technical barriers and interoperability gaps may limit cross-institutional reuse despite open publication of datasets, models and code.