An initiative bringing collectively universities, analysis organizations, business, and the open science group to make publicly funded analysis simpler to find, reuse, and construct upon.
Right now, we’re asserting Venture ProtoStar, an initiative to unlock the potential of publicly funded scientific analysis by making datasets, fashions, and code extra accessible and reusable, significantly for AI-driven analysis. ProtoStar brings collectively analysis establishments, business, and the open science group to help the discharge of scientific artifacts in AI-ready, machine-actionable codecs, and to exhibit their worth by means of a collection of open challenges. The primary problem, hosted by Carnegie Mellon College, opens on November 2. See How one can Take part under.
The initiative’s founding companions embrace Carnegie Mellon College, Purdue College, Georgia Institute of Know-how, Arizona State College, and The George Washington College, all working by means of the Open Discussion board for AI (OFAI) because the college coalition accomplice; the Astera Institute; Hugging Face because the infrastructure accomplice; and the Open Analysis Group Accelerator (ORCA), by means of its Fostering Innovation Community, because the group accomplice.
Collaborating organizations formalize their involvement by means of a memorandum of understanding supporting the discharge of appropriate scientific datasets, fashions, and code in AI-ready codecs on open infrastructure. Organizations eager about becoming a member of Venture ProtoStar can attain out to protostar@orcaopen.org.
Making publicly funded science extra usable
The U.S. federal authorities invests roughly $190 billion yearly in analysis and improvement, and federal coverage more and more expects the outcomes to be shared publicly. Companies such because the Nationwide Institutes of Well being (NIH) and Nationwide Science Basis (NSF) have strengthened public-access necessities for funded analysis publications and require researchers to plan for sharing the underlying information. Hugging Face is among the many ten generalist repositories listed by NIH for sharing analysis information.
However making analysis publicly obtainable doesn’t robotically make it usable. Scientific datasets, fashions, and code are sometimes scattered throughout incompatible methods, saved in static or legacy codecs, or launched with out the documentation, metadata, and provenance wanted to know and reuse them. These limitations make it more durable for researchers to construct on present work, and more durable nonetheless for AI methods to find, interpret, and mix scientific assets throughout disciplines.
Venture ProtoStar goals to assist shut this hole by encouraging systematic sharing of publicly funded scientific artifacts in AI-ready codecs, supported by open infrastructure and sensible publishing steering. It builds on efforts such because the Nationwide Most cancers Institute’s AI Knowledge-Readiness Problem and Hugging Face’s information to getting ready public-sector information for AI. The aim is to make it simpler for analysis organizations to fulfill present public-access expectations whereas rising the scientific worth of the work they launch.
The demand for reusable scientific assets is already evident. Hugging Face hosts greater than 17,000 public datasets throughout core scientific domains, together with genomics, protein design, chemistry, climate and local weather, Earth commentary, and astronomy. That determine approaches 29,000 when medication and arithmetic are included, alongside a comparable variety of scientific fashions.
The tempo of sharing can also be accelerating. Researchers shared greater than 3 times as many new science datasets on Hugging Face within the first 9 months of 2026 as in the identical interval of 2025, with progress remaining substantial even after excluding the biggest particular person uploader.
Determine 1. New public science datasets on Hugging Face per quarter, by area (Q1 2023–Q3 2026). Counts datasets nonetheless public, by creation date. Supply: Hugging Face Hub information, October 2026. Discover the information.
When organizations share analysis in reusable codecs, the scientific group can construct on it at scale. The 18 public repositories maintained by NIH’s Nationwide Heart for Biotechnology Data (NCBI) on Hugging Face have collectively acquired greater than 11 million downloads, whereas NASA IMPACT’s 27 repositories have acquired roughly 770,000 downloads, as of October 26, 2026. Documentation can also be an necessary a part of making these assets helpful: 90% of dataset downloads on Hugging Face are from totally documented datasets.
Publicly obtainable genomic information has proved necessary in contributing to analysis that additional advances fields similar to organic analysis; researchers at Hugging Face have been constructing on overtly obtainable scientific assets, together with current releases from HuggingFaceBio similar to Carbon-A and Carbon-X. By releasing the ensuing fashions, datasets, and instruments overtly, these efforts exhibit how present scientific assets can help new analysis that others can construct upon.
That is the form of exercise that we’re additional catalyzing through Venture ProtoStar by means of institutional insurance policies adopted by collaborating orgs. As an early instance of institutional dedication, CMU’s Open Supply Applications Workplace is encouraging its researchers to share appropriate datasets, AI fashions, and different analysis outputs on Hugging Face, whereas recognizing the continued significance of domain-specific repositories.
Open Science Challenges
Making scientific artifacts AI-ready is a crucial first step. We additionally need to exhibit what researchers, builders, and the broader scientific group can accomplish when these assets are made obtainable for reuse.
A central query for ProtoStar is: Can publicly funded scientific datasets, fashions, and code, launched in AI-ready codecs, be readily found, reused, mixed, and prolonged to speed up scientific discovery and public profit?
To discover this, ProtoStar will arrange a collection of open challenges, every hosted by a collaborating analysis establishment and constructed round scientific assets that the establishment has launched overtly. The challenges have two complementary objectives:
- Show scientific reuse. Produce concrete examples of researchers and builders constructing on publicly funded scientific artifacts, together with purposes that mix assets throughout establishments or disciplines.
- Be taught what makes reuse doable. Determine technical obstacles, documentation gaps, interoperability challenges, and profitable practices that may inform future analysis infrastructure, funding choices, and public coverage.
The ensuing tasks and classes discovered will assist construct a sensible proof base for making publicly funded science extra accessible and helpful.
The primary problem: Local weather Threat Intelligence
Hosted by Carnegie Mellon College | November 2–13, 2026
The inaugural ProtoStar problem invitations individuals to check whether or not AI can enhance predictions of city local weather circumstances utilizing historic climate information from 210 U.S. cities. Constructing on analysis by Carnegie Mellon College, groups will use local weather information by means of 2018 to foretell precise circumstances from 2019 to 2026, permitting their fashions to be evaluated in opposition to noticed climate somewhat than hypothetical future eventualities.
Contributors will forecast 4 indicators related to city planning: common annual temperature, whole annual precipitation, the most well liked day of the yr, and cooling diploma days, a measure of demand for air-con. Their predictions shall be in contrast in opposition to easy statistical baselines, together with historic averages and linear developments, to evaluate whether or not AI approaches provide significant enhancements in accuracy.
Past forecasting, individuals may even suggest a brief local weather adaptation coverage concept knowledgeable by what they discovered from the information. This element encourages groups to attach scientific evaluation with sensible public advantages, similar to getting ready cities for excessive warmth, enhancing vitality effectivity, or managing altering precipitation patterns.
Groups will submit a Hugging Face dataset with these outputs. Forecasting efficiency shall be assessed by means of a public leaderboard measuring prediction errors in opposition to established baselines, whereas coverage concepts will obtain separate consideration. Remaining forecasting outcomes shall be revealed in early 2027, as soon as the entire 2026 climate information can be found.
Discover the CMU Local weather Threat Intelligence Problem
The CMU problem is the primary on this collection. Different collaborating establishments will host their very own challenges sequentially, every highlighting totally different publicly funded scientific assets and alternatives for reuse. Future challenges shall be introduced by means of the Venture ProtoStar web page on Hugging Face.
How one can Be a part of Venture ProtoStar
Whether or not you are an educational researcher, pupil, unbiased developer, or firm constructing on open scientific infrastructure, ProtoStar affords alternatives to place publicly funded analysis to work and assist us be taught what is required to make AI-ready open science extra extensively helpful.
We additionally welcome universities, federal analysis companies, nationwide laboratories, nonprofits, and different analysis organizations eager about making their scientific assets extra accessible and reusable. To discover participation or collaboration, attain out to protostar@orcaopen.org.
Our hope is that ProtoStar will make it simpler for publicly funded scientific work to achieve the folks and instruments that may construct on it, whereas producing sensible proof concerning the infrastructure and practices wanted to help open science at scale.
