Project ProtoStar: Hugging Face Wants Publicly Funded Science to Be AI-Ready

Hugging Face has announced Project ProtoStar, an initiative to make publicly funded scientific datasets, models and code easier to find, reuse and combine, particularly for AI-driven research. It pairs five US universities, a research foundation and an open-science community with Hugging Face's hosting infrastructure, and it starts with a public challenge: forecasting urban climate for 210 US cities, hosted by Carnegie Mellon University from November 2 to 13, 2026.

For readers who build with AI, the practical point is data. Many of the best scientific datasets sit in static files with thin documentation, which makes them hard for people and for AI agents to use. ProtoStar tries to change that by agreeing on formats and by running challenges that test whether reuse actually works. We have followed Hugging Face's science push for months, including its Carbon-A DNA model and the Carbon-X variant-effects model. ProtoStar is the data-side counterpart.

TL;DR: the questions people ask first

What problem is ProtoStar trying to solve?

Hugging Face cites the figure that the US federal government invests roughly $190 billion a year in research and development, and notes that agencies such as the National Institutes of Health and the National Science Foundation have strengthened public-access rules for funded publications and require data-sharing plans. Hugging Face says it is among the ten generalist repositories listed by NIH for sharing research data; we could not open the NIH page ourselves to confirm the list, so treat that as Hugging Face's statement.

The post's argument is that public does not mean usable. Datasets, models and code are often scattered across incompatible systems, stored in legacy formats, or released without the documentation, metadata and provenance a stranger needs. That slows researchers and, in the post's words, makes it harder for AI systems to discover, interpret and combine scientific resources across disciplines. ProtoStar aims to encourage systematic sharing in AI-ready formats, backed by open infrastructure and practical publishing guidance, building on the National Cancer Institute's AI Data-Readiness Challenge and Hugging Face's guide to preparing public-sector data for AI.

Who is behind it?

The founding partners listed are Carnegie Mellon University, Purdue University, the Georgia Institute of Technology, Arizona State University and The George Washington University, all 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. Participating organizations formalize their involvement with a memorandum of understanding supporting release of suitable datasets, models and code in AI-ready formats on open infrastructure.

One early commitment is named: CMU's Open Source Programs Office is encouraging its researchers to share suitable datasets, AI models and other outputs on Hugging Face, while recognizing that domain-specific repositories still matter. That caveat is important: ProtoStar is not asking science to leave its existing archives, only to make artifacts reusable.

A box opening as green seeds drift out toward open hands, picturing Project ProtoStar releasing publicly funded research openly

The first challenge: Climate Risk Intelligence

Hosted by Carnegie Mellon, building on CMU research, the challenge asks whether AI can improve predictions of urban climate conditions using historical weather data from 210 US cities. Teams use records through 2018 to predict actual conditions from 2019 to 2026. Because evaluation is against observed weather rather than hypothetical future scenarios, there is no argument about which scenario was right.

Participants forecast four indicators relevant to urban planning:

1. Average annual temperature
2. Total annual precipitation
3. The hottest day of the year
4. Cooling degree days, a measure of demand for air conditioning

Predictions are compared with simple statistical baselines, historical averages and linear trends. That design is the most interesting part. In climate and weather work, a flashy model that cannot beat a linear trend is not progress, and the organizers have built that check into the leaderboard. Beyond forecasting, teams propose a short climate adaptation policy idea informed by the data, such as preparing cities for extreme heat, improving energy efficiency or managing changing precipitation. Teams submit a Hugging Face dataset with their outputs. Final forecasting results will be published in early 2027, once complete 2026 weather records are available, so expect the leaderboard to be provisional until then.

For context on the model side of weather AI, see our coverage of Google DeepMind's WeatherNext 3 and NVIDIA's StormScope. Those are pretrained weather models; this challenge is about city-level annual indicators from a specific open dataset, which is a different and more modest task.

What numbers does Hugging Face give, and how much do they prove?

The announcement includes several figures, mostly from Hugging Face's own Hub data:

• More than 17,000 public datasets across core scientific domains such as genomics, protein design, chemistry, weather and climate, Earth observation and astronomy, approaching 29,000 when medicine and mathematics are included, alongside a comparable number of scientific models.
• More than three times as many new science datasets in the first nine months of 2026 as in the same period of 2025, with growth remaining substantial even after excluding the largest single uploader.
• The 18 public repositories maintained by NIH's National Center for Biotechnology Information on Hugging Face have received more than 11 million downloads, and NASA IMPACT's 27 repositories about 770,000, as of October 26, 2026.
• 90% of dataset downloads are from fully documented datasets, a figure the announcement attributes to an arXiv paper (2401.13822).

These support a reasonable reading: researchers are using the Hub for science, and documentation correlates with use. They do not show that AI-ready formats cause better science, and they say nothing about the many datasets that live elsewhere. The 90% figure is a correlation, since popular datasets are likely to be better documented. Note also that the authors work for the platform being praised, which is normal for a launch post but worth keeping in mind.

Why this matters for builders

• Agents need clean data. A research agent can only combine datasets whose schema and provenance it can read. Machine-actionable metadata is the prerequisite for the agentic science tools we covered in Anthropic's AI for Science expansion and the Genesis Mission pledges.
• Benchmarks without leakage. Forecasting 2019 to 2026 from data through 2018 is a clean temporal split, a good template for your own evaluations.
• A route to visibility. If you maintain a public dataset or model from grant-funded work, the ProtoStar framing gives a checklist: documentation, metadata, provenance and an open home.
• Baselines matter. Whatever you build for the challenge, report how it compares with the linear trend first.

What does "AI-ready" actually mean?

The announcement uses the phrase "AI-ready, machine-actionable formats" without a formal standard, but the surrounding text points to a working definition. A dataset is AI-ready when a stranger, or an agent, can load it without emailing the authors: a documented schema, units and collection method, a license, a stable identifier, a fixed version, and a file format that tools can stream rather than a PDF table or a legacy archive. Provenance matters too, meaning where each record came from and how it was cleaned, because models trained on undocumented data are hard to audit.

This is the same lesson the model-release side has learned. A checkpoint without a model card, license or reproducible evaluation is difficult to trust, which is why we keep separating what a lab claims from what can be checked in our claimed-versus-verified posts. ProtoStar applies that discipline to data. The test it proposes, whether an outside team can discover, reuse, combine and extend a resource, is concrete and measurable, and the challenges are designed to record where that process breaks.

How to take part

A three-step podium with a green star on top, picturing the public leaderboard for the ProtoStar climate challenge

Anyone, from academic researchers and students to independent developers and companies, can join the challenges, which open November 2 on the CMU challenge page linked from the announcement. Organizations that want to contribute resources, including universities, federal research agencies, national laboratories and nonprofits, can email protostar@orcaopen.org. Further challenges will be hosted sequentially by other participating institutions around their own released resources and announced through the Project ProtoStar page on Hugging Face.

A practical way to prepare: pull the dataset when it opens, compute historical-average and linear-trend baselines for all four indicators, and only then try a learned model. If you cannot beat the baselines on hottest day, which is an extreme-value statistic and noisier than annual means, say so; the organizers say they want to learn what makes reuse possible, including technical barriers and documentation gaps.

What is not yet known

• No funding amounts, prizes or compute grants are mentioned in the announcement.
• The post does not say how many datasets the partners will release, or by when.
• The format standard is described as AI-ready and machine-actionable, but no schema is named.
• It does not say how the policy ideas will be scored.

We will update this article when the challenge page, leaderboard or further partners appear.

Related reading

• Hugging Face Carbon-A DNA model
• Hugging Face Carbon-X claims, checked
• Anthropic AI for Science expansion
• Genesis Mission industry pledges
• WeatherNext 3 forecasts
• NVIDIA StormScope
• Official: Hugging Face announcement

Dates and figures are accurate as of October 29, 2026 and come from the Hugging Face announcement.