Project ProtoStar Explained: AI-Ready Science and CMU's Climate Challenge

Hugging Face and research partners announced Project ProtoStar on October 29 to make publicly funded datasets, models and code easier to reuse in AI research. Its first scheduled challenge, hosted by Carnegie Mellon University, will test climate forecasting. This is a research initiative, not a new foundation model.

The launch announcement sets the CMU challenge window for November 2-13, 2026. As of October 30, the competition has been announced but its scheduled window has not begun.

What Project ProtoStar is building

The project's public page describes three connected activities: helping researchers publish usable artifacts, maintaining discoverable repositories and running public challenges that test whether others can build on them.

Founding partners include CMU, Purdue, Georgia Tech, Arizona State and George Washington University, alongside the Astera Institute. Hugging Face supplies open infrastructure, with the Open Forum for AI and Open Research Community Accelerator participating in the collaboration.

Participating organizations retain control over what they release. Their memoranda of understanding are non-binding and non-exclusive. This is not a promise that every dataset held by a partner university will become public, or that existing specialist repositories will be replaced.

That distinction matters for developers evaluating the initiative. The immediate value is an organized route to reusable research artifacts and people maintaining them, rather than unrestricted access to all institutional data.

CMU's first challenge tests climate predictions

The inaugural challenge uses historical weather records covering 210 US cities. Participants will forecast annual temperature, precipitation, the hottest day and cooling degree days, then propose a climate-adaptation policy idea.

Forecasting is compared with simple baselines such as historical averages and linear trends. The announcement says final results will arrive in early 2027, after complete 2026 weather observations become available.

The launch post and project page disagree on the forecast period's final year: 2026 versus 2025. Confirm the evaluation period in the CMU challenge rules before preparing a submission.

A simple baseline is important here. A more complicated model should demonstrate that it improves the forecast, not merely produce a more elaborate explanation of it. Separating prediction quality from the policy proposal also helps keep two different judgments from becoming one vague score.

What AI-ready data requires

Hugging Face's existing public-data preparation guide gives a concrete example of the work involved. Massachusetts education spreadsheets needed consistent headers, column names and data types before they could be combined across years. Other public records required extracting tables from PDFs.

Those are engineering tasks, not model-selection tasks. A forecasting team can lose time to an ambiguous unit, changing field definition or missing transformation history before training even begins. Publishing the conversion process alongside the result makes that work inspectable and reusable.

The important outcome is repeatable reuse. A successful challenge entry is useful, but the stronger test is whether another team can find the same data, understand its limitations and reproduce the preparation without institutional hand-holding. ProtoStar's early challenges should make those remaining gaps easier to see.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help research and developer-tool teams turn complex technical initiatives into clear stories that make participation easier to understand.

Book a 30-minute call.