Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready

What it claims

Project ProtoStar is a coalition that releases publicly funded science datasets, models and code on Hugging Face in AI-ready formats, then tests their reuse through open challenges. Members sign a memorandum of understanding to release suitable artifacts on open infrastructure. The five universities take part through the Open Forum for AI (OFAI). ORCA joins through its Fostering Innovation Network. Organisations can join by writing to protostar@orcaopen.org. The post frames this against roughly $190 billion in annual US federal R&D spending and stronger NIH and NSF public-access rules. Hugging Face is one of the ten generalist repositories NIH lists for sharing research data.

Evidence

The post gives these Hub statistics as of October 2026:

• Science datasets: about 29,000 when medicine and mathematics are included, with a comparable number of science models.
• Growth: new science datasets in the first nine months of 2026 were more than triple the same period of 2025. Growth stays substantial even after excluding the largest single uploader.
• NCBI: its 18 repositories have more than 11 million downloads (26 October 2026).
• NASA IMPACT: its 27 repositories have about 770,000 downloads (26 October 2026).
• Documentation: 90% of dataset downloads on the Hub are of fully documented datasets.

How to use it

The first event is the CMU Climate Risk Intelligence Challenge, running 2 to 13 November 2026.

Task. Teams train on historical weather records from 210 US cities up to 2018. They then forecast each year from 2019 to 2026, so predictions are scored against observed weather rather than scenarios. There are four targets:

• average annual temperature
• total annual precipitation
• hottest day of the year
• cooling degree days, a measure of air-conditioning demand

Scoring. A public leaderboard compares prediction errors with simple statistical baselines: historical averages and linear trends. The stated question is whether AI methods beat those baselines meaningfully.

Second deliverable. Each team also writes a short climate-adaptation policy idea, such as preparing for extreme heat. Policy ideas are judged separately from forecasts.

Submission. Teams submit everything as a Hugging Face dataset. Registration, data and submissions are handled on the challenge page on Hugging Face.

Later challenges. Other member institutions will host challenges one after another, each built on resources that institution has released. They will be announced on the Project ProtoStar page on Hugging Face.

Limits

• Results come late. Final forecasting results will be published in early 2027, once complete 2026 weather records exist. Leaderboard standings during the event therefore cannot cover the 2026 target year.
• Scoring is underspecified. The post does not name the error metric, the size or format of the training data, or how the policy ideas are scored.
• Unstated terms. The post gives no team-size rules, prizes or data licence.
• No artifact list. The post names no specific datasets that members have committed to release under ProtoStar. The only institutional policy it describes is CMU's Open Source Programs Office encouraging researchers to share on Hugging Face.