ResearchPapers, evidence & method
Hugging Face, five universities and research nonprofits launch Project ProtoStar to make publicly funded science AI-ready
Hugging Face, CMU, Purdue, Georgia Tech, ASU, GW, the Astera Institute and ORCA have formed a coalition committing members to publish publicly funded datasets, models and code in AI-usable formats, testing reuse through open challenges starting with a climate-forecasting contest.
From the Hugging Face article of October 29, 2026 and cited sources: Project ProtoStar was announced on October 29, 2026. [1] [4] [3]
SOURCES[1] Announcing Project ProtoStar (Hugging Face) huggingface.co[4] GW Is a Founding Partner in Project ProtoStar (Newswise) newswise.com[3] NCI CRDC AI Data-Readiness Challenge (NCI) datacommons.cancer.gov
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Hugging Face article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 29, 2026 article published
- Covered
- Oct 31 edition
- Evidence
- 7 sources across 7 sites
- 1Announcing Project ProtoStar (Hugging Face)huggingface.co
- 2CMU to Host Challenge to Put Scientific Data to Work With…cmu.edu
- 3NCI CRDC AI Data-Readiness Challenge (NCI)datacommons.cancer.gov
- 4GW Is a Founding Partner in Project ProtoStar (Newswise)newswise.com
- 5Accessing Scientific Data: Generalist Repositories…grants.nih.gov
- 6Generalist Repositories (NLM)nlm.nih.gov
QuietHow it’s measured
Hacker News · 1 like
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Project ProtoStar was announced on October 29, 2026. Members sign memoranda of understanding committing them to AI-ready releases of research outputs, and Hugging Face serves as the hosting partner. ProtoStar is a data-publishing and challenge program, not an AI model, and it presents NCI's 2024 AI data-readiness challenge as a predecessor. [1] [4] [3]
The first challenge, hosted by CMU from November 2 to 13, 2026, asks entrants to forecast four annual climate indicators for 210 U.S. cities for 2019–2026 using records only through 2018. Forecasts are scored against observed weather and compared with historical-average and linear-trend baselines. Policy proposals are judged separately, and final results are expected in early 2027. [1] [2]
Hugging Face says it hosts more than 17,000 public science datasets and that new science uploads more than tripled year over year. These are its own figures. Its claim to be one of the generalist repositories NIH lists conflicts with NIH's current page, which names nine repositories and does not include Hugging Face. [1] [5]
Analyst Ken Ashe argues that the hard problems are operational, not file formats: negative results, retractions, duplicate measurements and lab-specific notation. In his view, success should be judged by whether people can find data, confirm usage rights, reproduce transformations and cite sources. [7]
Evidence so far is commitments and one small challenge; no funding or prizes are disclosed.
Read the full assessment
If members follow through, standardized releases could ease cross-institution data reuse, but provenance and usage-rights metadata will likely decide practical and commercial value.
Executive brief
The ProtoStar coalition's first challenge has an unusual rule for scoring: it can't be fully scored until the weather has happened. Entrants forecast 2019–2026 climate indicators for 210 U.S. cities using records only up to 2018. Final results wait until the full 2026 weather data is in, in early 2027 (Hugging Face). Project ProtoStar was announced October 29, 2026 by Hugging Face, five universities, the Astera Institute and ORCA. Its aim is to get publicly funded datasets, models and code published in formats AI systems can find and use, and to test that reuse through open challenges (CMU News).
What changed and event timeline
NCI runs an AI data-readiness challenge
The National Cancer Institute asked participants to turn its cancer research data into AI-ready form, build models, and suggest improvements, with $50,000 in prizes. ProtoStar names it as a predecessor ().
Hugging Face takes a download snapshot
On this date, NIH's NCBI had 18 Hugging Face repositories with more than 11 million downloads, and NASA IMPACT had 27 repositories with about 770,000. Both are Hugging Face's own figures ().
ProtoStar launches
Founding partners are CMU, Purdue, Georgia Tech, ASU and GW (joining through the Open Forum for AI), plus Astera, Hugging Face and ORCA. Members sign memoranda of understanding committing to AI-ready releases (;).
CMU's Climate Risk Intelligence challenge runs
It is open to CMU-affiliated and outside participants. The focus is AI analysis of extreme-weather trends plus climate adaptation recommendations ().
Final forecasting results
These will be published once the complete 2026 weather records are available. Other member institutions will then host their own challenges, one after another ().
Capabilities and access
- No model is released. ProtoStar is a data-publishing and challenge program, not an AI system.
- Hosting: Hugging Face is the infrastructure partner. Challenge teams submit their outputs as a Hugging Face dataset (Hugging Face).
- Joining: Organizations sign an MOU. Contact is protostar@orcaopen.org.
Read the full section
- No model is released. ProtoStar is a data-publishing and challenge program, not an AI system.
- Hosting: Hugging Face is the infrastructure partner. Challenge teams submit their outputs as a Hugging Face dataset (Hugging Face).
- Joining: Organizations sign an MOU. Contact is protostar@orcaopen.org.
- CMU's policy: CMU's Open Source Programs Office encourages researchers to post suitable outputs on Hugging Face, while still recognizing domain-specific repositories.
- GW's plans: set up a Hugging Face presence, train researchers, and use open-weight models in research and coursework (Newswise/GW).
Technical analysis for researchers and developers
- Training data ends in 2018 and test data is observed weather from 2019–2026, so models are scored against reality, not simulated futures.
- Targets (per city): average annual temperature, total annual precipitation, hottest day of the year, and cooling degree days (a measure of air-conditioning demand).
- Scoring: a public leaderboard measures prediction error against simple baselines (historical averages and linear trends). Policy proposals are judged separately (Hugging Face).
Read the full section
- Evaluation design: a temporal holdout. Training data ends in 2018 and test data is observed weather from 2019–2026, so models are scored against reality, not simulated futures.
- Targets (per city): average annual temperature, total annual precipitation, hottest day of the year, and cooling degree days (a measure of air-conditioning demand).
- Scoring: a public leaderboard measures prediction error against simple baselines (historical averages and linear trends). Policy proposals are judged separately (Hugging Face).
- Not yet documented: the exact error metric, the source and licence of the weather data, and how 2019–2025 observations are kept out of submissions.
- Implication: the bar is beating a trend line, not beating other complex models.
Claims and evidence
- Growth: Hugging Face says it hosts more than 17,000 public science datasets (about 29,000 including medicine and math).
- Documentation: Hugging Face says 90% of dataset downloads go to fully documented datasets. That is a correlation; it doesn't show documentation causes downloads.
- NIH listing (conflicting): the blog says Hugging Face is one of ten generalist repositories NIH lists.
Read the full section
- Growth: Hugging Face says it hosts more than 17,000 public science datasets (about 29,000 including medicine and math). It also says new science datasets in January–September 2026 were more than triple the same period in 2025, even excluding the largest single uploader (Hugging Face). These figures come only from Hugging Face's own data.
- Documentation: Hugging Face says 90% of dataset downloads go to fully documented datasets. That is a correlation; it doesn't show documentation causes downloads.
- NIH listing (conflicting): the blog says Hugging Face is one of ten generalist repositories NIH lists. The NIH page reviewed names nine (Dataverse, Dryad, Figshare, IEEE Dataport, Mendeley Data, OSF, Synapse, Vivli, Zenodo) and does not include Hugging Face (NIH Grants).
- Coverage: the only other reporting found is university press material, which repeats the announcement. No independent corroboration was found.
Context and prior work
- ProtoStar describes itself as building on NCI's 2024 challenge.
- NIH's own generalist-repository program (GREI) funds established repositories such as Zenodo and Dryad (NLM).
- GW's open-source office dates from 2023 (Newswise/GW).
Read the full section
- ProtoStar describes itself as building on NCI's 2024 challenge. That challenge used participant feedback to tell NCI how much standardization and curation its data needed to be AI-ready (NCI CRDC).
- NIH's own generalist-repository program (GREI) funds established repositories such as Zenodo and Dryad (NLM). ProtoStar positions Hugging Face as an AI-oriented addition to these, not a replacement for domain repositories.
- GW's open-source office dates from 2023 (Newswise/GW).
Limitations, safety and contested findings
- The NIH-listing claim conflicts with NIH's current page, as noted above.
- Outside commentary: analyst Ken Ashe argues the hard part is operational, not file formats.
- Challenge scale: a two-week window and four annual indicators make it a small test of reuse, not a benchmark.
Read the full section
- The NIH-listing claim conflicts with NIH's current page, as noted above.
- Outside commentary: analyst Ken Ashe argues the hard part is operational, not file formats. Science data comes with negative results, retractions, duplicate measurements and lab-specific notation. He says success should be judged by whether people can find data, confirm usage rights, reproduce transformations and cite sources, not by the models built on it (Ken Ashe).
- Challenge scale: a two-week window and four annual indicators make it a small test of reuse, not a benchmark.
- No prizes or funding amounts are disclosed.
Business and practitioner implications
- Universities: expect institutional policies that point researchers to Hugging Face for meeting public-access rules. If NIH compliance matters, check NIH's current repository list first.
- Developers: the CMU challenge is a low-cost way to test whether ML forecasts actually beat simple statistical baselines on real outcomes.
- Vendors: standardized, machine-readable science releases make it easier to combine data across institutions, but provenance and rights metadata will decide whether commercial reuse is possible.
Read the full section
- Universities: expect institutional policies that point researchers to Hugging Face for meeting public-access rules. If NIH compliance matters, check NIH's current repository list first.
- Developers: the CMU challenge is a low-cost way to test whether ML forecasts actually beat simple statistical baselines on real outcomes.
- Vendors: standardized, machine-readable science releases make it easier to combine data across institutions, but provenance and rights metadata will decide whether commercial reuse is possible.
- Funders: each challenge's write-up of barriers may shape future infrastructure and data-policy decisions.