Contents
Project ProtoStar aims to make publicly funded datasets, models, and code easier to find and reuse for AI research. Announced on October 29, 2026, the initiative begins with a Carnegie Mellon University climate challenge starting November 2.
Key Points#
- ProtoStar brings universities, research institutions, and Hugging Face together to publish scientific resources in formats that people and software can reuse.
- The first challenge runs from November 2 to 13, 2026, using historical weather data from 210 U.S. cities.
- Participants will predict four climate indicators and compare their results with historical averages and linear trends.
- Final prediction results are expected in early 2027, once all 2026 weather records are available.
What ProtoStar Changes#
Public availability does not necessarily mean practical reusability. A dataset may be online yet lack the documentation, provenance, or consistent structure needed to interpret it. Code and models also become difficult to use when their inputs or dependencies are unclear.
ProtoStar addresses this gap by helping publish suitable scientific resources on open infrastructure in AI-ready, machine-readable formats. Public challenges will then test whether people can discover, combine, and extend those resources to produce useful research.
Founding partners include Carnegie Mellon University, Purdue University, Georgia Tech, Arizona State University, and George Washington University, participating through the Open Forum for AI. Astera Institute, Hugging Face, and the Open Research Community Accelerator are also involved, with Hugging Face providing infrastructure and ORCA serving as a community partner.
How Reuse Will Be Tested#
Each challenge centers on resources published by participating institutions. The aim is to demonstrate reuse in scientific research while identifying barriers such as missing documentation, incompatible formats, and uncertainty about data interpretation.
According to the announcement, Hugging Face hosts more than 17,000 public datasets across major scientific fields, rising to nearly 29,000 when medicine and mathematics are included. It also states that well-documented datasets account for 90% of dataset downloads. These figures show active use, but do not establish that documentation alone causes increased usage.
ProtoStar’s useful outcomes should therefore extend beyond winning submissions. Evidence showing which resources were reusable, which needed fixes, and what prevented successful combinations can help institutions improve future releases. The initiative sets out these goals, but has yet to demonstrate results.
The First Climate Challenge#
CMU’s Climate Risk Intelligence challenge asks participants to use weather records through 2018 to predict observations for 2019–2026 across 210 U.S. cities. This allows prediction accuracy to be tested retrospectively, though evaluating the final year requires complete records.
Participants will predict four indicators:
- Annual average temperature.
- Total annual precipitation.
- The hottest day of the year.
- Cooling degree days, a measure related to cooling demand.
Predictions will be compared with simple statistical baselines, including historical averages and linear trends. Each team must also propose a brief climate adaptation plan based on its analysis. Prediction errors will be evaluated on a public leaderboard, while policy proposals will be judged separately. Teams will submit their work as Hugging Face datasets.
What Participants Should Check#
Read the published challenge rules before choosing a model. Check the dataset structure, required submission components, evaluation metrics, and rules for using additional data. The announcement alone does not provide enough detail to determine these.
Practical preparation steps are:
1. Check data coverage and missing values by city, year, and target indicator. 2. Reproduce the specified statistical baselines before trying complex methods. 3. Follow the challenge rules and exclude observations from the evaluation period during model development. 4. Document the method and connect policy proposals to specific analytical findings.
Interpret results within their limits. Low prediction error can support claims of performance on this task, but does not by itself prove that a model improves urban planning. Likewise, adopting a policy proposal requires evidence beyond a leaderboard score. With final prediction results expected in early 2027, earlier evaluations should specify which observations are available.
Next Steps#
Visit the CMU Climate Risk Intelligence Challenge page on November 2, 2026, and review the submission requirements before preparing the dataset.
Source: https://huggingface.co/blog/announcing-project-protostar
Contents
