ProtoStar puts the reuse of public research to a climate-data test

Project ProtoStar is asking a practical question about open science: can someone outside the original research team actually use its data, models and code? Announced on 29 October, the collaboration will test that question through public challenges, beginning with a climate exercise hosted by Carnegie Mellon University.[1]

The founding partners include Carnegie Mellon, Purdue, Georgia Tech, Arizona State and George Washington universities, the Astera Institute and open-science organisations. Hugging Face supplies infrastructure; the universities participate through the Open Forum for AI, while the Open Research Community Accelerator contributes its research community.[1]

Forecasting with a fixed information cutoff

Carnegie Mellon’s inaugural challenge runs from 2 to 13 November and is open beyond the founding institutions. The project page asks entrants to use historical weather records for 210 US cities, with data restricted to 2018 and earlier, to forecast conditions in 2019–2025 and propose a climate-adaptation policy idea.[2]

The cutoff turns an already-observed period into a retrospective forecasting test. Its usefulness depends on keeping later information out of model inputs and reporting how the supplied records were processed. That is a different exercise from demonstrating a successful live forecast.[2]

The launch materials contain a date inconsistency: the announcement describes forecasts through 2026, while the project page ends the target period in 2025. Participants should check the current competition rules and submission template before preparing an entry.[1][2]

The work between a download and a reusable dataset

ProtoStar’s stated goals include identifying documentation gaps and interoperability problems, as well as demonstrating applications of publicly funded research. A missing dependency or unclear field can prevent reuse even when the source files are freely available. The challenges are intended to make those barriers visible.[1]

An earlier Hugging Face guide, which the programme builds upon, gives a concrete example from Massachusetts school-assessment data. The source came in spreadsheets, yet header positions, column names, numeric types and achievement-level definitions differed across years. A machine-readable file was only the starting point.[3]

Standardising a column name is a technical operation; establishing whether a score means the same thing in two years requires understanding the measurement. The guide also describes workforce figures extracted from presentation files and PDFs. In both cases, documentation has to preserve the context that makes the resulting table meaningful.[3]

Participating organisations retain control over which materials are suitable for release and how they are shared. The programme describes its memoranda as non-binding and non-exclusive, and pairs repositories with licensing, versioning and preservation practices.[2]

The lasting output will be more than a winning forecast if another team can follow the data, code and decisions afterwards. That is the test ProtoStar’s open-science infrastructure is being built to face.[2]

Report an error or suggest a correction