Hugging Face launches Project ProtoStar for AI-ready public science data

Hugging Face has announced Project ProtoStar, a collaborative initiative involving Carnegie Mellon University, Purdue, Georgia Tech, Arizona State, George Washington University, the Astera Institute, and ORCA to improve the discoverability and reusability of publicly funded scientific research for AI applications. The project addresses the gap between federal public-access mandates and actual usability, where research artifacts often remain scattered across incompatible systems, stored in legacy formats, or lack sufficient documentation and metadata for machine consumption.

Participating organizations commit through memoranda of understanding to release suitable scientific datasets, models, and code in AI-ready formats on open infrastructure, with CMU's Open Source Programs Office already encouraging researchers to share outputs on Hugging Face. The initiative will test its approach through a series of open challenges, beginning with CMU's challenge on November 2, to demonstrate whether standardized, publicly funded scientific resources can be effectively discovered, combined, and extended to accelerate research.

Institutional structure and AI-readiness requirements

• Open infrastructure mandate: Participating organizations must release scientific datasets, models, and code on open infrastructure, with Hugging Face serving as the designated infrastructure partner and one of ten generalist repositories approved by NIH for research data sharing.
• AI-ready format requirements: The initiative specifies machine-actionable formats with documentation, metadata, and provenance rather than static or legacy formats, addressing the gap between public availability and actual computational usability.
• Institutional policy integration: CMU's Open Source Programs Office is implementing internal policies encouraging researchers to share outputs on Hugging Face while maintaining domain-specific repository relationships, demonstrating how universities can operationalize open science commitments.
• Challenge-based validation: ProtoStar will use open challenges to empirically test whether AI-ready public scientific resources can be discovered, reused, combined, and extended, with the first challenge hosted by Carnegie Mellon University opening November 2.
• Growth baseline: Hugging Face reports more than 17,000 public datasets in core scientific domains (approaching 29,000 including medicine and mathematics), with science dataset uploads in the first nine months of 2026 more than triple the same period in 2025.

Why it matters

Project ProtoStar directly addresses a structural bottleneck in open science: the $190 billion annual U.S. federal research investment produces vast outputs that remain difficult to integrate into modern AI workflows due to format fragmentation and poor metadata. By establishing institutional MOUs that bind universities to specific release practices on open infrastructure, the initiative creates a template other institutions can adopt without negotiating proprietary platforms or restrictive licensing. For Linux and open-source AI developers, standardized access to genomic, climate, chemistry, and astronomical datasets on Hugging Face reduces data preparation overhead and enables reproducible pipelines that can run on commodity hardware or ROCm/CUDA backends without vendor-specific data extraction tools.