Publication of OmniFold Weights A Standardized Format for ML Based Unfolding Results

OmniFold produces per-event weights that cannot be reused without the original analysis environment. The original OmniFold paper deposited input datasets on Zenodo but not the learned weights, iteration metadata, or normalization contract, meaning independent reproduction requires rebuilding the full training setup from scratch. This project defines a publication standard to close that gap. The solution is a versioned metadata schema (YAML), a Parquet-based storage format for per-event weights, and a Python API for loading, validating, and applying published weights to compute observables. The package supports nominal weights, iteration metadata, and one uncertainty family as the required deliverable, with HEPData integration and extended systematics support as stretch goals. Deliverables: (1) formal schema specification with required and optional fields, (2) reader/writer package implementing the Parquet-based format, (3) validation tools covering schema compliance, normalization checks, and alignment integrity, (4) end-to-end reference example demonstrating export, reload, and observable reproduction, and (5) documentation and tutorial workflow. A working prototype with 17 passing tests is available at https://github.com/aashirvad08/omnifold-gsoc-eval.