Data Provenance in Training
Domain: Data & Datasets
π― Motivation
Tracking where every piece of training data came from: source, date, modifications, license. Essential for compliance.
π Concrete Example
Provenance: βArticle 1234: source: BBC News, date: 2024-03-15, original, license: CC-BY, no modifications.β Full traceability for every training example.
π Analogy
A diamondβs certificate - mine of origin, cut date, carat weight, chain of ownership. Data provenance is the certificate for each training example.