Aditya Goyal

WARP: Weight-Space Analysis for Recovering Training Data Portfolios

Tzu-Heng Huang, Aditya Goyal, John Cooper, Frederic Sala

Foundation models are routinely released to the public, yet the data recipes used to train them -- such as domain mixture weights that determine how different sources are sampled -- are rarely disclosed. This creates an access asymmetry: researchers study the resulting models but…Read more