Provable Training Data Identification for Large Language Models
The work formalizes training data identification as a set-level inference problem to provide provable control over false identification rates (FIR). It proposes Provable Training Data Identification (PTDI), a distribution-free method using conformal p-values and a novel Jackknife-corrected Beta boundary estimator. This approach allows researchers to select subsets of test data with guaranteed FIR control while maximizing the power to find true training samples.