training-dataset

OriginBlame is a record- and token-level data provenance system that propagates author identity through AI training data pipelines, enabling precise forget sets for machine unlearning. It eliminates over-deletion from dataset-level systems and improves unlearning effectiveness.