OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets

arXiv cs.AI · 18 hours ago

OriginBlame is a data provenance system that tracks authorship through data processing pipelines to enable precise identification of training records for removal requests. The system reduced over-deletion from 101x to 1.3x on 219,555 Wikipedia pages while adding 1.3-4.0% throughput overhead at HuggingFace. This allows model trainers to comply with data removal requests without unnecessarily retraining on unrelated data.