level: technical
when a data contributor asks for their data to be removed, model trainers face a problem. unlearning algorithms need a forget set, but no tool can find which training records belong to a specific author. current provenance systems work at file or dataset level, which forces too much deletion. originblame is a new system that tracks author identity through data processing pipelines. it can resolve removal requests into exact forget sets using deterministic queries.
tests on 219,555 wikipedia pages show that record-level provenance cuts dataset-level over-deletion from 101 times to 1.3 times the needed amount. adding this tracking adds a small overhead: 1.3 to 4.0 percent throughput drop on huggingface and 2.1 to 19.0 percent on datatrove for wiki data. on a 1.7 billion parameter model, forget sets built with provenance improve unlearning by 42 percent over random baselines.
originblame works by propagating author information at both record and token levels. this means it can pinpoint exactly which parts of the data belong to a person, not just whole files. the system integrates into common data processing tools, making it practical for real-world use. it helps comply with data removal requests without losing unrelated data, which is important for model accuracy and fairness.
why it matters: precise data removal helps meet legal and ethical requirements without degrading model performance, a key need for responsible ai deployment.