Researchers propose OriginBlame, a system for record- and token-level data provenance in AI training datasets

Researchers from arXiv cs.AI introduced OriginBlame (ob), a record- and token-level data provenance system designed to address a practical gap in AI model training: locating which training records belong to a given author when a data contributor requests removal.

Existing provenance systems operate at file or dataset level, often forcing model trainers to perform catastrophic over-deletion to comply with revocation requests. OriginBlame propagates author identity through data processing pipelines and resolves revocation requests into precise forget sets via deterministic queries.

In evaluation on 219,555 Wikipedia pages, the system demonstrated that record-level provenance reduces over-deletion from 101x to 1.3x compared to dataset-level approaches.

The researchers also measured integration overhead: on wiki data, throughput overhead ranged from 1.3% to 4.0% when using HuggingFace and from 2.1% to 19.0% when using Datatrove.

On a 1.7 billion parameter model, provenance-based forget sets improved unlearning performance by 42% over random baselines, indicating that precise forget sets can enhance unlearning effectiveness without sacrificing model quality.