OriginBlame Tracks Data Provenance for AI Unlearning.

When data contributors request the removal of their information, AI model trainers face a significant challenge: existing provenance systems typically operate at a coarse file or dataset level, leading to excessive data deletion during unlearning. To address this, a new system called OriginBlame (ob) has been developed. OriginBlame offers granular data provenance at both the record and token levels. It propagates author identity through data processing pipelines, allowing for deterministic queries that precisely identify which training records belong to a specific author. This capability enables the creation of accurate "forget sets" for unlearning algorithms. Evaluations show that OriginBlame drastically reduces over-deletion (from 101x to 1.3x) compared to dataset-level methods, with only a modest throughput overhead (1.3-4.0% for HuggingFace, 2.1-19.0% for Datatrove) during integration. For a 1.7B model, provenance-based forget sets improved unlearning effectiveness by 42% over random baselines.