OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets
06:00 · July 16, 2026 · arXiv cs.AI RSS

When a data contributor requests removal, model trainers face a practical gap: unlearning algorithms require a forget set, yet no tool can locate which training records belong to a given author. Existing provenance systems operate at file or dataset level, forcing catastrophic over-deletion. We present ob, a record- and token-level data provenance system that propagates author identity through data processing pipelines and resolves revocation requests into precise forget sets via deterministic queries. Evaluation on 219,555 Wikipedia pages demonstrates that record-level provenance eliminates dataset-level over-deletion (from 101x to 1.3x), while integration adds 1.3-4.0% throughput overhead (HuggingFace) and 2.1-19.0% (Datatrove) on wiki data. On a 1.7B model, provenance-based forget sets improve unlearning by 42% over random baselines.
Summary
OriginBlame addresses a practical disconnect in machine unlearning workflows: when an author requests data removal, trainers lack a reliable way to identify the exact records that originated from that contributor. Existing data version control and provenance systems track files or entire datasets, which forces either wholesale deletion or reliance on imprecise post-hoc inference. The system propagates author identity already present in source environments—such as revision histories—through tokenization and packing stages, storing metadata in plain JSONL files under a .ob directory.
Its architecture uses three content-addressable layers linked by SHA-256 hashes. An authors layer records identities and revocation status, a sections layer captures source-file copyright details, and a document-index layer maps individual output records to their contributing sections. An optional token-index layer extends attribution to token granularity. Integration requires a single track call per record, after which deterministic queries can produce forget sets at either record or token level without model access or gradient computations.
Evaluation on 219,555 Wikipedia pages shows that record-level provenance reduces over-deletion from a factor of 101× to 1.3× relative to dataset-level approaches. Integration overhead remains modest—between 1.3 % and 4.0 % on Hugging Face pipelines and 2.1 % to 19.0 % on Datatrove—while provenance-derived forget sets improve unlearning effectiveness by 42 % over random baselines on a 1.7 B parameter model. The design avoids machine-learning dependencies and central databases, relying instead on sharded hash buckets for efficient lookup and auditability.
Why it matters
Directly supports EU GDPR and AI Act compliance needs for data removal and transparency, actionable for Dutch AI teams building ethical pipelines; high technical depth and reproducibility for researchers.


