OriginBlame enables precise, record-level data removal from AI training
Stop deleting entire datasets when one author opts out — OriginBlame hits record-level precision.
When a data contributor requests removal, model trainers face a practical gap: unlearning algorithms require a precise 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 — deleting hundreds of records to remove one. Haolin Xue’s OriginBlame solves this by propagating author identity through data processing pipelines and resolving revocation requests into exact record- or token-level forget sets via deterministic queries.
Evaluated on 219,555 Wikipedia pages, OriginBlame reduces dataset-level over-deletion from 101x to just 1.3x — a 98.7% reduction. Integration overhead is modest: 1.3–4.0% on HuggingFace and 2.1–19% on Datatrove with wiki data. On a 1.7B-parameter model, using provenance-derived forget sets improves unlearning quality by 42% over random baselines. This opens the door for compliant data removal under regulations like GDPR without retraining from scratch or dumping entire datasets.
- Reduces over-deletion from 101x (dataset-level) to 1.3x (record-level) across 219K Wikipedia pages
- Adds only 1.3–4.0% throughput overhead on HuggingFace and 2.1–19% on Datatrove pipelines
- Improves unlearning accuracy by 42% on a 1.7B-parameter model compared to random forget-set selection
Why It Matters
Enables precise, regulation-compliant data removal without massive retraining or over-deletion of training datasets.