A new research paper introduces OriginBlame (ob), a data provenance system designed to bridge a critical gap in AI model training: the ability to locate specific records belonging to individual authors when they request data removal.
The Challenge of Over-Deletion
Current provenance systems typically operate at the file or dataset level. When a contributor requests their data be removed, model trainers often face a lack of precision, leading to "catastrophic over-deletion" where far more data is purged than necessary to satisfy the request. This occurs because trainers cannot pinpoint exactly which records in a processed pipeline belong to a specific author.
Precision through Record-Level Provenance
The ob system propagates author identity through data processing pipelines at the record and token level. This allows for deterministic queries that resolve revocation requests into precise "forget sets" for unlearning algorithms.
Key evaluation metrics from the research include:
- Eliminating Over-Deletion: In tests involving 219,555 Wikipedia pages, record-level provenance reduced dataset-level over-deletion from a factor of 101x down to 1.3x.
- Improved Unlearning: On a 1.7B parameter model, provenance-based forget sets improved the unlearning process by 42% compared to random baselines.
- Performance Overhead: Integration of the system resulted in a throughput overhead of 1.3-4.0% for HuggingFace and 2.1-19.0% for Datatrove when processing wiki data.
