OriginBlame provides record- and token-level data provenance, enabling precise deletion of contributor data without over-deletion, improving model unlearning by 42%.

1 min read

OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets

Introduction

FAQ

What is OriginBlame?

A record- and token-level data provenance system that precisely identifies training records belonging to a specific author.

How does OriginBlame solve data deletion?

It resolves revocation requests into precise forget sets via deterministic queries, avoiding dataset-level over-deletion.

What is the impact on model performance?

It improves unlearning by 42% on a 1.7B model with minimal overhead.

Can OriginBlame be used in MENA?

Yes, especially for organizations handling sensitive data and needing compliance with privacy regulations like GDPR.

Source: arXiv cs.AI

AI-assisted content, human-reviewed.