Research3h ago
OriginBlame Enables Precise Training Data Deletion
The OriginBlame system reduces data over-deletion from 101x to 1.3x by tracking author identity at the token level. It allows model trainers to resolve revocation requests into precise forget sets via deterministic queries. Integration adds up to 19% throughput overhead in Datatrove pipelines. This solves a critical gap in AI unlearning workflows.