OriginBlame brings record- and token-level provenance to AI training data

Introduction

When a data contributor asks for their material to be removed from an AI training corpus, the hard part is often not the policy decision but the plumbing. Machine unlearning methods typically need a forget set: a concrete list of records that should be removed or forgotten. Yet after web data has passed through cleaning, filtering, deduplication, formatting, and tokenization, a trainer may no longer know which training examples came from a particular author.

The arXiv paper “OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets” targets this gap. It introduces OriginBlame, or ob, as a provenance layer for AI training datasets. Rather than operating only at the file or dataset level, it propagates author identity through data processing pipelines and makes it possible to resolve revocation requests with deterministic queries.

Key points

• Finer-grained provenance: Existing systems often track provenance at the file or dataset level. That can force trainers to delete far more data than necessary when one contributor requests removal. OriginBlame focuses on record- and token-level provenance.
• From request to forget set: The system is designed to map an author’s removal request to the corresponding training records, producing a more precise forget set for downstream unlearning algorithms.
• Less over-deletion: On 219,555 Wikipedia pages, the paper reports that record-level provenance reduced dataset-level over-deletion from 101x to 1.3x.
• Measured pipeline cost: Integration added 1.3% to 4.0% throughput overhead in HuggingFace pipelines and 2.1% to 19.0% in Datatrove on wiki data.
• Unlearning benefit: In experiments with a 1.7B-parameter model, provenance-based forget sets improved unlearning by 42% over random baselines.

Why it matters

OriginBlame addresses a practical infrastructure problem behind data rights, copyright compliance, and model unlearning. If a training organization cannot identify the records tied to a contributor, it may have to remove entire files or datasets, wasting useful data and potentially harming model quality.

The paper’s broader message is that data governance has to be built into the preprocessing stack, not bolted on after a model is trained. Provenance metadata becomes most useful when it is preserved before records are transformed, mixed, and tokenized.

This work does not by itself settle legal questions about consent, licensing, or privacy, and its results are reported under the datasets, pipelines, and model scale studied in the paper. Still, it points toward an important direction for AI infrastructure: training data may need versioning and provenance controls as rigorous as those already used for software and model artifacts.

Source: arXiv