# Apache Sourcelume introduces JSON-LD metadata spec for AI model and dataset provenance

> **Key takeaway:** Apache Sourcelume provides a machine-readable framework for tracking the origin, licensing, and custody of AI training datasets and models.

**Published:** 2026-11-09T06:01:52+00:00 
**Source:** Apache Software Foundation News & Sponsorships 
**Category:** policy-community 
**Canonical URL:** https://fosswire.org/news/apache-sourcelume-introduces-json-ld-metadata-spec-for-ai-mo.html 

## Summary
The Apache Software Foundation has introduced Apache Sourcelume, a project designed to address the lack of verifiable transparency in AI training data. Currently, model provenance is often documented in non-machine-readable formats like PDFs or model cards, making it difficult to audit licensing and data origins. Sourcelume provides the technical plumbing to turn data documentation into a standard part of the machine learning build pipeline.

The project utilizes a versioned JSON-LD metadata specification to record dataset origins and custody chains. It features a reference registry built on Apache Atlas that serves these records via REST APIs. By introducing specific record types for both datasets and models, Sourcelume aims to create a functional equivalent to a Software Bill of Materials (SBOM) for the artificial intelligence ecosystem.

## Technical architecture of Sourcelume
- **Metadata Specification:** Uses a versioned JSON-LD format to define structured records for dataset origin, licensing, and custody chains.
- **Record Types:** Implements a `ProvenanceRecord` for datasets and a `TrainingProvenanceRecord` for models, the latter of which references datasets via canonical identifiers and includes token counts and loss metrics.
- **Registry and API:** Employs Apache Atlas as a reference registry to store and serve provenance records through standard REST APIs.
- **Pipeline Integration:** Designed to be integrated directly into training pipelines, allowing trainers to automatically write provenance records (including timestamps and data shares) during the model training run.
- **Disclosure States:** Supports explicit disclosure levels within records, categorized as complete, partial, or honestly undisclosed.

## Why it matters
By providing a machine-readable way to track training data, Sourcelume enables automated auditing of AI models for licensing compliance. This development is critical for software engineers and legal teams who need to verify the 'ingredients' of a model, moving the industry away from unverified prose toward structured, interoperable metadata.