Data Provenance in Voxel51 3.0: Versioning and Activity Logs

When a model underperforms, can you tell exactly what data went into training it? Which samples were included, what annotations they contained, and what's changed since? As datasets evolve through continuous annotation, curation, and automated workflows, answering those questions becomes increasingly difficult.

With Voxel51 3.0, teams can use Voxel51 as a persistent system of record for multimodal data. Dataset versioning captures immutable snapshots of the exact data used for training and evaluation, while activity logs and sample-level history track how that data changes over time.

This matters as data work scales beyond a handful of engineers to dozens of annotators, reviewers, and automated pipelines working simultaneously. Voxel51 connects the data to the people and processes that change it, giving teams the history they need to investigate model failures, reproduce results, and understand how their datasets evolved.

Architecture diagram showing how FiftyOne's app, operators, and backend services send events through fiftyone-mq to Activity Core, which powers audit logs, annotation metrics, and job monitoring. — FiftyOne's event architecture uses fiftyone-mq for job scheduling and Activity Core for event tracking, providing a shared foundation for audit logs, annotation metrics, and job monitoring.

Key takeaways

Voxel51 3.0 adds data provenance so teams can trace exactly what data trained a model and what has changed since.
Data provenance in Voxel51 works at four levels: a deployment-level activity log, a dataset-level activity log, sample-level edit history, and dataset versioning.
Dataset versioning in Voxel51 captures immutable snapshots of a dataset, so teams can reproduce training and evaluation runs.
Voxel51 permissions and workflow visibility connect each data change to the person or automated job that made it.

Data provenance needs to exist at multiple levels

Data provenance isn't a single log. Different questions require different levels of history.

Deployment-level activity log

See activity across your Voxel51 deployment: who performed an action, when it happened, and what it affected. This gives administrators an organization-wide view of activity for operational oversight and audit requirements.

Dataset-level activity log

Trace activity within an individual dataset. Teams can understand when datasets changed and what actions contributed to their current state, rather than reconstructing that history from separate systems.

Level	What it tracks	Question it answers
Deployment-level activity log	Who performed an action, when, and what it affected, across the deployment	What happened across our organization?
Dataset-level activity log	Actions within a single dataset	When did this dataset change, and what changed it?
Sample-level edit history	Changes to a sample’s labels, metadata fields, and other properties	How did this sample reach its current state?
Dataset versioning	Immutable snapshots of a dataset	What did the dataset look like at training time?

The four levels of data provenance in Voxel51 3.0, with what each level tracks and the question it answers.

Voxel51's dataset activity log showing a chronological history of sample edits, annotation tasks, reviews, and automated updates, with user attribution, timestamps, and links to affected samples. — Track every change to a dataset with a centralized activity log. See who modified samples, what changed, and when, with filters for edits, workflows, management actions, and runs.

Sample-level edit history

Follow the history of an individual sample. When a label, metadata field, or other property changes, you can see how it evolved over time and investigate the specific changes that led to its current state.

Voxel51's sample history panel showing a chronological record of annotation submissions, label changes, automated updates, and reviews, with user attribution and timestamps. — Trace the history of individual samples, including annotation changes, reviews, and automated updates. See who made each change and when, making it easier to investigate data quality issues and maintain a clear audit trail.

Dataset versioning

Sometimes you don't just need to know what changed. You need to know exactly what the dataset looked like at a particular point in time.

Dataset versioning lets teams capture immutable snapshots of a dataset so they can compare historical states, reproduce workflows, and identify the exact data used for training or evaluation.

Know who and what changed your data

Provenance is most useful when you can connect changes back to the people and processes responsible for them. Voxel51's permissions give administrators control over who can modify datasets, and workflow visibility shows the automated jobs operating on them.

Together, these capabilities give teams the context to understand not just what changed, but who or what changed it.

Provenance connects people, pipelines, and data

Together, these capabilities turn Voxel51 into more than a place to explore datasets. They give teams a durable record of how data changes—from the individual sample to the entire dataset—and the context needed to investigate issues, reproduce workflows, and understand exactly what data went into a model.

The result is a system of record for the full data lifecycle.

FAQ

Data provenance in Voxel51 is a record of how your multimodal data changes over time and who or what changed it. It combines activity logs, sample-level edit history, dataset versioning, permissions, and workflow visibility.

Dataset versioning in Voxel51 captures immutable snapshots of a dataset at a specific point in time. Teams use these snapshots to compare historical states, reproduce workflows, and identify the exact data used for training or evaluation.

The deployment-level activity log in Voxel51 shows who performed an action, when it happened, and what it affected across the whole deployment. The dataset-level activity log traces activity within a single dataset.

Yes. Sample-level edit history in Voxel51 shows how a label, metadata field, or other property of an individual sample changed over time.

Voxel51 3.0 Adds Data Provenance to Become Your System of Record

Try Voxel51 for yourself

Key takeaways