Research guide

Public vs. Licensed World-Model Data

How open research datasets differ from commercially licensed training data, and why download access is not resale permission.

For
Procurement, legal, data partnerships, and research leadership
Reviewed
July 24, 2026

Direct answer

Public access describes availability. A license describes granted rights. Commercial training, redistribution, customer delivery, and underlying game content may each require separate authority.

Open does not mean unrestricted

A repository can be publicly visible while its files remain gated, non-commercial, share-alike, research-only, or governed by separate platform terms. Read the dataset card, attached license, source-media terms, and access agreement together.

Code and data frequently use different licenses. An Apache-licensed loader does not make the media it downloads commercially reusable.

Separate listing from supply

A research directory can summarize and link to an external dataset without copying or selling it. The page must identify the creator, host, license, restrictions, and external access path.

A commercial listing requires a stronger record: rights holder, granted uses, customer sublicensing or access terms, territory, duration, attribution, withdrawal process, and evidence supporting each claim.

Review underlying content

Dataset authors may own annotations while third parties own video, game assets, trademarks, music, player likeness, or platform-hosted files. The license cannot grant rights the licensor does not control.

For gameplay data, distinguish private instrumented capture performed with publisher authority from public internet video referenced for research. They are different products even when their schemas look similar.

Use a fail-closed rights registry

Treat unreviewed material as unavailable for commercial delivery. A practical workflow moves records from discovered to quarantined, rights review, approved or rejected, ingested, released, and withdrawn.

Keep immutable evidence for the decision, including source terms, correspondence, license version, review owner, and date. Re-review when upstream terms or dataset versions change.

Bring us the training target.

We will map it to the observations, actions, state, outcomes, rights, and delivery evidence the dataset must contain.

Request access