Rights and contracts
Dataset purpose limitation metadata: bind permitted uses to each release
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
To associate permitted-use metadata with a delivered dataset, publish a policy record for each release that names the agreement, the permitted and prohibited purposes, the term and the duties, and bind it to the exact files by checksum and to every record by a policy pointer. A copy then carries its own terms, and any record can be traced to the release and clause that allow its use.
License terms
Dataset purpose limitation
Key takeaways
- A signed agreement in a contract folder does not travel with the files; the permitted uses have to be written into the delivery itself.
- Bind the terms at three levels: a policy record per release, a checksum for every file, and a policy pointer on every record.
- Express purposes as a structured policy with permissions, prohibitions, duties and a stated conflict rule, not as a free-text license sentence.
- Any change to the records, the terms or the excluded population produces a new release and a new policy record, never an edit in place.
- A purpose field records what the agreement allows; it does not create privacy permission, ownership or a licensing right on its own.
How do you attach permitted uses to a delivered dataset?
You attach them by shipping a machine-readable policy record inside every release and binding it to the release in two directions: the policy names the exact files by checksum, and each record names its policy by identifier. The agreement stays the legal source of truth, but the delivery now carries a faithful, versioned statement of what that agreement permits.
This matters because datasets move. A buyer's platform team copies files into object storage, a research group pulls a subset into a notebook, and an evaluation team merges two deliveries into one table. Each of those steps strips away the email thread and the PDF. If the permitted uses are not in the data, the next person to touch it has to guess.
The method below is written for the supplier, the business that owns the records and signs the license. It assumes the purposes have already been negotiated and approved; it does not decide what uses should be allowed. Its job is to make sure that whatever was agreed is still legible at the moment someone opens a file a year later.
Why a contract folder stops protecting the terms once files leave it
A contract folder protects terms only for people who know to open it, and the people handling training data rarely do. The link between a file and its agreement usually lives in someone's memory, a ticket or a folder name. None of those survive a copy to a different bucket.
Three failure patterns show up again and again. The first is the orphaned subset: a filtered extract no longer carries the delivery name, so nobody can say which agreement governs it. The second is the silent merge: records from an evaluation-only release are combined with a training release, and the stricter terms disappear. The third is the stale amendment: the agreement was amended to exclude a customer segment, but the files already distributed still contain those records and nothing in them says so.
Each pattern has the same root cause. The terms were attached to the transaction, not to the artifact. The fix is to make the artifact self-describing at the level where copies actually happen, which is the file and the record.
Three places the permitted uses have to live
Put the full terms once at release level, prove file identity with checksums, and give each record a short pointer back to the policy. Repeating the full terms on every row wastes space and drifts; keeping them only at release level loses them the moment a subset is cut.
Existing vocabularies already cover most of this. DCAT 3, the W3C catalog vocabulary, lets a dataset or distribution point to an ODRL policy through its has policy property and lets a distribution carry a checksum so a recipient can confirm the file has not changed; it also adds version and previous version properties. The Croissant RAI extension from MLCommons gives a slot, rai:dataUseCases, for the intended uses and the purposes for which data was collected. Using these names makes the record readable by catalog and ML tooling instead of only by your own scripts.
| Level | What it holds | Standard term to reuse | What breaks if it is missing |
|---|---|---|---|
| Release | Agreement reference, permitted and prohibited purposes, term, territory, duties, conflict rule | DCAT has policy pointing to an ODRL Agreement | Nobody can state the terms without the contract folder |
| File | Path, byte size, record count, SHA-256 digest, release identifier | DCAT checksum; Croissant FileObject sha256 | A modified or substituted file still looks covered |
| Record | Record identifier, policy identifier, exclusion flags | Field in the record schema (editorial) | A subset or merge loses its terms |
| Intended use note | Plain-language purpose and limitations for people | Croissant rai:dataUseCases and rai:dataLimitations | Reviewers read the code values without context |
| Release chain | Previous release and reason for the change | DCAT previous version; Croissant version | Amendments cannot be told apart from errors |
The permitted-use binding example, field by field
The record below is the asset this page exists to provide: a release-level policy record that a team can copy, rename and fill. Field names are illustrative and follow no single standard; the mapping column shows which published term each one corresponds to, so the record can be serialized as JSON-LD for a catalog or kept as a flat JSON file next to the data.
At a glance
The permitted-use binding example, field by field
Every value shown is invented. The identifiers use the urn:example namespace on purpose, so nobody mistakes them for a live agreement.
- Record-level pointer, illustrative: record_id TCK-000184, policy_uid urn:example:policy:support-tickets:r3, exclusion_flags none.
- Plain-language use note, illustrative: support tickets from a U.S. software company, de-identified before delivery, licensed for model training and evaluation; not for re-identification or resale of raw records.
- Store the policy record as policy.json at the root of the release and list it in the manifest with its own checksum, so the terms cannot be swapped without detection.
| Field | Illustrative value | Maps to | Validation rule |
|---|---|---|---|
| policy_uid | urn:example:policy:support-tickets:r3 | ODRL uid | Unique; never reused across releases |
| policy_type and parties | Agreement between urn:example:party:supplier and urn:example:party:licensee | ODRL Agreement, assigner and assignee | Agreement once signed, with both parties; an Offer grants nothing |
| agreement_ref | Data license schedule B, clause 4, as amended 2026-09-30 | ODRL Agreement metadata (editorial) | Must name a signed document and amendment date |
| permitted_purposes | model training; model evaluation | ODRL permission with a purpose constraint | Values from a versioned purpose list; no free text |
| prohibited_purposes | re-identification; redistribution of raw records | ODRL prohibition | At least one entry, or an explicit none |
| conflict_rule | prohibit | ODRL conflict | Must be set; the ODRL default voids the policy on conflict |
| term_and_territory | 2026-10-01 to 2029-09-30; United States and EU | ODRL constraints (dateTime, spatial) | End date and territory both present |
| duties | delete on termination; report suspected re-identification | ODRL duty; delete action | Each duty names a trigger |
| files | tickets-2019-2025.parquet with SHA-256 digest and record count | DCAT distribution with checksum | Digest recomputed on receipt must match |
| previous_release | urn:example:policy:support-tickets:r2, reason: amendment removed one customer segment | DCAT previous version | Required for any release after the first |
Write the permitted uses as a policy, not a sentence
A license sentence tells a lawyer what was agreed; a policy tells a pipeline. W3C's ODRL information model is built for this: a policy has a unique identifier and at least one permission, prohibition or obligation, each rule applies an action to a target asset, and constraints narrow when the rule applies.
Three ODRL details change how you fill the record. First, the policy type carries meaning. An Offer makes terms available but grants nothing, while an Agreement grants the rules between a named assigner and assignee, so a release should only ship with an Agreement. Second, ODRL's vocabulary includes a purpose left operand, defined as a defined purpose for exercising the action, which is the natural home for training versus evaluation. Third, the conflict property decides what happens when a permission and a prohibition overlap: perm lets permissions win, prohibit lets prohibitions win, and if nothing is set the default is invalid, which voids the whole policy.
For licensed business records, set conflict to prohibit. It matches how a supplier reads the agreement, where an exclusion beats a general grant, and it avoids a policy that silently becomes void. Use the ODRL delete action, defined as permanently removing all copies after use, to express a deletion duty on termination, and give it a trigger so it can be checked.
Tracing one record back to its terms: an illustrative release
A claim about a record is traceable when a reviewer can walk from the record to the policy, from the policy to the file digest, and from the policy to a named clause, without asking anyone. The case below is fictional and every number in it is an illustrative count.
Starting state: a software company delivers release r3 of its de-identified support tickets, 6 Parquet files and about 410,000 records, under an agreement that permits training and evaluation. Release r2 had been withdrawn after an amendment excluded one enterprise customer's tickets. Six months later, the licensee's evaluation team asks whether ticket TCK-000184, found in a merged evaluation table, may be used in a published benchmark.
The reviewer reads the record's policy pointer, opens policy.json for r3, recomputes the SHA-256 digest of the source file and gets a match, and sees that permitted purposes include evaluation but prohibited purposes include redistribution of raw records. The agreement reference points to schedule B, clause 4, as amended. Decision: the record may be used to run the evaluation, but the raw ticket text may not be published inside the benchmark. Outcome check: the evaluation team's release notes cite the policy identifier, and a later audit finds no raw r3 text in the public benchmark files.
Counterexample: a second merged table holds 2,300 records whose only metadata is a license column reading internal use. There is no policy pointer, no release identifier and no digest that matches any delivered file. Some of those records may come from the withdrawn r2. They fail the trace and are excluded, even if they were probably licensed, because nobody can show which terms apply to them.
Which changes force a new release of the policy record
Any change to the records, the terms or the excluded population needs a new release with its own policy identifier; only a change in packaging can keep the old one. Editing a policy in place breaks every trace that already cites it.
Croissant's versioning rules give a useful model. Under its MAJOR.MINOR.PATCH scheme, a PATCH change leaves the data the same but serialized differently, a MINOR change adds data while old records remain retrievable as before, and a MAJOR change means existing data was edited, removed or reshuffled. Croissant also lets each file carry a SHA-256 checksum. Map those levels onto rights changes as follows.
| Change | Example | New policy record? | What to record |
|---|---|---|---|
| Re-serialization only | Parquet files re-exported as JSON Lines | No; patch version, same policy | New file digests listed under the same policy |
| Records added under the same terms | A further quarter of tickets | Yes, new release identifier | Previous release and the added record range |
| Records removed | Opt-outs or a customer segment excluded by amendment | Yes, major version | Removed record identifiers and the amendment date |
| Purpose or term changed | Evaluation-only becomes training and evaluation | Yes, new Agreement | New agreement reference; old release keeps old terms |
| Error correction | Mislabelled field fixed in place | Yes, major version | Reason for change and affected record identifiers |
A purpose field is a record of terms, not a permission
Filling permitted_purposes with model training does not make training lawful or licensed; it only records what the agreement says. Four separate questions still sit behind every value: whether the team had access to the records, whether privacy law permits the new use, who owns the records and any rights in them, and what the signed contract grants.
In the EU, GDPR Article 5(1)(b) requires personal data to be collected for specified, explicit and legitimate purposes and not further processed in a way incompatible with them. That is a test applied to the processing, and metadata cannot pass it on anyone's behalf. De-identification can change the privacy analysis, but it does not by itself create a right to license the records. Counsel should review the specific facts before a purpose value is written into a release that will be used outside the company.
The practical rule for the record is narrow: never write a purpose that is not in a signed agreement, and never widen a purpose in metadata to make a pipeline run.
Bringing a bound release to a licensing conversation
SourceX supports a company through rights review, contracting, preparation, delivery and payment when it licenses selected records as an AI dataset. A supplier that already keeps release-level policy records and record pointers can show exactly how a licensed dataset would carry its terms, which shortens the rights conversation.
Under the SourceX process the company approves scope and terms and signs before any data is delivered, and the de-identified dataset is then used only as that agreement allows. Finding out whether a company qualifies starts with a short assessment and no upload of records.
Frequently asked questions
What is dataset purpose limitation metadata?
It is structured metadata that records which purposes a dataset may be used for, which are prohibited, under which agreement and for how long. It is useful only when it is bound to the actual files and records, so a copy carries its terms. It describes the agreement; it does not replace the legal analysis behind it.
Can ODRL express purposes such as model training or evaluation?
Yes. ODRL's vocabulary defines a purpose left operand that constrains when a rule applies, so a permission can be limited to a named purpose. Training and evaluation are not built-in values, so keep them in your own versioned purpose list and reference it consistently.
Is a license field on a dataset card enough?
Usually not for licensed business records. A single license field cannot express prohibited purposes, duties, term and territory, or the fact that one release replaced another. Use it as a pointer to the full policy record rather than as the terms themselves.
What should happen to records whose policy cannot be traced?
Treat them as unlicensed for the purpose in question and exclude them until the trace is repaired. A record without a policy pointer, release identifier and matching file digest cannot be shown to fall under any agreement, even if it probably does.
Do I need a new policy when records are removed after an opt-out?
Yes. Removing records changes the dataset, so publish a new release with its own policy identifier, record which record identifiers were removed and when, and link it to the previous release. The earlier release keeps its original policy so past uses can still be explained.
Sources
- An ODRL policy has a unique identifier and at least one permission, prohibition or obligation; an Agreement grants rules between assigner and assignee, an Offer grants none, and an unset conflict strategy defaults to invalid. Source
- The ODRL vocabulary defines a purpose constraint operand and a delete action meaning permanent removal of all copies after use. Source
- DCAT 3 lets a resource point to an ODRL policy, lets a distribution carry a checksum to show the file has not changed, and adds version and previous version properties. Source
- Croissant 1.1 defines a dataset license property, SHA-256 checksums on files and MAJOR.MINOR.PATCH versioning in which edited or removed records require a major version. Source
- Croissant RAI records intended uses and collection purposes in rai:dataUseCases and limits in rai:dataLimitations. Source
- GDPR Article 5(1)(b) requires personal data to be collected for specified, explicit and legitimate purposes and not further processed incompatibly with them. Source
See if your company qualifies
A short company assessment. No data uploads are needed.