Gated and custom-licensed datasets on model hubs: reading the terms before training

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

What a gate actually does, and what it does not

A gate is an access-control mechanism; it does not grant or certify any rights by itself. On Hugging Face, a gated repo shows a prompt such as "You need to agree to share your contact information to access this dataset," and the files become available only after the request is accepted [1]. The same Structured3D page then limits use to non-commercial research and education, which is the part that actually governs training [1].

The dataset card's YAML usually carries the gate configuration (extra_gated_prompt, extra_gated_fields, often a checkbox such as "I agree to use this dataset for non-commercial use only") and a license: field. When that field says other, the real terms live somewhere else: a LICENSE file in the repo, a paragraph in the README, a PDF agreement, or an external website. Those terms, not the green "access granted" banner, define your rights.

The gap between what a hub page says and what the data's terms actually are is well documented. The Data Provenance Initiative audited more than 1,800 text datasets and reported license omission rates above 70% and license error rates above 50% on popular hosting sites [2]. Treat the hub's license tag as a lead to verify, not a finding.

Common patterns in gated terms

Most gated research datasets restrict commercial use, and the restriction often reaches further than engineers expect. Four patterns recur across hub pages:

• Restated upstream terms. A gated ImageNet derivative restates the ImageNet terms: non-commercial research and educational use only, with the obligations extending to the researcher's employer [3][4]. A re-packaging does not loosen the original agreement.
• Contact-first commercial clauses. Some gates allow research access but require commercial users, explicitly including those training AI models, to contact the owner before any such use [5]. Accepting the gate does not cover the commercial case.
• Written-permission prohibitions. Challenge and benchmark datasets frequently bar commercial use without explicit written permission [6], which matters if your eval results feed product decisions.
• Custom model-style licenses. Bespoke licenses can reach downstream artifacts. The Llama 3.1 Community License, for example, attaches conditions to using outputs to create, train or fine-tune other AI models, including a naming requirement [7].

Not every custom license is restrictive. The Copernicus Sentinel data licence grants free, full and open access with broad reuse rights, subject to a liability waiver and an attribution requirement [9]. The point is that "custom" means you must read it; it does not signal either answer. For standard licenses such as CC BY or ODC-BY, use the open data license compatibility matrix instead.

Who in your company can accept the terms

Hub access requests are made from a personal account, so the person who clicks becomes the visible party to the terms [1]. Yet many gated agreements, like the ImageNet-style terms above, purport to bind the user's employer as well [3][4].

That creates two failure modes. An engineer with no signing authority may commit the company to non-commercial or attribution obligations it never reviewed. Conversely, a colleague who did not click may pull the same files from a shared bucket, outside any agreement the licensor recognizes.

A workable internal rule separates three cases:

1. Engineer may accept alone: terms limited to research or evaluation inside a sandbox, with no commercial model, no redistribution, and no clauses binding the employer beyond confidentiality.
2. Counsel or delegated reviewer approves first: any "license: other", any clause naming the employer, indemnities, governing-law or audit clauses, or any intended use in a model that will ship.
3. Do not accept; escalate to procurement: commercial training is prohibited or requires owner contact, or the terms require a signed agreement. That path leads to a negotiated license; see research-only datasets: getting commercial rights.

The custom dataset license review checklist

Each gated dataset needs answers to the same ten questions before it reaches a training run. Fill this in per dataset and store it with the run's data manifest.

Illustrative example: invented to show structure; it does not describe an available dataset.

Questions 4, 5 and 8 decide most commercial outcomes. If termination requires deleting "derivative works," ask counsel whether the licensor could argue that includes trained weights; the companion page on what happens to trained models when a data license ends covers that negotiation.

Recording acceptance so it survives an audit

An acceptance you cannot reproduce is close to no acceptance at all. Access approved through a gate can later be withdrawn by the owner [10], and terms pages can change after you clicked, so capture the evidence at the moment of access.

Illustrative example: invented to show structure; it does not describe an available dataset.

dataset_acceptance:
  repo_id: example-org/example-gated-corpus
  revision: 3f2c9e1            # commit hash pinned in the training config
  license_field: other
  terms_snapshot: s3://legal-evidence/datasets/example-gated-corpus/terms-2026-10-09.pdf
  terms_sha256: 9b1d...e4
  accepted_by: hub-user-jdoe   # individual account that clicked
  accepted_at: 2026-10-09T14:22:00Z
  approver: counsel-ticket-LEGAL-4182
  permitted_use: internal evaluation only
  restrictions: [no commercial training, no redistribution, attribution in paper]
  downstream_runs: [eval-suite-v7]

Pin the dataset revision rather than main, because a later commit can change both files and README terms. Link this record to your data lineage so any model card can trace back to the exact terms in force. The broader due diligence steps are in the AI training data due diligence checklist, and the open dataset license audit covers ungated sources.

When terms change after you have access

Assume the terms can move and plan a re-check. A dataset owner can edit the README, swap the LICENSE file, or tighten the gate prompt on a new commit, and an existing grant does not alert you. Your pinned snapshot is your evidence of what applied when you trained.

Regulators have also noticed unilateral terms changes, though from the other direction. FTC staff warned in February 2024 that companies quietly adopting more permissive data practices, such as AI training, through terms-of-service changes could be acting unfairly or deceptively [8]. That staff post is not a rule, but it reinforces a practical point: the version of the terms in force at collection and at use both matter.

Set a review trigger in your data catalog: when a pinned dataset's upstream repo gets a new commit touching README, LICENSE or the gate fields, flag every model trained on it for a terms diff. If the new terms are stricter, counsel decides whether your earlier acceptance still governs.

Gated research data versus a negotiated license

When the checklist turns up "non-commercial," "contact the owner," or "written permission required," the gate has done its job: it tells you that a separate commercial license is needed. Click-through terms rarely include warranties, indemnities or deletion terms you can negotiate; the ImageNet agreement, for instance, disclaims warranties outright [3]. For production SFT or eval data, buyers usually need a bilateral agreement that states records, uses, term and delivery.

At that point the options are to approach the owner, replace the dataset, or source comparable operational data under a negotiated license. The AI data license negotiation checklist and the licensing hub cover what to ask for. If you need operational data from US companies rather than research corpora, you can describe the data you need to SourceX: it sources such datasets on request, with each release approved by the supplying company, and a request does not guarantee a match.

Sourcing licensed data instead of a gated dataset

SourceX sources operational datasets from US companies, such as support and sales histories, engineering records and documents, and manages the licensing process. Every dataset is rights-reviewed and delivered under a license that defines records, uses, term and delivery, and nothing is contracted until a supplier agrees. Describe the data you need at sourcex.si/buyers.