A supplier says its model is open and available for commercial use. The buyer still needs to establish which model, which version, which licence and which intended activity the statement covers. Running an internal assistant, distributing a modified model and training another model on generated output can raise different contractual questions.
Training data provenance adds another layer. A model licence describes rights granted over specified materials. It does not, by its label alone, document every dataset used to create the model or settle the buyer's rights over its own fine-tuning and retrieval inputs. Procurement needs an evidence map that keeps those layers visible.
Build an inventory of components and uses
Start with the proposed deployment rather than the supplier's product category. List the base model, any adapters or fine-tuned weights, inference software, training datasets, retrieval sources and third-party services. Record the version or revision of each material component and who supplies it.
Then list the activities the institution plans to perform. Will it host the weights itself, provide an employee service, expose an API to clients, distribute a derivative or use output to train another system? An answer that is sufficient for one activity may not cover another. The inventory gives legal, data and technology teams the same factual object to assess.
Hugging Face's model card documentation describes metadata for licences, base models and training datasets. Those fields are useful discovery aids. They are statements attached to a repository, not a substitute for examining the referenced licence and underlying evidence. Capture the revision the institution evaluated so a later edit does not silently change the record.
Read the actual grant and its conditions
Different licences can govern downloadable models. The Apache License 2.0 text includes copyright and patent grants, redistribution conditions, and warranty and liability provisions. That does not mean every component packaged with a model is covered by that licence. The buyer needs to establish which materials the licensor has placed under it.
Meta's Llama 3.3 Community License is a specific example of a model licence with its own redistribution, attribution and acceptable-use conditions. Its text also addresses certain uses of model outputs to improve another distributed model. The point for procurement is to examine the chosen model's actual terms rather than treating all available weights as one licensing category.
Keep licence interpretation with the responsible legal team. The operational contribution is a precise description of deployment, distribution, dependencies and data movement. A legal review based on an internal prototype can become stale when a product team later adds client access or ships an adapter to a partner.
Provenance needs levels of evidence
A vendor may disclose complete dataset manifests, broad source categories or very little about pretraining. Record the level provided and the gaps. A list of source categories is different from document-level provenance, and neither should be represented as a guarantee about every training item.
Ask what the supplier can evidence about data acquisition, applicable terms, filtering, exclusions and rights-related complaints. Separate documented practices from contractual promises and independent findings. If a supplier cannot disclose material details, the institution needs an explicit decision about whether the remaining uncertainty fits the proposed use.
The buyer's own data deserves equal attention. A market-data contract might permit employee viewing while restricting redistribution, derived products or model training. Documents licensed for a research team may not be available to every business unit through an assistant. The fact that the institution can access a file does not establish every intended reuse.
A hypothetical procurement that changes scope
Hypothetical example: an asset manager tests an internally hosted model using public filings and licensed research. The prototype summarises documents for a small analyst group. After a successful trial, the product team proposes a client-facing service and a fine-tuned model that incorporates examples drawn from the research archive.
The initial assessment cannot simply be copied across. The audience has changed, the processing activity has changed and the role of the licensed research has changed. The buyer needs to check the relevant source contracts, model terms and supplier commitments against the revised use. A technical ability to export the weights answers none of those scope questions.
One workable choice might be to keep licensed research in a restricted retrieval layer and exclude it from training, if the source terms support that design. Another might be to obtain additional rights or narrow the product. Each choice changes cost, model performance and the ease with which data can later be removed. The decision needs both commercial and technical input.
Retrieval permissions also need an audience boundary. A source contract supporting one analyst group does not become an organisation-wide grant when its text enters a shared search index. Store source identity and access rules with the retrieved material so a later application can enforce the same scope.
Trace changes after the purchase
Assign an owner to the component and rights inventory. A new base model, a merged adapter, a dataset refresh or a change in distribution can trigger review. Make the trigger part of ordinary release management so that it does not depend on someone remembering an earlier procurement conversation.
Preserve licence texts, supplier responses and data-source agreements with their applicable versions. Ask how the supplier will notify the institution about a claim, a changed term or a component withdrawal. The institution should understand whether a notice requires investigation, a replacement model or suspension of a specific use.
Indemnities, warranties and liability limits need a factual reading alongside provenance. Who gives the commitment? What product and use does it cover? What exclusions apply when the buyer fine-tunes, combines sources or ignores deployment restrictions? A broad marketing statement cannot answer a narrow allocation of responsibility.
Price the cost of incomplete evidence
More disclosure does not automatically make one supplier the correct choice, and limited disclosure does not establish wrongdoing. The buyer is deciding what uncertainty it can accept for a defined activity. A low-impact internal drafting tool and a client product built around proprietary source material may justify different requirements.
Assess alternatives in operational terms. Replacing a model may require new evaluation, additional computing capacity and rewritten integration logic. Removing a retrieval source may be straightforward if the system records source relationships; removing its influence from trained weights is a different task. Procurement should ask for demonstrations proportionate to the exit it expects to use.
The distinction between access and reuse is developed in digital asset market data licensing. The need for usable records follows the same logic as vendor retention and export planning. An enterprise AI purchase is easier to defend when the institution can show what it acquired, what evidence supported the decision and which future uses require another review.