Licences, and what you may train on

Three separate permissions, often confused

Before a fine-tune there are three questions, and people routinely answer one and assume the others.

1. May I use these weights, in this way? That is the model licence.
2. May I train on this text? That is the data licence, plus copyright law, plus whatever you promised your own users.
3. What obligations attach to the thing I produce? That is the downstream clause, and it is the one that surprises people.

None of what follows is legal advice, and I am not in a position to give it. What I can do is show you the shape of each question so that you know which one you are in and what to hand a lawyer.

Model licences, in three families

Genuinely permissive. Apache-2.0 and MIT. Qwen’s recent releases, Mistral’s open models, OLMo, SmolLM, many others. You may fine-tune, merge, serve, sell and relicense the result. Apache-2.0 carries a patent grant and a notice requirement; that is essentially it. If you want the fewest questions, start here.

Custom community licences. Llama’s licence is the best-known: broadly free to use, with an acceptable-use policy, a naming requirement for derivatives, and a user-count threshold above which you must ask for separate terms. Gemma has its own terms with a use policy that follows the weights downstream. These are usable and widely used, but they are not open source in the OSI sense, and “open weights” is not a synonym for “do what you like”.

Research or non-commercial. Some weights, and rather more datasets, are released for research only. A fine-tune of a non-commercial model is a non-commercial model.

The practical habit: open the LICENSE file for the exact checkpoint you downloaded, not the family. Licences differ between sizes and between base and instruct variants more often than you would expect.

Dataset licences are the sharper edge

The famous instruction datasets are not uniformly licensed, and the difference matters:

• Some are CC-BY-SA, which is a share-alike licence: a copyleft obligation that arguably follows anything derived from it.
• Some are explicitly non-commercial, including several popular sets generated from hosted models under terms that were non-commercial at the time.
• Some are Apache-2.0 or CC0 and cause nobody any trouble. Databricks’ Dolly-15k and the No Robots set were both built specifically to be clean on this point, which is why they remain useful despite being small.

And scraped web text carries whatever rights its authors have, which is where the unresolved part begins.

The part that is not settled, and will not be settled by this lesson

Whether training a model on copyrighted material without permission is lawful is contested, and the answer currently differs by jurisdiction:

• In the United States the argument runs through fair use, and courts have reached different conclusions on different facts; the question of whether the source of the copies was lawful has mattered in at least one ruling as much as the training itself.
• The European Union has a text-and-data-mining exception with an opt-out that rightsholders can exercise, plus transparency obligations on general-purpose model providers under the AI Act.
• The United Kingdom has a research-only TDM exception and has been consulting on change for years without settling it.
• Japan’s exception is comparatively broad; several other jurisdictions have nothing specific at all.

So the honest summary is: this is live litigation and live legislation, the answers differ by country and by facts, and anybody who tells you it is simple is describing their preference rather than the law. What you can do is reduce your exposure: prefer sources with explicit permission, record provenance per example so you can answer the question later, and escalate the commercially significant cases to counsel rather than to a forum.

Outputs of other models

Several hosted providers’ terms restrict using their outputs to develop competing models. The exact wording varies, the scope of “competing” is not obvious, and some providers have loosened these terms while others have not. If your distillation plan depends on a particular provider’s outputs, read that provider’s current terms and get an opinion.

The clean way around it is to distil from an Apache-2.0 open-weight teacher you run yourself. A 70B model rented for a day produces a large dataset, and a permissive licence imposes no downstream restriction on what you do with it.

What to write down, per example

Provenance is cheap at collection time and impossible to reconstruct later. One extra column:

{"messages": [...], "source": "support_logs", "licence": "internal",
 "consent_basis": "tos_v4", "collected": "2026-08-14"}

Six months on, when somebody asks whether the model was trained on customer data and under what basis, that column is the difference between an answer and an investigation.