{"course":{"slug":"fine-tuning","title":"Fine-Tuning, and When Not To","url":"https://addaly.com/learn/fine-tuning","level":"advanced","track":"deep"},"lesson":{"slug":"licences-and-what-you-may-train-on","title":"Licences, and what you may train on","url":"https://addaly.com/learn/fine-tuning/licences-and-what-you-may-train-on","position":8,"of":84,"minutes":10,"answer":"Model licence, data licence and downstream obligations are three separate permissions, the underlying copyright question is genuinely unsettled and differs by country.","keyIdea":"Model licence, data licence and downstream obligations are three separate permissions, the underlying copyright question is genuinely unsettled and differs by country, and per-example provenance recorded at collection time is the only thing that makes the question answerable later."},"body":{"format":"markdown","text":"## Three separate permissions, often confused\n\nBefore a fine-tune there are three questions, and people routinely answer one and assume the others.\n\n1. **May I use these weights, in this way?** That is the model licence.\n2. **May I train on this text?** That is the data licence, plus copyright law, plus whatever you promised your own users.\n3. **What obligations attach to the thing I produce?** That is the downstream clause, and it is the one that surprises people.\n\nNone 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.\n\n## Model licences, in three families\n\n**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.\n\n**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\".\n\n**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.\n\nThe 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.\n\n## Dataset licences are the sharper edge\n\nThe famous instruction datasets are not uniformly licensed, and the difference matters:\n\n- Some are CC-BY-SA, which is a share-alike licence: a copyleft obligation that arguably follows anything derived from it.\n- Some are explicitly non-commercial, including several popular sets generated from hosted models under terms that were non-commercial at the time.\n- 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.\n\nAnd scraped web text carries whatever rights its authors have, which is where the unresolved part begins.\n\n## The part that is not settled, and will not be settled by this lesson\n\nWhether training a model on copyrighted material without permission is lawful is contested, and the answer currently differs by jurisdiction:\n\n- 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.\n- 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.\n- The United Kingdom has a research-only TDM exception and has been consulting on change for years without settling it.\n- Japan's exception is comparatively broad; several other jurisdictions have nothing specific at all.\n\nSo 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.\n\n## Outputs of other models\n\nSeveral 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.\n\nThe 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.\n\n## What to write down, per example\n\nProvenance is cheap at collection time and impossible to reconstruct later. One extra column:\n\n```json\n{\"messages\": [...], \"source\": \"support_logs\", \"licence\": \"internal\",\n \"consent_basis\": \"tos_v4\", \"collected\": \"2026-08-14\"}\n```\n\nSix 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."},"check":{"question":"A team fine-tunes an Apache-2.0 base model on a mix of internal tickets and one popular public instruction dataset released under CC-BY-SA. Why does the permissive base licence not settle what they may do with the result?","options":[{"text":"Apache-2.0 requires derivative models to be released under Apache-2.0, which conflicts with the internal tickets being confidential.","correct":false,"tempting":"Misstates Apache-2.0, which is permissive and imposes no copyleft on derivatives; the confusion is with share-alike licences."},{"text":"The share-alike obligation on the training data is a separate permission, and it is the one constraining the output.","correct":true},{"text":"Fine-tuning voids a base model licence, because the resulting weights are legally a new work with no inherited permissions.","correct":false,"tempting":"Asserts a clean legal break that no licence grants and no court has established; obligations are generally analysed as following the material used."},{"text":"Public instruction datasets may only be used for research, so no commercial fine-tune can include one.","correct":false,"tempting":"Over-generalises from the subset of datasets that are non-commercial; several widely used sets are explicitly permissive."}]},"prev":{"slug":"the-cost-of-owning-a-model","title":"The cost of owning a model","url":"https://addaly.com/learn/fine-tuning/the-cost-of-owning-a-model"},"next":{"slug":"the-decision-record","title":"Writing the decision down","url":"https://addaly.com/learn/fine-tuning/the-decision-record"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Licences, and what you may train on', in Fine-Tuning, and When Not To. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/fine-tuning/licences-and-what-you-may-train-on (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Licences, and what you may train on","courseTitle":"Fine-Tuning, and When Not To","module":{"index":1,"slug":"deciding-whether-to-fine-tune","title":"Deciding whether to fine-tune at all"},"answer":null,"keyIdea":"Model licence, data licence and downstream obligations are three separate permissions, the underlying copyright question is genuinely unsettled and differs by country, and per-example provenance recorded at collection time is the only thing that makes the question answerable later.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"three-separate-permissions-often-confused","runs":[{"text":"Three separate permissions, often confused"}]},{"t":"p","runs":[{"text":"Before a fine-tune there are three questions, and people routinely answer one and assume the others."}]},{"t":"ol","start":1,"items":[[{"text":"May I use these weights, in this way?","b":true},{"text":" That is the model licence."}],[{"text":"May I train on this text?","b":true},{"text":" That is the data licence, plus copyright law, plus whatever you promised your own users."}],[{"text":"What obligations attach to the thing I produce?","b":true},{"text":" That is the downstream clause, and it is the one that surprises people."}]]},{"t":"p","runs":[{"text":"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."}]},{"t":"h2","id":"model-licences-in-three-families","runs":[{"text":"Model licences, in three families"}]},{"t":"p","runs":[{"text":"Genuinely permissive.","b":true},{"text":" 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."}]},{"t":"p","runs":[{"text":"Custom community licences.","b":true},{"text":" 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”."}]},{"t":"p","runs":[{"text":"Research or non-commercial.","b":true},{"text":" Some weights, and rather more datasets, are released for research only. A fine-tune of a non-commercial model is a non-commercial model."}]},{"t":"p","runs":[{"text":"The practical habit: open the "},{"text":"LICENSE","code":true},{"text":" 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."}]},{"t":"h2","id":"dataset-licences-are-the-sharper-edge","runs":[{"text":"Dataset licences are the sharper edge"}]},{"t":"p","runs":[{"text":"The famous instruction datasets are not uniformly licensed, and the difference matters:"}]},{"t":"ul","items":[[{"text":"Some are CC-BY-SA, which is a share-alike licence: a copyleft obligation that arguably follows anything derived from it."}],[{"text":"Some are explicitly non-commercial, including several popular sets generated from hosted models under terms that were non-commercial at the time."}],[{"text":"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."}]]},{"t":"p","runs":[{"text":"And scraped web text carries whatever rights its authors have, which is where the unresolved part begins."}]},{"t":"h2","id":"the-part-that-is-not-settled-and-will-not-be-settled-by-this","runs":[{"text":"The part that is not settled, and will not be settled by this lesson"}]},{"t":"p","runs":[{"text":"Whether training a model on copyrighted material without permission is lawful is contested, and the answer currently differs by jurisdiction:"}]},{"t":"ul","items":[[{"text":"In the United States the argument runs through fair use, and courts have reached different conclusions on different facts; the question of whether the "},{"text":"source","i":true},{"text":" of the copies was lawful has mattered in at least one ruling as much as the training itself."}],[{"text":"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."}],[{"text":"The United Kingdom has a research-only TDM exception and has been consulting on change for years without settling it."}],[{"text":"Japan’s exception is comparatively broad; several other jurisdictions have nothing specific at all."}]]},{"t":"p","runs":[{"text":"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."}]},{"t":"h2","id":"outputs-of-other-models","runs":[{"text":"Outputs of other models"}]},{"t":"p","runs":[{"text":"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."}]},{"t":"p","runs":[{"text":"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."}]},{"t":"h2","id":"what-to-write-down-per-example","runs":[{"text":"What to write down, per example"}]},{"t":"p","runs":[{"text":"Provenance is cheap at collection time and impossible to reconstruct later. One extra column:"}]},{"t":"code","lang":"json","label":"JSON","text":"{\"messages\": [...], \"source\": \"support_logs\", \"licence\": \"internal\",\n \"consent_basis\": \"tos_v4\", \"collected\": \"2026-08-14\"}"},{"t":"p","runs":[{"text":"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."}]}],"check":{"question":[{"text":"A team fine-tunes an Apache-2.0 base model on a mix of internal tickets and one popular public instruction dataset released under CC-BY-SA. Why does the permissive base licence not settle what they may do with the result?"}],"options":[{"text":[{"text":"Apache-2.0 requires derivative models to be released under Apache-2.0, which conflicts with the internal tickets being confidential."}],"correct":false,"tempting":[{"text":"Misstates Apache-2.0, which is permissive and imposes no copyleft on derivatives; the confusion is with share-alike licences."}]},{"text":[{"text":"The share-alike obligation on the training data is a separate permission, and it is the one constraining the output."}],"correct":true,"tempting":null},{"text":[{"text":"Fine-tuning voids a base model licence, because the resulting weights are legally a new work with no inherited permissions."}],"correct":false,"tempting":[{"text":"Asserts a clean legal break that no licence grants and no court has established; obligations are generally analysed as following the material used."}]},{"text":[{"text":"Public instruction datasets may only be used for research, so no commercial fine-tune can include one."}],"correct":false,"tempting":[{"text":"Over-generalises from the subset of datasets that are non-commercial; several widely used sets are explicitly permissive."}]}],"order":[0,3,2,1],"why":null},"prev":{"slug":"the-cost-of-owning-a-model","title":"The cost of owning a model","minutes":9},"next":{"slug":"the-decision-record","title":"Writing the decision down","minutes":8}}}