Guide
Research-Only vs Commercial Dataset Licenses
Short answer
A research-only dataset license permits academic or non-commercial use but prohibits training a model intended for commercial deployment. A commercial license explicitly grants that right. A buyer-owned commercial license goes further: it is perpetual and irrevocable, with the resulting trained model owned by the buyer rather than licensed back from the provider.
Last reviewed:
What separates a research-only license from a commercial one?
A research-only license permits use for academic study, benchmarking, and non-commercial research, but says nothing grants the right to train a model that will ship in a commercial product. A commercial license explicitly extends that permission. The distinction is not about data quality or how well a dataset covers a skill — it is a legal grant that either exists in the license text or does not.
This is general information about how dataset licenses are typically structured, not legal advice. Confirm the current license text of any specific dataset directly before relying on it for a commercial roadmap.
Does training a model count as a restricted use?
Common non-commercial license families (such as Creative Commons NonCommercial variants) prohibit commercial use of the licensed material itself, and NoDerivatives variants separately prohibit distributing modified or derivative versions. Training a model intended for commercial deployment is generally treated as a commercial use of the underlying data under an NC term, independent of how the resulting model is later licensed. Whether a trained model itself is a "derivative work" of the training data is a separate and unsettled question that varies by jurisdiction — which is exactly why relying on a research-only or NC/ND-licensed dataset for a commercial model carries risk that surfaces after deployment, not before.
Research-only vs commercial vs buyer-owned, compared
| License type | Commercial training | Redistribution of derivatives | Model ownership |
|---|---|---|---|
| Research-only / non-commercial (e.g. CC BY-NC) | No | Attribution required, non-commercial only | N/A — commercial training not permitted |
| Non-commercial, no-derivatives (e.g. CC BY-NC-ND) | No | No derivatives, even non-commercial | N/A — commercial training not permitted |
| Commercial dataset license (per agreement) | Yes, per the specific terms | Per the specific terms | Per the specific terms — often licensed, not owned |
| Firsthand buyer-owned commercial license | Yes | Buyer-owned; derived models are the buyer’s | Perpetual and irrevocable, owned by the buyer |
General license categories, not a specific dataset’s current terms. See /compare/public-dataset-licenses for how named public datasets are positioned.
Why does the license type matter for a commercial roadmap?
A model trained on research-only data can perform well in evaluation and still be unshippable, because the license was never structured to permit the commercial use the team ultimately needs. That risk surfaces after the model is already built and integrated into a product, which is the most expensive point to discover it. Confirming commercial rights and documented consent before training, not after, is what avoids that outcome.
Firsthand ships a perpetual, buyer-owned commercial license with consent artefacts and a data card per batch, so this question is settled before a single hour is captured. See /provenance for the licensing and consent chain.
FAQ
Research-Only vs Commercial Dataset Licenses, answered.
01Can I use a research-only dataset to pretrain a model I later fine-tune commercially?
Generally no — a research-only or non-commercial license restricts commercial use of the data itself, and pretraining a model that will eventually be deployed commercially is typically treated as a commercial use regardless of the later fine-tuning step. Confirm the specific license text before relying on this.
02Is a trained model a "derivative work" of its training data?
It is a contested legal question that varies by jurisdiction and by the specific license, not a settled rule. Because it is unsettled, the safer path for a commercial product is data whose license explicitly grants commercial training rights, rather than relying on an argument about derivative status.
03What is the difference between a commercial license and a buyer-owned commercial license?
A commercial license grants permission to use the data commercially under whatever terms the specific agreement sets — which can still be limited or revocable. A buyer-owned commercial license, like Firsthand’s, is perpetual and irrevocable, and the resulting trained model is owned by the buyer rather than licensed back.
04Does a strict license matter less if the dataset’s content quality is excellent?
No. License restrictions are independent of data quality. A dataset can have strong coverage of exactly the skill a team needs and still be fully blocked for commercial use by its license terms.
05How does this relate to the validated-hours vs raw-hours distinction?
They are separate axes. Validated vs raw is about whether an hour of footage has passed quality checks; research-only vs commercial is about what the license permits doing with that hour once it is validated. A validated hour under a research-only license still cannot be used to train a commercial model.
Sources
Where is this documented?
- CC BY-NC 4.0 license text (opens in a new tab)Official Creative Commons NonCommercial license deed creativecommons.org
- CC BY-NC-ND 4.0 license text (opens in a new tab)Official Creative Commons NonCommercial-NoDerivatives license deed creativecommons.org
Check it against the sample pack.
40 episodes across 4 environments, delivered in the exact schema these guides describe.