Glossary
What Is a Training Data License?
Last reviewed: October 2026
Short answer
A training data license is the contract that sets what a buyer may do with a dataset when training and commercializing AI models. It covers scope of use, term, exclusivity, ownership of derived models, redistribution, and the consent and provenance behind the data. This page is general information, not legal advice.
01
Why it matters
The license decides whether a dataset can legally train a model you ship. A dataset that fits your task but carries a research-only or non-commercial grant can leave a product without usable training data, and a vague grant can leave derived-model ownership unclear.
02
How Firsthand handles it
Firsthand delivers custom egocentric data under a perpetual, irrevocable, worldwide, buyer-owned license that is non-exclusive by default, with a warranted consent chain and a data card listing consent artefact IDs per episode. Exclusivity is an added term, and the signed order form controls.
03
What is a training data license?
A training data license is the legal permission attached to a dataset. It states what the buyer may do with the data when building and selling AI models, and what the buyer may not do. No single standard form exists: the terms come from a public license text, a dataset agreement or a negotiated order form.
This page is general information, not legal advice. Have your own counsel review the actual document before you sign or train.
04
What terms does a training data license usually cover?
Most licenses answer the same handful of questions: scope of use, term, revocability, exclusivity, ownership of derived models, redistribution, and the consent and provenance behind the data. The first table below lists each term, what it controls and the question a buyer should ask.
Scope of use usually decides whether a dataset can go into a shipped product at all, so read it first.
05
How do public dataset licenses differ from custom-collection licenses?
A public dataset has one set of terms for everyone, written by its publisher. Some are permissive and some restrict use. The Creative Commons Attribution-NonCommercial 4.0 deed, for example, says: “You may not use the material for commercial purposes.” A research-only grant may therefore not cover a commercial model.
A custom-collection license is written for one buyer around data collected to a skill spec, so scope, exclusivity and consent records can be agreed before capture. The second table compares the two in general terms and names no particular dataset.
06
What does perpetual, irrevocable and non-exclusive mean for a buyer?
Perpetual means the grant has no end date. Irrevocable means the grantor cannot withdraw it. Non-exclusive means the grantor may license the same or similar data to others. The three are independent: a buyer can hold a perpetual, irrevocable grant that is still non-exclusive.
On this site, buyer-owned describes a license grant, not a copyright assignment. Exclusivity, if you want it, is a separate term written into the order before capture.
07
How do consent and provenance relate to the license?
A license can only grant what the data owner had permission to grant, so the consent chain is the ceiling on any license. Provenance is the record that shows it: who recorded, what they agreed to and where capture happened.
Firsthand states that every contributor signs a written commercial release before capture, that faces and license plates are blurred and audio is scrubbed, and that the batch data card lists consent artefact IDs per episode and jurisdictions of capture.
08
What license terms does Firsthand state?
Firsthand delivers custom egocentric data under a perpetual, irrevocable, worldwide, buyer-owned license that is non-exclusive by default. The third table lists each stated term. Anything it does not list is not stated, and the signed order form controls over the website.
09
How do you get one with fast, on-demand custom collection?
Put license scope, exclusivity and consent requirements in the skill spec before capture begins. Where coverage is deep, a first validated batch lands in a median of 48 hours from signed spec; thinner environments typically take three to four weeks.
10
What common license terms control (general, not legal advice)
| Term | What it controls | Question to ask |
|---|---|---|
| Scope of use | What you may do: training, evaluation, fine-tuning, deployment; commercial or research only | Does it cover a shipped commercial model? |
| Term | How long the grant lasts | Is it perpetual or does it expire? |
| Revocability | Whether the grantor can withdraw the grant | Is it irrevocable? |
| Exclusivity | Whether others may receive the same data | Who else can get it? |
| Derived models | Who holds models trained on the data | Is there any royalty or reporting duty? |
| Redistribution | Whether the data itself can be passed on | Can you share the raw data? |
| Consent and provenance | The permissions behind the data | Is there a consent record per item? |
12
Public-dataset licenses vs custom-collection licenses (general comparison)
| Aspect | Public dataset license | Custom-collection license |
|---|---|---|
| Who writes the terms | The dataset publisher, one set for everyone | Agreed with one buyer in an order form |
| Commercial use | Varies; some terms, such as Creative Commons NonCommercial, bar it | Set by the contract |
| Exclusivity | None; open to anyone who accepts the terms | Optional added term |
| Consent record | Whatever the publisher documented | Can be required in the contract |
13
What this site states about Firsthand license terms
| Term | What the site states |
|---|---|
| Grant | Perpetual, irrevocable, worldwide, non-exclusive by default |
| Permitted use | Training, evaluating, fine-tuning and deploying models, including commercial models |
| Derived weights | Yours; no royalty, revenue share or reporting obligation |
| Redistribution | Not permitted for the footage itself; unlimited for models derived from it |
| Term | No expiry; a lapsed program does not revoke data already delivered |
| Ownership | Buyer-owned license; not a copyright assignment |
| Exclusivity | Available per skill line or per environment as an added term |
| Consent and warranty | Consent chain warranted for every delivered episode, with indemnity against claims arising from it |
| Governing document | The signed order form controls over the website |
| Exclusivity duration or price | Not stated |
| Sublicensing, liability caps, certifications | Not stated |
14
Frequently asked questions
- What is a training data license?
- The contract or license text that sets what a buyer may do with a dataset when training and commercializing AI models. It covers scope, term, exclusivity, derived-model ownership, redistribution and consent.
- Is a license the same as owning the data?
- No. A license is a permission to use. On this site, buyer-owned describes a perpetual, irrevocable license grant, not a copyright assignment.
- Can I train a commercial model on a public dataset?
- It depends on that dataset license. The Creative Commons NonCommercial deed says you may not use the material for commercial purposes. Read the actual terms and ask your counsel.
- Who owns models trained on Firsthand data?
- The site states derived weights are yours, with no royalty, revenue share or reporting obligation. The footage itself cannot be redistributed; models derived from it can, without limit.
- Does a license cover consent?
- A license cannot exceed the consent behind the data. Firsthand states written releases before capture, a warranted consent chain with indemnity for claims arising from it, and consent artefact IDs per episode in the data card.
- What does Firsthand not state about its license?
- Exclusivity duration or price, sublicensing, liability caps and certifications are not stated. They belong in the signed order form, which controls. This is not legal advice.
Related terms