No exhibits filed yet
Paste a batch of URLs and press Check. Each one opens its own exhibit card here.
Paste dataset or repo URLs. Get a plain-English flag before you train on them.
View on GitHubNo exhibits filed yet
Paste a batch of URLs and press Check. Each one opens its own exhibit card here.
Whether you can train an AI or machine learning model on a dataset comes down to its license, and the license is not always the SPDX tag at the top of the repo. Provenance reads the actual LICENSE and README text behind each GitHub or Hugging Face URL you paste and returns a plain answer: clear, caution, or restricted, with the exact clause quoted so you can check it yourself.
Standard open-source licenses like MIT and Apache 2.0 mostly say nothing about machine learning training one way or the other. The risk sits in the exceptions. A growing number of datasets and model repos add non-standard terms straight to a README or a custom LICENSE file: "not for use in training machine learning models", OpenRAIL-style behavioral restrictions, or a NonCommercial license applied to data in a way that is ambiguous for commercial training. Generic license scanners built for software compliance skip these, because "restricts AI training" is not part of any SPDX taxonomy. Provenance scans for that specific phrasing with a maintained pattern library and flags every hit.
Each URL resolves to one of three stamps. CLEAR means no training restriction was found. CAUTION means the terms are ambiguous for machine learning, most often a NonCommercial license or a research-only grant. RESTRICTED means an explicit no-training clause was matched, and the card shows the sentence and the file it came from. No verdict is a black box: if Provenance flags a source, it points at the words that caused it.
It depends on the dataset's license and any extra terms in its README. If the license is permissive (MIT, Apache 2.0, BSD) and carries no added restriction, training is generally allowed. A NonCommercial or research-only license, or an explicit no-AI-training clause, changes that answer. Paste the URL above to see which case applies.
Yes. The MIT license permits use, modification, and distribution with no field-of-use limits, so training a model on MIT-licensed data is allowed as long as you keep the copyright notice. Provenance still scans the README, because some MIT repos bolt on a separate training restriction the license text alone would not reveal.
Creative Commons NonCommercial licenses allow non-commercial and research training but are ambiguous for commercial machine learning use, since "NonCommercial" is not defined for model training. Provenance flags the CC-BY-NC family as CAUTION rather than guessing.
Not always. An OSI-approved license covers copying and distributing the work, but a dataset published under it can still carry a separate usage note that restricts training. Always read the README as well as the LICENSE, which is exactly the pair Provenance checks for you.