Licence and trademarks¶

Code: Apache License 2.0¶

In short, you may use, modify and redistribute the code, including commercially, provided that:

you keep the NOTICE file, which names this project as the origin, in redistributions and derivative works, forks included (Section 4(d));
you mark the files you changed (Section 4(b));
you include the licence.

Contributions are accepted under the same licence (Section 5). This summary is not legal advice. The licence text is what applies.

Name and logo: not licensed¶

The Apache licence does not grant rights to the name OpenTernary, the OPEN-TERNARY wordmark or the three-tomoe symbol (Section 6). The details are in TRADEMARKS.md:

Allowed without asking	Not allowed without written permission
redistributing unmodified releases, which contain the marks	naming a fork, product or service "OpenTernary" or something confusingly similar
referring to the project: "based on OpenTernary", "a fork of OpenTernary", "works with OpenTernary"	using the logo or wordmark as the identity of a fork, product or service
showing the unmodified logo next to such a reference, without suggesting endorsement	suggesting a modified version is official or endorsed

A public fork should use its own name and logo. Change src/openternary/ui/brand.py and the assets, and keep the NOTICE file.

Models and data you use¶

OpenTernary does not include any model weights or datasets. Models you quantize and datasets you evaluate on keep their own licences, and a quantized model is a derivative of the original. In particular:

Gemma models are subject to Google's Gemma terms, which you accept on Hugging Face before downloading.
The research quality dataset is built from WikiText (CC BY-SA 3.0 / GFDL), SQuAD (CC BY-SA 4.0) and JGLUE/JSQuAD (CC BY 4.0). Those terms apply to the built file.
wiki-tiny calibration downloads WikiText from Hugging Face.

Check the licence of every model and dataset before you publish anything derived from it.