In short
- The Amada weights: CC BY-NC-ND 4.0. Free to use for non-commercial purposes, with credit.
- No commercial use, and no sharing of modified or fine-tuned versions.
- Amada is built on mmBERT-base (MIT license) by Johns Hopkins University (CLSP).
- The website itself (story, images, code): all rights reserved by LiamDB.
- Fonts: SIL Open Font License.
At a glance
| Part | License | You can | You can't |
|---|---|---|---|
| Amada model weights | CC BY-NC-ND 4.0 | use it, share unchanged copies, with credit, non-commercially | sell it, use it in a paid product, share modified or fine-tuned versions |
| mmBERT-base (base network) | MIT | almost anything, keeping its copyright notice | remove its notice |
| Website: story, texts, images, videos, code | all rights reserved | look, share links, screenshots with credit | reuse them elsewhere without permission |
| Fonts | SIL OFL 1.1 | see the license of each font | sell the fonts on their own |
The Amada model
The Amada model weights are released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International license (CC BY-NC-ND 4.0). In plain words:
- Attribution (BY): when you share or use Amada publicly, credit it as "Amada by LiamDB" and link to this license.
- NonCommercial (NC): you may not use it mainly for money: no paid product, no paid service, no selling it.
- NoDerivatives (ND): you may use it as it is, but you may not share a changed version (for example a fine-tuned or merged model).
Research, learning, personal projects and tests are welcome. For any other use (commercial, or sharing a modified version), write to the contact address: permission can be discussed.
What Amada is built on
Amada is a fine-tuned version of mmBERT-base, a multilingual encoder made by the Center for Language and Speech Processing of Johns Hopkins University (jhu-clsp/mmBERT-base). mmBERT-base is released under the MIT license, which allows reuse as long as its copyright and license notice are kept. Anyone who shares the Amada files must keep that notice with them.
Training data
Amada learned from a mix of public datasets and synthetic examples. Being transparent about it:
- Public texts: Wikipedia, FineWeb-2, Project Gutenberg books, government reports, arXiv and PubMed abstracts.
- Public labelled datasets: QuALITY, ContractNLI, Qasper (long documents), CLINC150 (user requests).
- Synthetic data: a large part of the training examples (questions, answers, label lists, trap texts) was written by large language models from LiamDB's instructions, then filtered.
Each dataset keeps its own license and terms, available on its original page. Amada does not contain these texts as such: it learned patterns from them.
Never trained on, used only to test: Banking77, AG News, dair-ai/emotion, tweet_topic, deepset/prompt-injections, SST-2 and When2Call. This is why the benchmark results on the Site are "zero-shot".
The website
The story, texts, design, images, videos and code of the Site are © 2026 LiamDB, all rights reserved. You may look at them, share links, and share screenshots with credit (see the Terms of use).
The images and videos were generated with AI on OpenArt and then edited. In some countries, images made mostly by AI may not be protected by copyright in the same way as human work. Even so, please respect the work behind them and ask before reusing them.
Fonts
- Leckerli One by Gesine Todt: SIL Open Font License 1.1.
- Nanum Pen Script by Sandoll Communication: SIL Open Font License 1.1.
- IBM Plex Mono by IBM: SIL Open Font License 1.1.
The fonts are served by Google Fonts. The license text is at openfontlicense.org.
Names and trademarks
Windows and Microsoft are trademarks of Microsoft Corporation. YouTube is a trademark of Google LLC. OpenArt, Vercel, Anthropic and Claude belong to their owners. These names are used only to describe things; their owners do not sponsor or approve this Site.