A Marigold registry for the Grasmere models: Gemma 4 E4B fine-tuned for Welsh, Irish, Scottish Gaelic and Scots.
| Application | Directory | What it does |
|---|---|---|
chat |
chat | chat with the Grasmere models through Open WebUI |
evals |
evals | the same Welsh prompts to each Grasmere model and stock Gemma 4, answers side by side |
The names are deliberately those of applications in other registries,
such as chat in marigold-examples:
a host with both registries resolves chat to whichever it lists
first.
The registry is published at https://grasmere-chat.marigold.run.
Requires Marigold, Docker and, for these models, a GPU with about 24 GB of memory.
git clone https://github.com/bayinfosys/grasmere-chat
cd grasmere-chat
marigold application create chat -o dist/
marigold application install dist/chat-local.tar.gz
marigold cache populate chat
marigold application start chat # Open WebUI at http://localhost:3000
The evaluation, the same way:
marigold application create evals -o dist/
marigold application install dist/evals-local.tar.gz
marigold cache populate evals # needs HF_TOKEN for google/gemma-4-e4b-it
marigold application start evals
marigold application logs evals
Its results are written to the application’s outputs directory under the Marigold cache.
Each application is one directory below the root, holding its
marigold.toml, and the directory’s name is the application’s name. marigold-registry.toml at the root describes the
registry. Pushing a v* tag publishes every application under that
tag.
The Grasmere models are by jonpojonpo. They are Gemma 4 fine-tunes, used under the Gemma terms. This repository holds only Marigold applications, under the MIT licence.