grasmere-chat

chat

Chat in Welsh, Irish, Scottish Gaelic and Scots through Open WebUI, with the Grasmere models. Open WebUI is added to the platform as a service and answers at http://localhost:3000.

Models

Model Use
jonpojonpo/grasmere-e4b-ftpo everyday chat; the FTPO fix reduces repetition loops in long answers
jonpojonpo/grasmere-e4b the highest benchmark scores; loops more often in long answers
sentence-transformers/all-minilm-l6-v2 Open WebUI’s document embeddings

Both Grasmere models are Gemma 4 E4B fine-tunes, about 16 GB each in bf16. A 24 GB GPU holds one with room for its context. The worker keeps one model loaded at a time, so switching between them in Open WebUI reloads the weights.

Settings in Open WebUI

The model card recommends these, set per model under Workspace, Models:

Setting Value
Temperature 0.3
Top P 0.9
Custom parameter chat_template_kwargs {"enable_thinking": false}

Title, tag and follow-up generation run on the same worker as chat. To keep the queue for chat, turn them off under Admin Panel, Settings, Interface.

Replies arrive whole when generation finishes; Marigold does not stream token by token.

Comparing the models

The evals application runs a fixed set of Welsh prompts against each Grasmere model and a stock Gemma 4 baseline, and writes the answers side by side.