README / README.md
Dev-Noble's picture
Luganda eval sample includes Kimera corpus
67b6e8a verified
|
Raw History Blame Contribute Delete
2.23 kB
metadata
title: README
emoji: 🌍
colorFrom: blue
colorTo: red
sdk: static
pinned: false

CommonMind

Civic discourse spans languages. Our infrastructure makes sense of it.

CommonMind builds open language models, datasets, and benchmarks for African languages. Our Data Lab is where volunteers create, annotate, and verify civic language datasets, starting with the Civic Corpus.

Models

Swahili scores are on FLORES-200 devtest, English β†’ Swahili. Each model card has full evaluation details and usage code.

Model Language Task BLEU chrF++ License
Swahili-nllb-1.3b Swahili Translation (LoRA on NLLB-200 1.3B) 38.74 63.37 CC BY-NC 4.0
Swahili-nllb-600m Swahili Translation (LoRA on NLLB-200 600M) 35.97 61.26 CC BY-NC 4.0
Swahili-gemma-1b-it Swahili Translation + instruction following (Gemma 3 1B) 28.86 54.77 Gemma + NC
Luganda-gemma-1b-it Luganda Translation + chat (LoRA on Gemma 3 1B) 13.85* 46.59* Gemma

* Luganda scores are on a 100-example held-out sample of its training data (SALT and the Kimera corpus) in both directions, not FLORES, so they are not directly comparable to the Swahili scores.

Try them without installing anything in the Playground. Swahili-gemma-1b-it also runs locally with Ollama:

ollama run amplifiedaccessorg/Swahili-gemma-1b-it

Get involved

  • Contribute data: join the Data Lab to help build civic language datasets.
  • Contribute code: the platform is open source at github.com/Amplified-Access/common-mind.
  • Report a problem with a model: open a discussion on that model's page.

🌐 thecommonmind.org