Understanding civic discourse
Open research advancing language resources, models, and evaluations for civic AI across underrepresented languages.
- CIVIC AI
- LANGUAGE
- OPEN RESEARCH
The Civic Corpus
ResearchAn open foundation for multilingual civic understanding, connecting local expressions, civic terminology, and structured language data to shared concepts across communities.
- Luganda
- Terminology
Purpose-built vs. general-purpose models for Swahili translation
Technical ReportA comparative evaluation of translation-specific architectures and instruction-tuned language models, demonstrating the performance advantages of targeted adaptation on shared benchmarks.
- Swahili
- Model Evaluation
Swahili-nllb-1.3B and Swahili-nllb-600M
Model ReleaseTwo LoRA-adapted NLLB-200 models advancing English-to-Swahili translation through efficient, task-specific fine-tuning.
- Swahili
- Translation
Continued pretraining for Swahili language adaptation
Technical ReportAn investigation into how continued pretraining followed by supervised fine-tuning improves translation performance compared with direct instruction tuning.
- Swahili
- Model Adaptation
Swahili-gemma-1B-it
Model ReleaseA Gemma 3 1B model adapted through continued pretraining and instruction tuning for English-to-Swahili translation and Swahili instruction-following.
- Swahili
- Language Modelling
Adapting Gemma 3 1B for Luganda translation
Technical ReportA study of parameter-efficient language adaptation, improving English–Luganda translation from 0.06 to 13.85 BLEU using QLoRA and limited computational resources.
- Luganda
- QLoRA
Luganda-gemma-1B-it
Model ReleaseA QLoRA-adapted Gemma 3 1B model supporting bidirectional English–Luganda translation and Luganda-language interaction.
- Luganda
- Translation
Open Language Datasets
DatasetsExplore 11 open datasets supporting multilingual model development, training, and evaluation across Luganda and Swahili.
- Multilingual
- Open Data
