CultureSEEDS East Africa: Culturally Grounded Swahili Safety Evaluation
Constructing community-sourced cultural and safety evaluation suites across Swahili and nine regional East African languages to prevent false-positive refusals and culturally blind safety failures in frontier models.
Scientific Context & Approach
Safety filters trained predominantly on Western English norms routinely misclassify everyday coastal Swahili idioms, agricultural advice, traditional ecological stewardship, and East African civic discourse as adversarial or unsafe, whereas they overlook genuine regional harms.
As East Africa Lead on the Google DeepMind and Google.org CultureSEEDS consortium, SAAIL Lab leads participatory prompt curation across 10 African languages anchored in Swahili. In parallel, our IITMZ cultural preservation grant with Hope K. McCoy documents oral community narratives across Zanzibar and mainland Tanzania.
Core Technical Objectives
- Assemble validated cultural safety prompts written directly by native speakers across coastal Swahili and regional languages.
- Distinguish benign cultural, religious, and agricultural expressions from genuine contextual safety risks.
- Benchmark frontier open and closed LLMs on cross-lingual refusal calibration and East African cultural fidelity.
- Archive community oral storytelling corpora under participatory data stewardship protocols.
Verified Deliverables & Outputs
- 100+ audited East African CultureSEEDS evaluation prompts across 10 languages
- Zanzibar oral narrative and coastal Swahili domain test suites
- Cross-model false-refusal and cultural grounding audit methodology