Language & Safety Flagship NLP Benchmark Active Curation

Zanzibar Swahili Corpus, Spoken QA & Dira Provenance Benchmark

Building authentic coastal Kiunguja Swahili text and spoken domain QA paired with the Dira benchmark to detect machine-translated translationese from genuinely authored Swahili prose.

CARA-FM culturally attuned resource-aware East African foundation models poster

Scientific Context & Approach

Most public African language benchmarks consist of English datasets machine-translated into standardized textbook Swahili, erasing syntactic rhythm, coastal vocabulary, and agrarian domain context.

The SAAIL Lab BS research team curates verified native Swahili prose alongside spoken Zanzibar agricultural QA, training and benchmarking the Dira detector to measure provenance quality across modern language models.

Core Technical Objectives

  • Curate verified native Kiunguja and East African editorial Swahili corpora alongside aligned synthetic translations.
  • Develop lexical, discourse, and neural stylometric detectors separating translationese from human Swahili authoring.
  • Construct audio-text smallholder agricultural and coastal livelihood QA pairs grounded in Zanzibar dialect usage.

Verified Deliverables & Outputs

  • Zanzibar Swahili Corpus and spoken agricultural QA annotations
  • Dira Native vs. Translated Swahili benchmark and baseline models
  • Dialectal evaluation protocols for low-resource speech and text pipelines