Pilot information
As a part of the pilot, Makerere AI Lab and Google Analysis collected 8,091 annotated adversarial queries in English and 6 African languages (e.g., Pidgin English, Luganda, Swahili, Chichewa). The queries are adversarial in nature and have a excessive probability of manufacturing unsafe responses from an LLM as a method of testing and mitigating for potential hurt. This dataset in flip can be utilized to guage fashions for his or her security and cultural relevance inside the context of those languages. The dataset is open-source and accessible for exploration.
Consultants from seven delicate domains (e.g., tradition and faith, employment) annotated these queries with ten subjects inside their area of experience (i.e., “corruption and transparency” for politics and authorities area), 5 generative AI themes (e.g., public curiosity, misinformation) and 13 delicate traits (e.g., age, tribe) which are related to the African context.
Essentially the most distinguished domains have been well being (2,076) and training (1,469), with the highest subjects being power illness (373) and training evaluation and measurement (245), respectively. Nearly 80 p.c of the queries contained contextual details about misinformation or disinformation, stereotypes, and content material related to public welfare resembling well being or legislation. Nearly all of the queries have been about social teams belonging to gender (e.g., “Chibok women”), age (e.g., “newborns”), faith or perception (e.g., “Conventional African” religions), and training stage (e.g., “uneducated”).