hngondoki/ge_sheng_guardrails_project

The goal of this project is to explore model options that can classify text and ensure that generative AI systems respond only with vetted, contextually relevant content, thereby enhancing both the accuracy and safety of conversational AI in with low resource languages. # Sheng Guardrails Project ## Introduction In Kenya, Sheng, an evolving informal language that fuses Swahili and English, is widely used, especially among the youth. While this linguistic hybrid fosters cultural expression and social cohesion, it poses significant challenges for Natural Language Processing (NLP) systems trained predominantly on standard English. Current chatbot models, and other service providers provide In-Topic (IT) guardrails to determine whether user queries fall within predefined subject boundaries. However, these models often have limited linguistic comprehension for low resource languages such as Sheng due to limited training data and insufficient linguistic representation. This leads to poor topic classification, decreased user trust, and reduced overall utility of AI-driven services in underrepresented communities. ## Objective The goal of this project is to explore model options that can classify text and ensure that generative AI systems respond only with vetted, contextually relevant content, thereby enhancing both the accuracy and safety of conversational AI in with low resource languages. ## Data WARNING: The information contained in the datasets is of a sensitive nature. ⚠️ Content Advisory: Please be advised that the datasets referenced in this presentation contain sensitive and potentially explicit material, particularly related to mental, sexual, and reproductive health. This content is included strictly for academic and research purposes, and is presented with the intent to foster informed, respectful, and evidence-based discussion. Viewer discretion is advised. ### Data description - wz_qa.csv: This is a file with 308 sheng data (questions and answers) that has been vetted and labelled already. The whole data set is IT and the questions + Answers will be combined to provide the vocabulary within the sentence embeddings. It has 2 columns, question and answer - wz_eng_labeled.csv: This contains 117 rows of data in english w …