A Domain-Adapted Multilingual Large Language Model for Mining Safety in African Languages

The African mining industry continues to face a serious communication gap between technical safety documentation and the language used by the workforce. Safety manuals, operating procedures, and regulatory guidance are often produced in English or French, while many workers communicate in isiZulu, chiShona, and other local languages. This mismatch is not just inconvenient; it can be dangerous in safety-critical settings where misunderstandings may contribute to preventable incidents. This paper reports on Talk MINING, a domain-adapted multilingual large language model built on InkubaLM and trained on mining-specific text, question–answer pairs, and multimodal hazard data. The model was developed using Hugging Face tools and evaluated with both quantitative and qualitative methods. Results show a marked improvement over the baseline model, with overall accuracy rising from below 50% to 82%, and strong gains across isiZulu, chiShona, and English. The project also includes a web-based prototype that allows users to take photos of mining conditions and receive guidance in a local language. This paper presents the full pipeline for Talk MINING, including document preparation, translation and localization, model training, evaluation, deployment, monitoring, and validation. Beyond benchmark performance, the system is designed for practical deployment in environments where connectivity is limited and literacy levels vary—supporting both text queries and image-based hazard input through a mobile-ready prototype. The work contributes a reproducible pipeline, a curated multilingual mining corpus, and an open evaluation framework that can serve as a foundation for extending domain-adapted AI to other African languages and safety-critical industries. Ultimately, Talk MINING demonstrates that closing the language gap in mining safety is not only technically feasible but achievable with relatively modest model size when domain adaptation is done rigorously.