Event language
UI language
Adapting Large Language Models to low-resource languages presents distinct technical challenges. Drawing from the SoroLLM project (a UNICEF collaboration for Tajik language), this talk examines four key obstacles: insufficient training data, limited computational resources, catastrophic forgetting, and suboptimal tokenization. We will discuss practical mitigation strategies and the inherent trade-offs in each approach, providing actionable insights for researchers and engineers working on domain adaptation for underrepresented languages.
