Most NLP tasks depend on the “single ground truth” assumption, so NLP datasets are annotated for a single ground truth. However, this assumption does not hold for subjective NLP tasks such as hate speech classification, where hate speech text is subjective to different annotators. For a fair evaluation, it is essential to consider the opinions of different annotators and how they diverge on specific topics. This work aims to study conflicting perspectives from annotators with various demographic backgrounds and how state-of-the-art models can be improved for multi-perspective hate speech datasets for low-resource language.