Detecting hate and offensive spans in Vietnamese content presents unique challenges due to the complex linguistic structure of the language. This study introduces VHateX, an ensemble-based framework developed to detect hate and offensive spans in Vietnamese text using the PhoBERT model. The primary contribution of this work is the integration of PhoBERT with ensemble learning techniques to significantly improve the model’s performance in the problem of Vietnamese hate and offensive spans detection within the ViHOS dataset. To enhance model robustness and diversity, the training set of the ViHOS dataset is divided into three bootstrap subsets, each of which is randomly drawn from 80% of the original training data. This study trained three independent classification models (PhoBERT) on three different bootstrap subsets of the ViHOS dataset and combined their predictions using soft voting. The results indicate that VHateX outperforms existing methods, achieving higher Precision (8.1% increase), Recall (8.1% and 10.2% increase), and F1 Score (10.9% and 12.2% increase) on both syllable-based and word-based models. These results demonstrate the effectiveness of VHateX in real-world applications for Vietnamese hate and offensive span detection.