Indonesian mangosteen exports have increased markedly, highlighting the necessity for precise ripeness determination in harvesting, sorting, and quality evaluation. The lack of real-world datasets constrains computer vision-based classification. This research presents a new dataset for categorizing export-grade mangosteen ripeness into four categories. Statistical methods, such as Chi-Square, ANOVA, Tukey HSD, and Levene's test, assess dataset characteristics and class distribution, whilst EDA and HSV colour analysis further investigate its structure. Deep learning models (DenseNet121, InceptionV3, ResNet50, EfficientNetB3) are evaluated based on accuracy, precision, recall, F1-score, and loss. The highest accuracy of 92.15 % is attained without data augmentation or transfer learning, illustrating the dataset's efficacy. This dataset establishes a benchmark for classifying mangosteen ripeness, addressing data scarcity, and supporting the evaluation of deep learning models in export quality control. © 2025 IEEE.