Glaucoma, a leading cause of irreversible blindness worldwide, demands urgent and accurate di-agnosis. Existing AI models for glaucoma detection face challenges due to the limited and non-diverse nature of training datasets and insufficient validation in real- world scenarios. Bridging this gap, we introduce NETRA, an extensive and diverse dataset of 519 retinal images from Nepalese glaucoma patients, covering all stages of the disease. This contribution enriches the available resources for model training and enables a more comprehensive evaluation of AI performance in glaucoma detection. We trained eight state-of-the-art deep learning models using the open-source RIGA dataset and fur-ther validated their performance on NETRA. Notably, the EfficientNetbO-Unet++ model demonstrated superior performance, achieving accuracy rates of 95.70% and 81.60 % in identifying optic discs and cups, respectively. Furthermore, the ResNet34-Unet++ model attained the highest AUC of 0.92, signaling the most accurate cup- to-disc ratio calculation, a critical metric in glaucoma screening. Our study underscores the value of diverse datasets, robust deep learning methodologies, and rigor-ous real-world validation in achieving precise and reliable AI-based glaucoma diagnoses. These findings offer a significant advancement towards the practical clinical application of AI in glaucoma diagnosis.