Worked on stabilizing the Gemma3VL training pipeline within the NVIDIA/NeMo repository, focusing on resolving a critical bug that caused interruptions during model training. Addressed backward compatibility issues in dataset batch loading, ensuring that legacy workflows continued to function seamlessly with updated code. Improved the model export process by correcting vocabulary size handling, which prevented mismatches and export-time errors. The work involved deep learning and machine learning techniques, leveraging Python for both model training and export tasks. These targeted improvements enhanced the reliability and maintainability of the training workflow, directly supporting ongoing development and deployment of deep learning models.