Worked on the nvidia-cosmos/cosmos-rl repository to refactor image input handling for Qwen2.5-VL, focusing on aligning data schemas with Hugging Face transformers. The main contribution involved renaming the image input key from pixel_values_images to pixel_values throughout the data packing and model forward processes, which resolved a persistent mismatch and clarified image data flow. This update improved the robustness and maintainability of the data pipeline, reducing potential runtime errors and easing future integrations. The work demonstrated practical application of Python, PyTorch, and Hugging Face Transformers, with careful attention to minimizing disruption in existing machine learning training workflows.