Worked on the apple/axlearn repository to enhance loss-function configurability and consistency for deep learning workflows. Refactored core loss functions in Python and JAX to support both float and Tensor temperature parameters, enabling flexible experimentation and compatibility with mixed-precision training. Established a standardized protocol for implementing contrastive losses, reducing code duplication and improving maintainability across the codebase. Improved the instantiation of loss functions within the CLIPFusionNetwork, making it easier to configure and extend for various machine learning experiments. These changes streamlined experimentation, reduced maintenance overhead, and provided a scalable foundation for future contrastive learning development within the project.