Worked on the nod-ai/SHARK-Platform repository to enhance perplexity evaluation for language models, focusing on both feature development and stability improvements. Addressed issues in the Perplexity class by refining sequence length handling and cache state access, and updated the initialization logic for cache paging in TorchGenerator. Improved the continuous integration workflow by adjusting test parameters, enhancing logging, and introducing dedicated test prompts, which streamlined pull request validation and increased test reliability. Utilized Python, PyTorch, and YAML to implement these changes, resulting in more robust model evaluation and faster feedback cycles, while maintaining a strong emphasis on testing and CI/CD best practices.