Worked on the sglang repository to enhance the reliability of benchmark data processing by addressing a subtle but impactful issue with argument naming conventions. Focused on benchmarking and code refactoring using Python, the work involved renaming generated shared prefix benchmark arguments from 'gen_' to 'gsp_' within the data processing script. This change ensured that cache keys were generated correctly and that dataset sampling behaved as intended, directly improving the stability of benchmark results. The contribution demonstrated careful attention to detail in maintaining data integrity and reproducibility, reflecting a methodical approach to resolving bugs in performance-critical benchmarking workflows.