Worked on the EvolvingLMMs-Lab/lmms-eval repository to enhance the reliability of ScienceQA evaluation by addressing a bug in the post-processing logic. Focused on improving the accuracy of predicted versus target answer comparisons, the solution introduced case-insensitive exact-match evaluation and added support for predictions beginning with a letter followed by a period. This adjustment reduced false mismatches and improved the trustworthiness of model benchmarking metrics. Leveraged Python for bug fixing and data processing, applying natural language processing techniques to refine evaluation logic. The work resulted in a more robust evaluation pipeline, supporting faster and more data-driven model refinement decisions.