Worked on the facebookresearch/fairseq2 repository to address a critical issue in the WerMetric class, focusing on accurate word error rate (WER) metric computation. Corrected the construction of hyp_seqs_list by ensuring hyp_seqs, rather than ref_seqs, was used, thereby eliminating misleading metric signals and improving the reliability of model evaluation. Collaborated closely with another contributor to deliver a lightweight, well-documented fix that aligns with the project’s QA and release processes. Utilized Python for data processing and metrics implementation, applying machine learning knowledge to enhance the trustworthiness of model comparisons and support research and product decision-making within the codebase.