Worked on stabilizing data pipeline sampling in the facebookresearch/fairseq2 repository, addressing a bug that affected sampling accuracy when allow_repeats was set to false. Introduced a binary search algorithm to ensure only active pipelines were considered during sampling, effectively preventing incorrect data selection and potential leakage across multiple pipelines. Developed comprehensive unit tests to validate the correctness of the new approach, reducing the risk of regression in complex data pipeline environments. Utilized C++ and Python to implement and test the solution, applying skills in algorithm optimization and data pipeline development to improve the reliability of training data sampling processes.