Worked on the PaddlePaddle/PaddleFormers repository to calibrate and update baseline values for ERNIE-21B SFT loss metrics, addressing baseline drift in model training workflows. Focused on improving the accuracy and reliability of training metric monitoring by applying targeted patches and aligning the metrics pipeline with updated baselines. Utilized data analysis and model training skills, working primarily with JSON to ensure consistent metric reporting across training runs. This effort enhanced the fidelity of model evaluation and supported faster iteration cycles by providing more accurate feedback on training performance, ultimately contributing to more reliable decision-making in machine learning development processes.