Worked on backend reliability and configuration flexibility for deep learning systems, focusing on Python-based projects. In the volcengine/verl repository, addressed transformer configuration stability by fixing a TypeError that occurred when multiple keyword arguments were overridden in Megatron integration, ensuring robust application of configuration overrides for model training and inference. In the modelscope/ms-swift repository, improved dataset provisioning workflows by making the vp_stage argument optional in the swift_datasets_provider, resolving an AssertionError and enhancing compatibility with vpp and mcore 0.15. Demonstrated expertise in Python, deep learning, and model configuration, with a focus on targeted bug fixes that improved system robustness.