Worked on PaddleNLP and Paddle repositories to enhance distributed deep learning workflows, focusing on reinforcement learning and model training stability. Addressed training issues in RLHF reward modeling by refactoring the flashmask reward training setup and improving data processing, which increased experiment reliability. Updated documentation and training configurations to clarify data formats and streamline onboarding. In distributed training scenarios, unified FuseLoss handling across Qwen2 and Qwen3 models, ensuring correct hidden state gathering and reshaping for efficiency and correctness. Improved pipeline parallelism robustness in Paddle by adding checks for None tensors. Utilized Python, deep learning frameworks, and parallel computing techniques throughout.