Worked on stabilizing the multi-turn training data flow in the modelscope/ms-swift repository by addressing a bug in the inputs_to_rolloutrequest function. The solution involved refactoring the data_dict merging logic to prevent nested structures and ensure that new and existing data were combined correctly across repeated calls. This backend development effort, implemented in Python, enhanced the robustness of the rollout data pipeline and reduced the risk of data corruption during training. The changes maintained compatibility with existing interfaces and improved traceability through clear commit documentation, demonstrating a focus on bug fixing, data handling, and careful integration within established workflows.