Worked on stabilizing distributed training workflows in the metatensor/metatrain repository by addressing a critical DataLoader edge-case. Focused on resolving a training crash that occurred when the batch size exceeded the dataset size, the solution involved conditionally setting the DataLoader’s drop_last parameter based on the relationship between batch size and dataset size. This approach improved the reliability of machine learning experiments, particularly for configurations involving large batch sizes. The fix was implemented in Python and reinforced with automated testing to ensure robustness. The work emphasized careful handling of data loading and testing to prevent interruptions during model training in distributed environments.