Worked on turbo-llm/turbo-alignment to optimize the Reward Model training pipeline, focusing on data processing and model training using Python. Refactored the training workflow to implement sequential packing, which improved efficiency and throughput. Enhanced data handling by moving batching logic from the dataset to the data collator, enabling more precise batch construction and padding. Addressed tensor size mismatches by standardizing sequence lengths and improved boundary logic for maintainability. Expanded documentation and logging, and set up the groundwork for sequence-parallel execution and DPO readiness. Collaborated closely with peers to ensure robust implementation and code quality throughout the development process.