Worked on stabilizing the dataset preparation workflow for the unslothai/unsloth repository, focusing on improving the reliability of the data ingestion pipeline. Addressed a runtime NameError in the _prepare_dataset function by importing the psutil library, which enabled proper access to system and process utilities during dataset preparation. This targeted bug fix reduced training downtime and prevented failures that previously disrupted production workflows. Related updates were made in rl.py to maintain compatibility and minimize edge-case regressions. The work was delivered as a concise, well-documented patch using Python, drawing on data science and machine learning expertise to enhance developer experience.