Pablo implemented an optional data_dir argument for the Huggingface DataLoader in the tracel-ai/burn repository, enabling users to specify relative paths for subset datasets and supporting manual download workflows. This feature required coordinated updates to both the Rust-based DataLoader and the Python importer, ensuring seamless integration across the data provisioning pipeline. By focusing on data loading and dataset management, Pablo addressed the need for more flexible dataset handling, particularly for workflows involving partial or manually curated datasets. Although no major bugs were fixed during this period, the work laid a foundation for broader adoption of improved dataset management practices within the project.