mlr3 hugging face

This project proposes mlr3hf, an R package for integrating Hugging Face datasets into the mlr3 ecosystem. The package will allow users to download datasets from Hugging Face, cache them locally, and convert them into mlr3 tasks efficiently. To support large datasets, the package will use Parquet + DuckDB for lazy, disk-based querying, minimizing RAM usage while preserving scalability. It will also support fallback handling for canonical formats such as CSV, JSON, and TSV, along with split-aware task creation for train/test/validation datasets. The final outcome will be a practical, memory-efficient, and CRAN-ready workflow for using Hugging Face datasets in mlr3.