skops

Why we included this project

Scikit-learn's default way of saving models, Python's pickle, gets brittle fast once you step outside a single notebook. A model saved in one environment often fails to load in another after a library version bump, and loading pickled data from someone you do not trust can execute arbitrary code. skops gives sklearn users a safer persistence format that checks objects before loading, so a trained estimator can move to a teammate or a CI pipeline without those risks. It also generates model cards whose metadata Hugging Face Hub recognizes, which makes sharing a fitted model feel like pushing a repo rather than packaging a fragile artifact. Anyone who has hit a version-mismatch crash on a reloaded joblib file will see this as a practical way out.