scikit-learn
France's Inria took on a library that gave classical machine learning one interface in Python: fit, predict, score.
Why it matters
The shared interface mattered more than any individual algorithm: methods became interchangeable in a single line of code.
Before scikit-learn every implementation had its own conventions, and comparing two methods meant rewriting code. One interface made comparison trivial and processing pipelines composable. The library remains the most used tool for problems where deep learning is excessive, which is most applied problems. Development has been largely European.