I had a short look at scikit-learn for a hobby classifier I'm writing (it aims to email me interesting HN articles so I don't waste much time reading things I don't like or refreshing HN all the time), and it looks fantastic. I haven't looked at the code or used it very much, but the documentation is very thorough and informative, and the library itself is very extensive.

There are some things missing, (e.g. The naive Bayes implementation is offline, which is not very useful to me), but the library is fantastic for prototyping, at least. Congratulations to the people working on it!