Nice article and an interesting approach. thanks for writing it up.

  The model can only get better with more training data,
  which requires me to judge whether I like stories or not.
  I do this occasionally [using] the above command-line tool,
  but maybe I’ll come up with something better in the future.

Well, you could analyse your server logs to see which stories you really did click on and which you skipped over (You'll probably want to only consider pages that have at least one click for the edge case of "I didn't even look at that page". Need a cookie or login to make sure it only counts your clicks)

Also consider scraping your HN "saved stories" list as a positive source.

Don't recall if you mentioned it in your article, but you'll probably want to randomly insert the occasional low-scoring article as a check for under-weighting.

I really wish our profile pages supplied a (private) log of up/down votes for comments and flags for stories in addition to the the up votes for stories. It would make for some interesting datamining