SpamBayes 1.1a6 is now available! (This includes both the source archives and a Windows binary installer).
See the download page for more info or select an appropriate version from the table below.
You may also like to see what other people have been saying about us in the press and elsewhere.
Locate the row which contains your operating system and mail program to see which version of SpamBayes is right for you. If you can test any of the configurations, please let us know. Note that installing a source release is more involved than the binary releases.
Please try the test releases if at all possible. While they are still labelled as "alpha", they are really quite stable. We're just extremely conversative/lazy about doing beta/final releases. ;-)
What is SpamBayes?
The SpamBayes project is working on developing a statistical (commonly, although a little inaccurately, referred to as Bayesian) anti-spam filter, initially based on the work of Paul Graham. The major difference between this and other, similar projects is the emphasis on testing newer approaches to scoring messages. While most anti-spam projects are still working with the original graham algorithm, we found that a number of alternate methods yielded a more useful response. This is documented on the background page.
SpamBayes is not a single application. The core code is a message classifier, however there are several applications available as part of the SpamBayes project which use the classifier in specific contexts. For the most part, the current crop of applications all operate on the client side of things, however, a number of people have experimented with using SpamBayes on mail servers to classify incoming mail for multiple users. The table below outlines the main applications which are part of the SpamBayes distribution.
That's great, but what's SpamBayes?
(the non-technical hand-waving answer)
SpamBayes will attempt to classify incoming email messages as 'spam', 'ham' (good, non-spam email) or 'unsure'. This means you can have spam or unsure messages automatically filed away in a different mail folder, where it won't interrupt your email reading. First SpamBayes must be trained by each user to identify spam and ham. Essentially, you show SpamBayes a pile of email that you like (ham) and a pile you don't like (spam). SpamBayes will then analyze the piles for clues as to what makes the spam and ham different. For example; different words, differences in the mailer headers and content style. The system then uses these clues to examine new messages.
For instance, the word "Nigeria" appears often in spam, so you could use a spam filter which identifies anything with that word in it as spam. But what if your business involves writing a guidebook on Nigerian Wildlife Conservation? Clearly a more flexible approach is necessary. Additionally spammers will adapt their content over time and will no longer use the word "Nigeria" (or the words "Lose Weight Fast", or any number of other common lines). Ideally the software will be able to adapt as the spam changes.
So, that is what SpamBayes does. It compares the spam and the ham and calculates probabilities. For instance, for me, the word "weight" almost never occurs in legitimate email, but it occurs all the time in 'lose weight fast' spam. SpamBayes can then look at incoming email, extract the most significant clues and combine the probabilities to produce an overall rating of "spamminess". It flags the messages so that your mailer can handle the different message types. You might set it up so that ham goes straight through untouched, spam goes to a folder that you ignore (or delete without checking) and the unsure messages go to another folder which you can review for errors.
How is SpamBayes different?
There are a number of similar projects to SpamBayes - most are just using the original Paul Graham algorithm. Examining the Graham technique with careful testing showed that it did a remarkably good job, but there was considerable room for improvement. (See the background page for more.)
The SpamBayes team tinkered with new algorithms, tweaking existing algorithms, and, most importantly, did enormous test runs, slamming tens of thousands of messages against each other, in an attempt to quantify whether or not a change to the system was beneficial.
The new algorithm is a combination of work from Gary Robinson and Tim Peters, and provides not just a 'spam' and 'ham' rating, but also an 'unsure' rating, for those messages where it can't work out how to rate the message.
See the background page for more, well, background.
The code (implemented in Python) is currently available from a variety of methods from the downloads page.
There are now a couple of end-user applications available for those excited by the bleeding edge - these are detailed on the Applications page, and available as part of the source download.
Most of the heavy lifting on this project was done by Tim Peters, with the cast of spambayes obsessive-compulsives providing ideas, heckling, and testing. Gary Robinson provided a lot of the serious maths and theory, as well as his essay on "how to do it better" (see the background page for a link). Rob Hooft also contributed maths/stats clues. Mark Hammond amazed the world with the Outlook2000 plug-in (with Tony Meyer, Sean True, and Adam Walker making significant contributions), and Richie Hindle, Neale Pickett, Tim Stone worked on the end-user applications.
(Thanks also to Rachel Holkner for turning Anthony's gibberish into something closer to actual English, although all mistakes are Anthony's.)