Rescore Mass-Check

This is the procedure we use to generate new scores. It takes quite a while and is labour-intensive, so we do it infrequently.

We generate new scores by analyzing a massive collection of mail (a "corpus"), and running software to create a score-set that gets the best possible set of scores, so that the maximum possible number of mails in that corpus are correctly classified (ie. so that SA thinks the ham messages are nearly all ham, and the spam messages are nearly all spam).

Summary

The corpus consists of many (approximately 1 million pieces) of real-world, hand sorted mail.

Basically a smallish number of people (about 15), including the developers themselves, work as volunteer "corpus submitters". They hand classify their mail and then run mass-check over it. They submit the output logs mass-check generates. Occasionally people review the submitted logs for obvious mistakes, but it is largely a trust system.

If you want to see the statistics from the last corpus run, check the STATISTICS.txt files that come in the SA tarball. It will tell you how many emails were used, and what the hit rates of all the rules were.

Procedure

Here's the process for generating the scores as of SpamAssassin 3.1.0:

1. heads-up

Inform everyone in advance on the -users and -dev lists that we will be starting mass-checks shortly, and they should get their corpora nice and clean (see CorpusCleaning) and sign up for RsyncAccounts.

2. announce mass-check

See MassCheck. The mass-check for both scoresets can be done in one command, e.g.

  cd masses
  mkdir -p spamassassin
  rm -f spamassassin/*

  cat > spamassassin/user_prefs
  bayes_auto_learn 0
  lock_method flock
  bayes_store_module Mail::SpamAssassin::BayesStore::SDBM
  use_auto_whitelist 0
  [hit Control-D]

  mass-check -j 4 --bayes --net --restart=400 --learn=35 --reuse [all targets]

  (note if a --after flag is part of the announcement, please add it as well)

Here's the full announcement text for this phase: RescoreDetails

We then take the log files rsync'd up to the server, and use those logs for all 4 score sets. The initial logs are for score set 3 (the fourth), sets 0, 1, and 2 can be generated from set 4 by stripping out the network tests and/or the Bayes tests.

3. allow several days to complete (it takes a really long time!)

Provide enough time, including a weekend if possible, giving people enough time to get around to running it given that they may be busy with day-job stuff.

4. generate scores for score sets

See RunningPerceptron.

Once this is complete, update rules/50_scores.cf with the generated scores.