1SA-LEARN(1) User Contributed Perl Documentation SA-LEARN(1)
2
6 sa-learn - train SpamAssassin's Bayesian classifier
7
9 sa-learn [options] [file]...
10 sa-learn [options] --dump [ all | data | magic ]
12 Options:
14 --ham Learn messages as ham (non-spam)
16 --spam Learn messages as spam
17 --forget Forget a message
18 --use-ignores Use bayes_ignore_from and bayes_ignore_to
19 --sync Synchronize the database and the journal if needed
20 --force-expire Force a database sync and expiry run
21 --dbpath <path> Allows commandline override (in bayes_path form)
22 for where to read the Bayes DB from
23 --dump [all|data|magic] Display the contents of the Bayes database
24 Takes optional argument for what to display
25 --regexp <re> For dump only, specifies which tokens to
26 dump based on a regular expression.
27 -f file, --folders=file Read list of files/directories from file
28 --dir Ignored; historical compatibility
29 --file Ignored; historical compatibility
30 --mbox Input sources are in mbox format
31 --mbx Input sources are in mbx format
32 --max-size <b> Skip messages larger than b bytes;
33 defaults to 256 KiB, 0 implies no limit
34 --showdots Show progress using dots
35 --progress Show progress using progress bar
36 --no-sync Skip synchronizing the database and journal
37 after learning
38 -L, --local Operate locally, no network accesses
39 --import Migrate data from older version/non DB_File
40 based databases
41 --clear Wipe out existing database
42 --backup Backup, to STDOUT, existing database
43 --restore <filename> Restore a database from filename
44 -u username, --username=username
45 Override username taken from the runtime
46 environment, used with SQL
47 -C path, --configpath=path, --config-file=path
48 Path to standard configuration dir
49 -p prefs, --prefspath=file, --prefs-file=file
50 Set user preferences file
51 --siteconfigpath=path Path for site configs
52 (default: /etc/mail/spamassassin)
53 --cf='config line' Additional line of configuration
54 -D, --debug [area=n,...] Print debugging messages
55 -V, --version Print version
56 -h, --help Print usage message
57
59 Given a typical selection of your incoming mail classified as spam or
60 ham (non-spam), this tool will feed each mail to SpamAssassin, allowing
61 it to 'learn' what signs are likely to mean spam, and which are likely
62 to mean ham.
63 Simply run this command once for each of your mail folders, and it will
65 ''learn'' from the mail therein.
66 Note that csh-style globbing in the mail folder names is supported; in
68 other words, listing a folder name as "*" will scan every folder that
69 matches. See "Mail::SpamAssassin::ArchiveIterator" for more details.
70 SpamAssassin remembers which mail messages it has learnt already, and
72 will not re-learn those messages again, unless you use the --forget
73 option. Messages learnt as spam will have SpamAssassin markup removed,
74 on the fly.
75 If you make a mistake and scan a mail as ham when it is spam, or vice
77 versa, simply rerun this command with the correct classification, and
78 the mistake will be corrected. SpamAssassin will automatically
79 'forget' the previous indications.
80 Users of "spamd" who wish to perform training remotely, over a network,
82 should investigate the "spamc -L" switch.
83
85 --ham
86 Learn the input message(s) as ham. If you have previously learnt
87 any of the messages as spam, SpamAssassin will forget them first,
88 then re-learn them as ham. Alternatively, if you have previously
89 learnt them as ham, it'll skip them this time around. If the
90 messages have already been filtered through SpamAssassin, the
91 learner will ignore any modifications SpamAssassin may have made.
92 --spam
94 Learn the input message(s) as spam. If you have previously learnt
95 any of the messages as ham, SpamAssassin will forget them first,
96 then re-learn them as spam. Alternatively, if you have previously
97 learnt them as spam, it'll skip them this time around. If the
98 messages have already been filtered through SpamAssassin, the
99 learner will ignore any modifications SpamAssassin may have made.
100 --folders=filename, -f filename
102 sa-learn will read in the list of folders from the specified file,
103 one folder per line in the file. If the folder is prefixed with
104 "ham:type:" or "spam:type:", sa-learn will learn that folder
105 appropriately, otherwise the folders will be assumed to be of the
106 type specified by --ham or --spam.
107 "type" above is optional, but is the same as the standard for
109 ArchiveIterator: mbox, mbx, dir, file, or detect (the default if
110 not specified).
111 --mbox
113 sa-learn will read in the file(s) containing the emails to be
114 learned, and will process them in mbox format (one or more emails
115 per file).
116 --mbx
118 sa-learn will read in the file(s) containing the emails to be
119 learned, and will process them in mbx format (one or more emails
120 per file).
121 --use-ignores
123 Don't learn the message if a from address matches configuration
124 file item "bayes_ignore_from" or a to address matches
125 "bayes_ignore_to". The option might be used when learning from a
126 large file of messages from which the hammy spam messages or spammy
127 ham messages have not been removed.
128 --sync
130 Synchronize the journal and databases. Upon successfully syncing
131 the database with the entries in the journal, the journal file is
132 removed.
133 --force-expire
135 Forces an expiry attempt, regardless of whether it may be necessary
136 or not. Note: This doesn't mean any tokens will actually expire.
137 Please see the EXPIRATION section below.
138 Note: "--force-expire" also causes the journal data to be
140 synchronized into the Bayes databases.
141 --forget
143 Forget a given message previously learnt.
144 --dbpath
146 Allows a commandline override of the bayes_path configuration
147 option.
148 --dump option
150 Display the contents of the Bayes database. Without an option or
151 with the all option, all magic tokens and data tokens will be
152 displayed. magic will only display magic tokens, and data will
153 only display the data tokens.
154 Can also use the --regexp RE option to specify which tokens to
156 display based on a regular expression.
157 --clear
159 Clear an existing Bayes database by removing all traces of the
160 database.
161 WARNING: This is destructive and should be used with care.
163 --backup
165 Performs a dump of the Bayes database in machine/human readable
166 format.
167 The dump will include token and seen data. It is suitable for
169 input back into the --restore command.
170 --restore=filename
172 Performs a restore of the Bayes database defined by filename.
173 WARNING: This is a destructive operation, previous Bayes data will
175 be wiped out.
176 -h, --help
178 Print help message and exit.
179 -u username, --username=username
181 If specified this username will override the username taken from
182 the runtime environment. You can use this option to specify users
183 in a virtual user configuration when using SQL as the Bayes
184 backend.
185 NOTE: This option will not change to the given username, it will
187 only attempt to act on behalf of that user. Because of this you
188 will need to have proper permissions to be able to change files
189 owned by username. In the case of SQL this generally is not a
190 problem.
191 -C path, --configpath=path, --config-file=path
193 Use the specified path for locating the distributed configuration
194 files. Ignore the default directories (usually
195 "/usr/share/spamassassin" or similar).
196 --siteconfigpath=path
198 Use the specified path for locating site-specific configuration
199 files. Ignore the default directories (usually
200 "/etc/mail/spamassassin" or similar).
201 --cf='config line'
203 Add additional lines of configuration directly from the command-
204 line, parsed after the configuration files are read. Multiple
205 --cf arguments can be used, and each will be considered a separate
206 line of configuration.
207 -p prefs, --prefspath=prefs, --prefs-file=prefs
209 Read user score preferences from prefs (usually
210 "$HOME/.spamassassin/user_prefs").
211 --progress
213 Prints a progress bar (to STDERR) showing the current progress. In
214 the case where no valid terminal is found this option will behave
215 very much like the --showdots option.
216 -D [area,...], --debug [area,...]
218 Produce debugging output. If no areas are listed, all debugging
219 information is printed. Diagnostic output can also be enabled for
220 each area individually; area is the area of the code to instrument.
221 For example, to produce diagnostic output on bayes, learn, and dns,
222 use:
223 spamassassin -D bayes,learn,dns
225 For more information about which areas (also known as channels) are
227 available, please see the documentation at:
228 C<http://wiki.apache.org/spamassassin/DebugChannels>
230 Higher priority informational messages that are suitable for
232 logging in normal circumstances are available with an area of
233 "info".
234 --no-sync
236 Skip the slow synchronization step which normally takes place after
237 changing database entries. If you plan to learn from many folders
238 in a batch, or to learn many individual messages one-by-one, it is
239 faster to use this switch and run "sa-learn --sync" once all the
240 folders have been scanned.
241 Clarification: The state of --no-sync overrides the
243 bayes_learn_to_journal configuration option. If not specified, sa-
244 learn will learn to the database directly. If specified, sa-learn
245 will learn to the journal file.
246 Note: --sync and --no-sync can be specified on the same
248 commandline, which is slightly confusing. In this case, the
249 --no-sync option is ignored since there is no learn operation.
250 -L, --local
252 Do not perform any network accesses while learning details about
253 the mail messages. This will speed up the learning process, but
254 may result in a slightly lower accuracy.
255 Note that this is currently ignored, as current versions of
257 SpamAssassin will not perform network access while learning; but
258 future versions may.
259 --import
261 If you previously used SpamAssassin's Bayesian learner without the
262 "DB_File" module installed, it will have created files in other
263 formats, such as "GDBM_File", "NDBM_File", or "SDBM_File". This
264 switch allows you to migrate that old data into the "DB_File"
265 format. It will overwrite any data currently in the "DB_File".
266 Can also be used with the --dbpath path option to specify the
268 location of the Bayes files to use.
269
271 There are now multiple backend storage modules available for storing
272 user's bayesian data. As such you might want to migrate from one
273 backend to another. Here is a simple procedure for migrating from one
274 backend to another.
275 Note that if you have individual user databases you will have to
277 perform a similar procedure for each one of them.
278 sa-learn --sync
280 This will sync any outstanding journal entries
281 sa-learn --backup > backup.txt
283 This will save all your Bayes data to a plain text file.
284 sa-learn --clear
286 This is optional, but good to do to clear out the old database.
287 Repeat!
289 At this point, if you have multiple databases, you should perform
290 the procedure above for each of them. (i.e. each user's database
291 needs to be backed up before continuing.)
292 Switch backends
294 Once you have backed up all databases you can update your
295 configuration for the new database backend. This will involve at
296 least the bayes_store_module config option and may involve some
297 additional config options depending on what is required by the
298 module. (For example, you may need to configure an SQL database.)
299 sa-learn --restore backup.txt
301 Again, you need to do this for every database.
302 If you are migrating to SQL you can make use of the -u <username>
304 option in sa-learn to populate each user's database. Otherwise, you
305 must run sa-learn as the user who database you are restoring.
306
308 (Thanks to Michael Bell for this section!)
309 For a more lengthy description of how this works, go to
311 http://www.paulgraham.com/ and see "A Plan for Spam". It's reasonably
312 readable, even if statistics make me break out in hives.
313 The short semi-inaccurate version: Given training, a spam heuristics
315 engine can take the most "spammy" and "hammy" words and apply
316 probabilistic analysis. Furthermore, once given a basis for the
317 analysis, the engine can continue to learn iteratively by applying both
318 the non-Bayesian and Bayesian rulesets together to create evolving
319 "intelligence".
320 SpamAssassin 2.50 and later supports Bayesian spam analysis, in the
322 form of the BAYES rules. This is a new feature, quite powerful, and is
323 disabled until enough messages have been learnt.
324 The pros of Bayesian spam analysis:
326 Can greatly reduce false positives and false negatives.
328 It learns from your mail, so it is tailored to your unique e-mail
329 flow.
330 Once it starts learning, it can continue to learn from SpamAssassin and
332 improve over time.
333 And the cons:
335 A decent number of messages are required before results are useful for
337 ham/spam determination.
338 It's hard to explain why a message is or isn't marked as spam.
339 i.e.: a straightforward rule, that matches, say, "VIAGRA" is easy
340 to understand. If it generates a false positive or false negative,
341 it is fairly easy to understand why.
342 With Bayesian analysis, it's all probabilities - "because the past
344 says it is likely as this falls into a probabilistic distribution
345 common to past spam in your systems". Tell that to your users!
346 Tell that to the client when he asks "what can I do to change
347 this". (By the way, the answer in this case is "use whitelisting".)
348 It will take disk space and memory.
350 The databases it maintains take quite a lot of resources to store
351 and use.
352
354 Still interested? Ok, here's the guidelines for getting this working.
355 First a high-level overview:
357 Build a significant sample of both ham and spam.
359 I suggest several thousand of each, placed in SPAM and HAM
360 directories or mailboxes. Yes, you MUST hand-sort this - otherwise
361 the results won't be much better than SpamAssassin on its own.
362 Verify the spamminess/haminess of EVERY message. You're urged to
363 avoid using a publicly available corpus (sample) - this must be
364 taken from YOUR mail server, if it is to be statistically useful.
365 Otherwise, the results may be pretty skewed.
366 Use this tool to teach SpamAssassin about these samples, like so:
368 sa-learn --spam /path/to/spam/folder
369 sa-learn --ham /path/to/ham/folder
370 ...
371 Let SpamAssassin proceed, learning stuff. When it finds ham and
373 spam it will add the "interesting tokens" to the database.
374 If you need SpamAssassin to forget about specific messages, use the
376 --forget option.
377 This can be applied to either ham or spam that has run through the
378 sa-learn processes. It's a bit of a hammer, really, lowering the
379 weighting of the specific tokens in that message (only if that
380 message has been processed before).
381 Learning from single messages uses a command like this:
383 sa-learn --ham --no-sync mailmessage
384 This is handy for binding to a key in your mail user agent. It's
386 very fast, as all the time-consuming stuff is deferred until you
387 run with the "--sync" option.
388 Autolearning is enabled by default
390 If you don't have a corpus of mail saved to learn, you can let
391 SpamAssassin automatically learn the mail that you receive. If you
392 are autolearning from scratch, the amount of mail you receive will
393 determine how long until the BAYES_* rules are activated.
394
396 Learning filters require training to be effective. If you don't train
397 them, they won't work. In addition, you need to train them with new
398 messages regularly to keep them up-to-date, or their data will become
399 stale and impact accuracy.
400 You need to train with both spam and ham mails. One type of mail alone
402 will not have any effect.
403 Note that if your mail folders contain things like forwarded spam,
405 discussions of spam-catching rules, etc., this will cause trouble. You
406 should avoid scanning those messages if possible. (An easy way to do
407 this is to move them aside, into a folder which is not scanned.)
408 If the messages you are learning from have already been filtered
410 through SpamAssassin, the learner will compensate for this. In effect,
411 it learns what each message would look like if you had run
412 "spamassassin -d" over it in advance.
413 Another thing to be aware of, is that typically you should aim to train
415 with at least 1000 messages of spam, and 1000 ham messages, if
416 possible. More is better, but anything over about 5000 messages does
417 not improve accuracy significantly in our tests.
418 Be careful that you train from the same source -- for example, if you
420 train on old spam, but new ham mail, then the classifier will think
421 that a mail with an old date stamp is likely to be spam.
422 It's also worth noting that training with a very small quantity of ham,
424 will produce atrocious results. You should aim to train with at least
425 the same amount (or more if possible!) of ham data than spam.
426 On an on-going basis, it is best to keep training the filter to make
428 sure it has fresh data to work from. There are various ways to do
429 this:
430 1. Supervised learning
432 This means keeping a copy of all or most of your mail, separated
433 into spam and ham piles, and periodically re-training using those.
434 It produces the best results, but requires more work from you, the
435 user.
436 (An easy way to do this, by the way, is to create a new folder for
438 'deleted' messages, and instead of deleting them from other
439 folders, simply move them in there instead. Then keep all spam in
440 a separate folder and never delete it. As long as you remember to
441 move misclassified mails into the correct folder set, it is easy
442 enough to keep up to date.)
443 2. Unsupervised learning from Bayesian classification
445 Another way to train is to chain the results of the Bayesian
446 classifier back into the training, so it reinforces its own
447 decisions. This is only safe if you then retrain it based on any
448 errors you discover.
449 SpamAssassin does not support this method, due to experimental
451 results which strongly indicate that it does not work well, and
452 since Bayes is only one part of the resulting score presented to
453 the user (while Bayes may have made the wrong decision about a
454 mail, it may have been overridden by another system).
455 3. Unsupervised learning from SpamAssassin rules
457 Also called 'auto-learning' in SpamAssassin. Based on statistical
458 analysis of the SpamAssassin success rates, we can automatically
459 train the Bayesian database with a certain degree of confidence
460 that our training data is accurate.
461 It should be supplemented with some supervised training in
463 addition, if possible.
464 This is the default, but can be turned off by setting the
466 SpamAssassin configuration parameter "bayes_auto_learn" to 0.
467 4. Mistake-based training
469 This means training on a small number of mails, then only training
470 on messages that SpamAssassin classifies incorrectly. This works,
471 but it takes longer to get it right than a full training session
472 would.
473
475 sa-learn and the other parts of SpamAssassin's Bayesian learner, use a
476 set of persistent database files to store the learnt tokens, as
477 follows.
478 bayes_toks
480 The database of tokens, containing the tokens learnt, their count
481 of occurrences in ham and spam, and the timestamp when the token
482 was last seen in a message.
483 This database also contains some 'magic' tokens, as follows: the
485 version number of the database, the number of ham and spam messages
486 learnt, the number of tokens in the database, and timestamps of:
487 the last journal sync, the last expiry run, the last expiry token
488 reduction count, the last expiry timestamp delta, the oldest token
489 timestamp in the database, and the newest token timestamp in the
490 database.
491 This is a database file, using "DB_File". The database 'version
493 number' is 0 for databases from 2.5x, 1 for databases from certain
494 2.6x development releases, 2 for 2.6x, and 3 for 3.0 and later
495 releases.
496 bayes_seen
498 A map of Message-Id and some data from headers and body to what
499 that message was learnt as. This is used so that SpamAssassin can
500 avoid re-learning a message it has already seen, and so it can
501 reverse the training if you later decide that message was learnt
502 incorrectly.
503 This is a database file, using "DB_File".
505 bayes_journal
507 While SpamAssassin is scanning mails, it needs to track which
508 tokens it uses in its calculations. To avoid the contention of
509 having each SpamAssassin process attempting to gain write access to
510 the Bayes DB, the token timestamps are written to a 'journal' file
511 which will later (either automatically or via "sa-learn --sync") be
512 used to synchronize the Bayes DB.
513 Also, through the use of "bayes_learn_to_journal", or when using
515 the "--no-sync" option with sa-learn, the actual learning data will
516 take be placed into the journal for later synchronization. This is
517 typically useful for high-traffic sites to avoid the same
518 contention as stated above.
519
521 Since SpamAssassin can auto-learn messages, the Bayes database files
522 could increase perpetually until they fill your disk. To control this,
523 SpamAssassin performs journal synchronization and bayes expiration
524 periodically when certain criteria (listed below) are met.
525 SpamAssassin can sync the journal and expire the DB tokens either
527 manually or opportunistically. A journal sync is due if --sync is
528 passed to sa-learn (manual), or if the following is true
529 (opportunistic):
530 - bayes_journal_max_size does not equal 0 (means don't sync)
532 - the journal file exists
533 and either:
535 - the journal file has a size greater than bayes_journal_max_size
537 or
539 - a journal sync has previously occurred, and at least 1 day has passed
541 since that sync
542 Expiry is due if --force-expire is passed to sa-learn (manual), or if
544 all of the following are true (opportunistic):
545 - the last expire was attempted at least 12hrs ago
547 - bayes_auto_expire does not equal 0
548 - the number of tokens in the DB is > 100,000
549 - the number of tokens in the DB is > bayes_expiry_max_db_size
550 - there is at least a 12 hr difference between the oldest and newest
551 token atimes
552 EXPIRE LOGIC
554 If either the manual or opportunistic method causes an expire run to
555 start, here is the logic that is used:
556 - figure out how many tokens to keep. take the larger of either
558 bayes_expiry_max_db_size * 75% or 100,000 tokens. therefore, the goal
559 reduction is number of tokens - number of tokens to keep.
560 - if the reduction number is < 1000 tokens, abort (not worth the
561 effort).
562 - if an expire has been done before, guesstimate the new atime delta
563 based on the old atime delta. (new_atime_delta = old_atime_delta *
564 old_reduction_count / goal)
565 - if no expire has been done before, or the last expire looks "weird",
566 do an estimation pass. The definition of "weird" is:
567 - last expire over 30 days ago
568 - last atime delta was < 12 hrs
569 - last reduction count was < 1000 tokens
570 - estimated new atime delta is < 12 hrs
571 - the difference between the last reduction count and the goal
572 reduction count is > 50%
573 ESTIMATION PASS LOGIC
575 Go through each of the DB's tokens. Starting at 12hrs, calculate
576 whether or not the token would be expired (based on the difference
577 between the token's atime and the db's newest token atime) and keep the
578 count. Work out from 12hrs exponentially by powers of 2. ie: 12hrs *
579 1, 12hrs * 2, 12hrs * 4, 12hrs * 8, and so on, up to 12hrs * 512
580 (6144hrs, or 256 days).
581 The larger the delta, the smaller the number of tokens that will be
583 expired. Conversely, the number of tokens goes up as the delta gets
584 smaller. So starting at the largest atime delta, figure out which
585 delta will expire the most tokens without going above the goal
586 expiration count. Use this to choose the atime delta to use, unless
587 one of the following occurs:
588 - the largest atime (smallest reduction count) would expire too many
590 tokens. this means the learned tokens are mostly old and there needs
591 to be new tokens learned before an expire can occur.
592 - all of the atime choices result in 0 tokens being removed. this means
593 the tokens are all newer than 12 hours and there needs to be new tokens
594 learned before an expire can occur.
595 - the number of tokens that would be removed is < 1000. the benefit
596 isn't worth the effort. more tokens need to be learned.
597 If the expire run gets past this point, it will continue to the end. A
599 new DB is created since the majority of DB libraries don't shrink the
600 DB file when tokens are removed. So we do the "create new, migrate old
601 to new, remove old, rename new" shuffle.
602 EXPIRY RELATED CONFIGURATION SETTINGS
604 "bayes_auto_expire" is used to specify whether or not SpamAssassin
605 ought to opportunistically attempt to expire the Bayes database. The
606 default is 1 (yes).
607 "bayes_expiry_max_db_size" specifies both the auto-expire token count
608 point, as well as the resulting number of tokens after expiry as
609 described above. The default value is 150,000, which is roughly
610 equivalent to a 6Mb database file if you're using DB_File.
611 "bayes_journal_max_size" specifies how large the Bayes journal will
612 grow before it is opportunistically synced. The default value is
613 102400.
614
616 The sa-learn command is part of the Mail::SpamAssassin Perl module.
617 Install this as a normal Perl module, using "perl -MCPAN -e shell", or
618 by hand.
619
621 spamassassin(1) spamc(1) Mail::SpamAssassin(3)
622 Mail::SpamAssassin::ArchiveIterator(3)
623 <http://www.paulgraham.com/> Paul Graham's "A Plan For Spam" paper
625 <http://www.linuxjournal.com/article/6467> Gary Robinson's f(x) and
627 combining algorithms, as used in SpamAssassin
628 <http://www.bgl.nu/~glouis/bogofilter/> 'Training on error' page. A
630 discussion of various Bayes training regimes, including 'train on
631 error' and unsupervised training.
632
634 "Mail::SpamAssassin"
635
637 The SpamAssassin(tm) Project <http://spamassassin.apache.org/>
638perl v5.16.3 2018-10-15 SA-LEARN(1)