Read Me
#
# README - master README for CRM114
#
# Copyright 2009 William S. Yerazunis.
# This file is under GPLv3, as described in COPYING.
#
Congratulations!!! You got this far. First things first.
THIS SOFTWARE IS LICENSED UNDER THE GNU PUBLIC LICENSE
IT MAY BE POORLY TESTED.
IT MAY CONTAIN VERY NASTY BUGS OR MISFEATURES.
THERE IS NO WARRANTY.
THERE IS NO WARRANTY WHATSOEVER!
A TOTAL, ALMOST KAFKA-ESQUE LACK OF WARRANTY.
Y O U A R E W A R N E D ! ! !
Now that we're clear on that, let's begin.
----- News This Release: -----
March 27, 2009 - BlameSteveJobs
This is a large set of changes. Old .css files will not be compatible
with this software, and must be rebuilt. The code is now fully 64-bit
compatibile. Support GNU regex has been entirely discarded. The
copyright has been updated to GPLv3, except for those functions moved
to the library, which are copyrighted as LGPLv3 to allow use in non-GPL
applications.
March 26, 2008 - BlameSentansoken
This release is basically a bugfix release and includes some (but not
all) fixes to the Hyperspace classifier, the vector tokenizer (VT),
and the neural network. Many thanks to Paolo, Ger, and the other mailing
list denizens who found the bugs. The neural network can now learn nonlinear
problems in reasonable times (such as the classic XOR problem). It
also includes a new example - alternating_example_neural.crm. This is
a good example of how to turn a single big example file into a number
of small ones (which is recommended). If you run it as "crm -t
alternating_example_neural.crm" you also get an amusing look at the
neural net as it trains (and no, I did not intentionally make it into
a "The Matrix" style output- that was actually the natural way to
express the state of the training). Upgrade to this version only if
you are using Hyperspace or the Neural Network.
Feb 13, 2008 - BlameJoeLangeway_VT
This release includes an improved vector tokenizer (that is, VT),
VT in the Hyperspace classifier, and a VT-enabled and much-improved
neural net as well. Bug fixes include the pR=5.0 change to work
versus microspams better, as well as some other bugfixes in both
the code and the test set. The neural network is now 8192 x 8 x 8 x 2
(that is 8192 retina slots feeding 8 input neurons; each of the input
neurons feeds all 8 of the hidden layer neurons, all 8 hidden layer
neurons feed both of the two output neurons) for minimum disk footprint
of just over half a megabyte before you staart adding documents. For
large data sets you should expand the network with -s NNN where NNN is
the number of retina slots (the rest of the net will scale reasonably;
the minimum is 1024 retina slots feeding 4 x 4 x 2). The algorithm
now uses both backpropagation and stochastic updates to avoid getting
caught in local minima.
As before, the neural net works best if you train in both in-class and
out-of-class (i.e. REFUTE) examples. Default is to update weights
only on the end of each training loop epoch; use BYCHUNK to update
after each example document. APPEND turns off autotraining after a
LEARN. The default cycle limits for LEARN are 5000 cycles if FROMSTART
is used and 250 cycles incrementally (experimentally determined to
work OK). To run another 250 cycles, just LEARN with a null text
to learn- this won't add a document but _will_ run 250 more backprop
cycles.
One trick that's not in the C code yet but that users can implement
in the CRM114 program is to alternate positive and negative chunks of
example text. By alternating examples, the BYCHUNK epoch learner can
converge on this rather difficult problem in not-unreasonable time
(an hour or three, depending on your CPU).
Running the neural net with -t can be illuminating; each line starts
with E (for epoch) and then the epoch number. After that, each "."
means one chunk of text that was successfully classified "strongly"
(3:1 neural response ratio). A # means a positive (in-class) example
that was classified as out of class, while a + means positive that was
classified positive but not strongly; an X means a negative
(out-of-class) example that was classified as a positive example, and
an x (lower case!) as an out-of-class example that was classified
"correctly" but not strongly so. It's rather fun to watch the system
as it learns the small differences between two very similar files (the
INTRO and QUICKREF files, taken in chunks of 1Kbyte+ characters rather
than as two big chunks as done in megatest. This takes a while, versus
the same problem run as only two examples, which runs in about
fifteen seconds.)
May 5, 2007 - BlameSpamConf
This is a mixed enhancement/bugfix release. You can now use
INSERT localfile and also INSERT [var_expanded_file] ; note that
the expansion only uses command-line and environment variables
such as :_env_HOME: (this isn't run-time expansion). The algebraic
math evaluator has now been switched over to a much nicer system.
Several other bugs are also stomped.
Feb 16, 2007 - BlameBaltar
This is a bugfix version. The much-maligned :datadir: is now gone;
everything is expected to be in either the :fileprefix: or the
:mailreaver: directories (don't combine them, otherwise if you reave
old files out of the reaver cache by age, your configuration files
will be the first to die!). You can now use start/len on INPUT on
stdin. Arithmetic now respects >=, <=, and != . Hyperspace has been
de-normalized, entropic classification is now somewhat smarter.
Mailreaver now really does put links in for prob_good and prob_spam.
Megatest is now more detailed and captures runtime errors better. The
"--" args have been removed from mailfilter, mailreaver, and
mailtrainer, because of problems they caused for non-X86 Linux users.
A PPC GCC compiler bug was found; PPC users need to use at least a
3.4+ GCC and preferably a 4.0++ GCC (sorry!). A cute little bug in
SYSCALL got stepped on (mostly), as well as one that affected any program
using two or more classifiers. A bad initialization in the
bit-entropy classifier was fixed, and entropy disk space required has now been
cut by another factor of almost 2. An incorrect result variable
parsing in TRANSLATE was fixed. Megatest now runs the local version
of crm114 (as built) rather than whatever's in /usr/bin, and the
variable :_pgm_text: contains a matchable copy of the program source,
post-preprocessing. Matthias's bugfixes are in. The highly experimental
Gatling sorter is included, but may exercise bugs (be warned!)
November 3, 2006 - BlameDalkey
This version is yet further bugfixes. It has Paolo's "BSD doesn't
have logl() so do this instead" bugfix, as well as the entropy sixfold
segfault bugfix, the missing mailfilter.cf backslash bugfix, the
DEFAULT doesn't default bugfix, and the bit-entropy FIR-prior
threshold is now automagically scaled to work decently for different
sized .ben files. The .BEN structure has been cleaned up (uses less
disk space at no accuracy cost!), and both Hyperspace and bit-entropy
pR values have been "rescaled" so that a pR within +10 to -10 gives
good results for thick threshold training (so now you don't have to
change thresholds when you change classifiers). The only remaining hitch
is that multi-class OSB sometimes gives wierdish results.
September 20, 2006 - BlameNico
This version is mostly a bugfix version - Hyperspace is now "fixed",
and some typos are cleaned up. This also introduces the _very_
experimental bit-entropic classifier. This classifier uses bits as
features, forms the Markov model of the features that minimizes
entropy of the system, and then uses the compression of that model as
the relative gauge of match quality. It's very VERY experimental, but
seems to work very well. It's not as fast as Hyperspace or OSB
(1/10th as fast), and uses significantly more memory (default is 64 megs
per class) but in long runs like TREC06-P at 90+Kmsgs it is about
two or three times as accurate. Currently it works best as
<entropy unique crosslink>
and trained with SSTTR (single-sided thick threshold training) with a
threshold of just 0.03 to 0.04 (because pR for bit-entropic is not yet
calibrated to the nominal scale.)
Read the long comments at the top of crm_bit_entropy.c to see how the
classifier works inside and how it's different from Andrej Bratko's
and Gordon Cormack's designs (in short, it uses < 1/50th the memory
and never overflows the node list thereby never needing to be dumped
and restarted; it also runs much faster because of this).
It's still HIGHLY experimental, so use bit-entropy only if you want to
be a lab rat. :-) The .ben (Bit-ENtropic) files are NOT compatible
with .css files or with any other classifier.
July 4, 2006 - BlameRobert
Mailreaver is working very well; this is a final cleanup release
for Mailreaver to go public as "supported and recommended". Robert wins
the award for this one as he whacked the sole remaining known bug in
the runtime system (and found another nearby!). The only thing I'm
contemplating further at this point is putting "autotrain anything in
the following region" back in. So, speak now if you find a bug, else
this one goes onto the webpage.
June 16, 2006 - ButterBeast
Release status: Mailreaver is now mostly debugged. No other changes
from BlameTheBeast except I don't know of anything that doesn't work
now, hence this is the Version of the Beast, with yummy butter added.
Lots of little twiddles and documentation fixes, too. (thanks,
Regis!)
June 6, 2006 - BlameTheBeast
Release Status: First testable "mailreaver-by-default" release
This is the first test release with a usable (we think) version of
mailreaver. As far as the user is concerned, mailreaver is pretty
much interchangeable with mailfilter, except that mailreaver by
default caches all email, tags all email with an sfid (Spam Filter ID)
which is all that's needed to recover the original, un-mangled text,
and thus doesn't need intact headers or mail editing in order to
train. This will make it much easier to write plugins and use
CRM114-based filters with MTAs and MUAs that insist on screwing with
the headers.
That said, this is *still* an experimental release; be aware
of that if you install it. There will be bugs and rough spots;
be prepared to endure, discuss, and help solve.
[[ the one-character memory leak is still here in crm_var_hash_table.c,
if you find it, please let me know!!! The bug is completely benign in
any but an intentional "tickler" program, and at worst simply forces
an error after a few million reclamations, so it is highly unlikely to
affect real apps, but it's an ugly wart. -- WSY ]]
------------ HOW TO USE MAILREAVER ( instead of mailfilter ) -----
Mailreaver.crm is the "next generation" mailfilter; use it instead of
mailfilter.crm. (and, please note, right now mailreaver.crm is "field
test" but the plan is that it will eventually become the recommended
production system, and mailfilter.crm will become nothing more than
a good-looking corpse.)
Mailreaver.crm takes all of the same flags and command lines (or at
least it should), and the default is to use cache. It also has the
new option --dontstore which means "do NOT put this text into the
reavercache". Mailreaver.crm also has the possibility of becoming
faster than mailfilter.crm, because it doesn't need to JIT any of the
learning code unless learning is actually needed. (future plan: mailreaver
will become a "wrapper" program for a set of mail service programs)
Mailreaver.crm uses the same old mailfilter.cf configuration file;
those things that don't make sense don't get used, and a few new
options (like :thickness:) do get used. IT IS RECOMMENDED that
you save your old mailfilter.cf file as mailfilter.cf.old, and use the
NEW one which has the new options already set up. (defaults are
OSB Unique Microgroom for the classifier cnfiguration, a thick
threshold of 10 pR units, simple word tokenization, a decision
length of 16000 bytes, 0/0/1 exit codes for good/spam/program
fault, cacheing enabled, rewrites enabled, a spam-flag subject string of
ADV:, and an unsure flag subject string of UNS:.
The big advantage of mailreaver.crm is that now all mailtraining
goes through mailtrainer.crm, which opens the door to some very
powerful training techniques.
Note - if you use the -u option, you must make sure that
mailtrainer.crm, maillib.crm and shuffle.crm are in the directory you
are -u'ing into. ( --fileprefix is not used for this location).
Note 2 - this version does not include a "cache flusher" option; the
full text of emails stored in the cache will remain there until you
manually delete them; one month is probably OK for emails not in the
knownspam or knowngood directory but keep anything in those directories
(don't worry, we hardlink to the files on *nix systems, and make copies
on Windozen). You can do this cache flushing with a cron job easily
enough, if you really need the disk space that badly..
April 22, 2006 - ReaverReturn
Release Status: for testing and bug chasing.
This release is primarily to verify we've whacked a few bugs in the
"css version" and reserved-bucket code, as well as improved the
mailtrainer code and documentation. This new version also has the
DEBUG statement (drops you immediately to the debugger from program
code).
It is suggested you install this release ONLY if you have an outstanding
issue or bug. If you are currently happy and are not out "looking for
a fight", this is not the release for you.
Feb 6, 2006 - ReaversSecondBreakfast
Release status: Looks good, but has major new (mal)features!
This is the transition release to the new "Reaver" format for storing
and caching the learned texts. The "Reaver" format keeps all incoming
mail in a maildir-like cache (but _not_ your maildir; we don't touch
that, and if you don't use maildir, that's just fine). The advantage
is that the incoming mail is saved exactly as it needs to be trained,
and a CacheID is added to the mail that you finally see.
As long as the CacheID header is intact, you can just reference this
stored copy (assuming it hasn't been purged away) You say "use the
cached text" with --cache for command line users as in:
bash: crm mailfilter.crm --cache --learnspam < text_with_cacheid_header
or (for mail-to-myself users - forward to yourself with the CacheID
header intact):
command my_secret_password learnspam cache
and you will train *exactly* the text should be trained, with no worry
as to whether something in the system added headers, deleted
suspicious text, or whatever. The "Reaver" cache hardlinks your
trained data into "known spam" and "known nonspam" subdirectories, so
even when the cache gets purged, you don't lose training data.
Note that the :thickness: parameter now is meaningful! Anything
scoring less than this amount (plus or minus) is tagged as UNSURE
and will be delivered along the "good" mail path with the extra
header of "Please train me!" Default thickness is 10 pR units, which
gives really good results for me.
Another advantage of the Reaver cache is that you can use
mailtrainer.crm to run advanced training (like DSTTTR) on the full
contents of your "known" and "probably" directories, to get really
fast accurate .css files. After testing, we've changed the defaults
on mailtrainer to only do one pass of DSTTTR.
Notes to MeetTheReavers users who have already used mailtrainer.crm -
we've added the --reload option and the default repeat count is now
1 pass (rather than 5).
------ How To Use mailtrainer.crm -----------
Mailtrainer.crm is a bulk mail statistics file trainer. It allows you to
feed in bulk training of your own example email and get an optimized
filter in a very few minutes - or to test variations if you want to
play around with the different classifiers, thickness settings, etc.
Mailtrainer by default uses whatever settings are in your current
mailfilter.cf file, so you'll get .css files that are optimized for
your standard setup including mime decoding, normalization, classifier
flags, etc.
Mailtrainer.crm uses DSTTTTR (Double Sided Thick Threshold Training with
Testing Refutation) which is something I didn't come up with (Fidelis
is the top of the list of suspects for this). The good news is that this can
more than double accuracy of OSB and similar classifiers. I'm seeing
better than 99.7% accuracy with mailtrainer's DSTTTTR on the 2005 TREC
SA test corpus, with 10-fold validation and the default thick threshold
of 5.0 pR units and the classifier set to OSB Unique Microgroom.
This is substantially better than any other result I've gotten. Six of
the ten runs completed with only ONE error out of the 600 test cases.
It is safe to run mailtrainer.crm repeatedly on a .css fileset and
training data; if the data doesn't need to be trained in, it won't be.
All you will waste is CPU time. The examples need to be one example
per file. The closer these files are to what mailfilter.crm wil see
in real life the better your training will be. Preferably the headers
and text will be complete, intact, and unmutilated.
The mailtrainer.crm options are as follows. You *must* provide --spamdir
and --gooddir; the other flags are optional.
Required:
--spamdir=/directory/full/of/spam/files/one/per/file
--gooddir=/directory/full/of/good/files/one/per/file
Optional:
--thick=N - thickness for thick-threshold training- this
overrides the thickness in your mailfilter.cf file.
--streak=N - how many successive correct classifications
before we conclude we're done. Default is 10000.
--repeat=N - how many times to pass the training set through
the DSTTTR training algorithm.
--reload - if marked, then whenever either the spam or good
mail training set is exhausted, reload it
immediately from the full training set. The default is
no reload- to run alternate texts and when
either good or spam texts are exhausted, run
only the other type until all of those have been
run as well. --reload works a little better
for accuracy but takes up to twice as long.
--worst=N - run the entire training set, then train
only the N worst offenders, and repeat. This is
excruciatingly slow but produces very compact
.css files. Use this only if your host machine
is dreadfully short of memory. Default is not
to use worst-offender training. N=5% of your total
corpus works pretty well.
--validate=regex_no_slashes - Any file with a name that matches
the regex supplied is not used for training; instead
it's held back and used for validation of the
new .css files. The result will give you an
idea of how well your .css files will work.
Example:
Here's an example. We want to run mailtrainer.crm against a bunch of
examples in the directory ../SA2/spam/ and ../SA2/good/. Quit when
you get 4000 tests in a row correct, or if you go through the entire
corpus 5 times. Use DSTTTR, with a training thickness of just .05 pR
units. Don't train on any filename that contains a "*3.*" in the
filename; instead, save those up and use them as a "test corpus" for
N-fold validation, and print out the expected accuracy. For this
particular corpus, that's about 10% of the messages.
Here's the command:
crm mailtrainer.crm --spamdir=../SA2/spam/ --gooddir=../SA2/good/ \
--repeat=5 --streak=4000 --validate=[3][.] --thick=0.05
This will take about eight minutes to run on the SA2 (== TREC SA) corpus
of about 6000 messages; 1000 messages a minute is a good estimate
for 5 passes of DSTTTTR training.
Notes:
* If the .css statistics files don't exist, they will be created for you.
* If the first test file comes back with a pR of 0.0000 exactly, it is
assumed that these are empty .css statistics files, and that ONE file
will be trained in, simply to get the system "off center" enough that
normal training can occur. If there is anything already in the files,
this won't happen.
* When running N-fold validation, if the filenames are named as in
the SA2 corpus, there's an easy trick: use a regex like [0][.] for
the first run, [1][.] for the second run, [2][.] for the third, and
so on. Notice that this a CRM114-style regex, and _not_ a BASH-style
file globbing as *3.* would be.
* If you want to run N-fold validation, you must remember to delete the
.css files after each run, otherwise you will not get valid results.
* N-fold validation does NOT run training at all on the validation set,
so if you decide you like the results, you can do still better by
running mailtrainer.crm once again, but not specifying --validate.
That will train in the validation test set as well, and hopefully
improve your accuracy still more.
December 31, 2005 - BlameBarryAndPaolo
This is a bugfix/bugchase release; it should remedy or at least
yield information on the strange var-name bug that a few people with
very-long-running demons have encountered. It also has some bugfixes
(especially in the W32 code, from both Barry and JesusFreke) and in the
microgroomer, bugfixes, and accuracy improvements.
Upgrade is advised if you are having bug problems, otherwise, it's
not that big an issue.
October 2, 2005 - BlameRaulMiller
This is a new-features release. The new feature is the TRANSLATE
statement- it is like tr() but allows up- and down- ranges, repeated
ranges, inversion, deletion, and uniquing. The Big Book has been
updated with TRANSLATE and Asian Language stuff (it shows as
"highlighted in yellow" in the PDF but it hasn't been indexed yet..)
Next version of the code will have JesusFreke's Windows latest
bugfixes, and the fixes to microgrooming (they didn't make this
release; sorry)
September 10, 2005 - BlameToyotomi
This is mostly a docmentation/bugfix release. New features are that
the Big Book is now very close to "final quality", some improvements
in speed and orthogonality of the code, bugfixes in the execution
engine and in mailfilter.crm, and allowing hex input and output in
EVAL math computations ( the x and X formats are now allowed, but are
limited to 32-bit integers; this is for future integration of libSVM
to allow SVM-based classifiers).
Upgrade is recommended only if you're coding your own filters (to
get the new documentation) or if you are experiencing buggy behavior
with prior releases.
July 21, 2005 - BlameNeilArmstrong
This release is an upgrade for two new classification options -
UNIGRAM and HYPERSPACE. It also contains some bugfixes (including
hopefully a bugfix for the mmap error-catching problem), and the
new flag DEFAULT for conditionally setting an ISOLATED var only if it
hasn't had any value yet.
The <default> flag for ISOLATE is designed to be an executable
statement, rather than a compile-time default value. If a variable
has never been set in the program, ISOLATE <default> will set it;
otherwise that ISOLATE statement does nothing. This is inspired by
JesusFreke's <once> patch).
Using <unigram> classification effectively turns CRM114 into a normal
Bayesian classifier, as <unigram> tells the classifers to use only
unigrams (single words) as features. This is so people can do an
apples-to-apples comparison of CRM114's Markovian, OSB Winnow, and OSB
Hyperspace classifiers versus Typical Bayesian classification and not
have to write tons of glue code; just change your classify/learn flag.
(note - <unigram> does not cause a top-N decision list as used in A
Plan For Spam - CRM114 still does not throw anything away)
The other (and maybe bigger) news is the new hyperspatial classifier.
The hyperspatial classifier is most easily explained by analogy to
astronomy. Each known-type example document represents a single star
in a hyperspace of about four billion dimensions. Similar documents
form clusters of light sources, like galaxies of stars. Each class of
document is therefore represented by several galaxies of stars (each
galaxy being documents that are hyperspatially very similar). The
unknown document is an observer dropped somewhere into this
four-billion-dimensional hyperspace; whichever set of galaxies appears
to be brighter (hence closer) on this observer is the class of the
unknown document.
What's amazing is that this hyperspace approach is not only faster
than Bayesian OSB, it's also more accurate (26 errors total in the
final 5000 texts in the SA ten-way shuffle, better even than Winnow)
and uses only about 1/40th the disk space of Markovian (300 Kbytes
v. 12 Mbytes for the SA test corpus) and about 6 times faster ( 26
seconds seconds versus 3min 11 sec on the same Pentium-M 1.G GHz for
one pass, in "full demonized" mode).
It should be pointed out that the fully demonized hyperspace code
takes just 26 seconds for the 4147-text SA corpus is 6.2 milliseconds
per text classified, including time to do all learning.
The only downside of Hyperspace math is that it needs single-sided
thick threshold training (SSTTT), with a thickness of 0.5 pR units.
With straight TOE, it's still barely faster than Markovian but not as
accurate. It still only uses 400K per .css file though,
Best accuracy and speed on the SA corpus is achieved with only two
terms of the OSB feature set- digram and trigram; that's the default
for hyperspace, and in the code. The only downside is that this is
still an experimental design; use it for fun, not for production, as
the file format will undoubtedly change in the future and if you don't
keep your training set as individual disjoint documents you'll have to
start again. Activate it by using <hyperspace> as the classify/learn
control flag; you can also use <unique> and <unigram> if you want.
As usual, "prime" each of the Hyperspace statistics files by LEARNing
a very short example text into each one. This creates the files with
the proper headings, etc. Alternatively, create a null hyperspace file
with the commands:
head \/dev\/zero --bytes=4 > spam.css
head \/dev\/zero --bytes=4 > nonspam.css
---- What YOU Should Do Now -----
Contents:
1) "What Do You Want?"
2) If you want to write programs...
3) How to "make" CRM114
1) "What Do You Want?"
**** If you just want to use CRM114 Mailfiltering, print out the
CRM114_Mailfilter HOWTO and read _THAT_. Really; we will help a LOT.
The instructions in the HOWTO are much more in-depth and up to
date than whatever you can glean from here.
2) If you want to write programs, read the introduction file INTRO.txt
That file will get you started.
Remember, this is a wierd-ass language, you _don't_ understand it
yet. (okay, wiseguy, what does a "LIAF" statement do? :-) )
Then, print out and read the QUICKREF.txt (quick reference card).
You'll want this by your side as you write code until you get
used to the language.
3) CRM114 (as of this writing) does not have a fully functional
.config file. There is a beta version, but it doesn't work
on all systems.
Until that work is finished, you have a couple of recommended options:
1) run the pre-built binary release,
or
2) use the pre-built Makefile to build from sources.
Caution: if you are building from sources, you should install
the TRE regex library ***first***. TRE is the recommended regex
library for CRM114 (fewer bugs and more features than Gnu Regex).
You will need to give the TRE .configure the --enable-static
argument, i.e. " ./configure --enable-static " .
The reason for the "static" linking recommendation is that many people
don't have root on their site's mail server and so cannot install the
TRE regex library there. By making the default standard CRM114 binary
standalone (static linked), it's possible for a non-root user to run
CRM114 on the host without deep magic.
Here are some useful Makefile targets:
"make clean" -- cleans up all of the binaries that you have
that may or may not be out of date. DO NOT
do a "make clean" if you're using a binary-only
distribution, as you'll delete your binaries!
"make all" -- makes all the utilities (both flavors of crm114,
cssutil, cssdiff, cssmerge), leaving them in
the local directory.
"make install" -- as root will build and install CRM114 with
the TRE REGEX libraries as /usr/bin/crm . If you
want a "stripped install" (cuts the binary sizes
down by almost a factor of two) you will need
to edit the Makefile- look at the options for
"INSTALLFLAGS"
"make uninstall" -- undoes the installation from "make install"
"make megatest" -- this runs the complete confidence test of
your installed CRM114. Not every code path can
be tested this way (consider- how do you test multiple
untrappable fatal errors? :) ), but it's a good
confidence test anyway.
"make install_gnu -- as root will build and install CRM114 with
the older GNU REGEX libraries. This is
obsolete but still provided for those of us
with a good sense of paranoid self-preservation.
Not all valid CRM114 programs will run under GNU;
the GNU regex library has... painful issues.
"make install_binary_only -- as root, if you have the binary-only
tarball, will install the pre-built, statically
linked CRM114 and utilities. This is very handy if
you are installing on a security-through-minimalism
server that doesn't have a compiler installed.
"make install_utils" -- will build the css utilities "cssutil",
"cssdiff", and "cssmerge".
cssutil gives you some insight into
the state of a .css file, cssdiff lets you
check the differences between two .css files,
and cssmerge lets you merge two css files.
"make cssfiles" - given the files "spamtext.txt" and
"nonspamtext.txt", builds BRAND NEW spam.css
and nonspam.css files.
Be patient- this can take about 30seconds per 100Kbytes
of input text! It's also destructive in a sense - repeating
this command with the same .txt files will make the
classifier a little "overconfident". If your .txt
files are bigger than a megabyte, use the -w option to
increase the window size to hold the entire input.
*******************************************************
*** Utilities for looking into .css files.
This release also contains the cssutil utility to take a look at
and manage .css spectral files used in the mailfilter.
Section 8 of the CRM114_Mailfilter_HOWTO tells how to use these
utilities; you _should_ read that if you are going to use the
CLASSIFY funtion in your own programs.
If you are using the OSB classifier, you can use the cssutil program
because the file formats of default SBPH Markov and OSB Markov
are compatible. However, the OSBF classifier, the Winnow
classifier, and the Corellative classifier all have their own
(incompatible) file formats. For the OSBF classifier, you can
use the osbf-util program to look inside; for Winnow you can use
the cssutil program but only the bucket use counts and the chain
length counts will be correct.
*** How to configure the mailfilter.crm mail filter:
The instructions given here are just a synopsys- refer to the CRM114
Mailfilter HOWTO, included in your distribution kit.
You will need to edit mailfilter.cf , and perhaps a few other
files. The edits are quite simple, usually just inserting a username,
a password, or choosing one of several given options.
*** The actual filtering pipeline:
- If you have requested a safety copy file of all incoming mail, the
safety copy is made.
- An in-memory copy of the incoming mail is made; all mutilations
below are performed on this copy (so you don't get a ravaged
tattered sham of email, you get the real thing)
- If you have specified BASE64 expansion (default ON), any base64 attachments
are decoded.
- If you have specified undo-interruptus, then HTML comments are
removed.
- The rewrites specified in "rewrites.mfp" get applied. These
are strictly "from>->to" rewrites, so that your mail headers
will look exactly like the "canonical" mail headers that were
used when the distribution .css files were built. If you build
your own .css files from scratch, you can ignore this.
- Filtration itself starts with the file "priolist.mfp' . Column 1
is a '+' or '-' and indicates if the regex (which starts in column 2)
should force 'accept' or 'reject' the email.
- Whitelisting happens next, with "whitelist.mfp" . No need for a + or
a - here; every regex is on it's own line and all are whitelisting.
- Blacklisting happens next, with "blacklist.mfp" . No need for + or -
here either- if the regex matches, the mail is blacklisted.
- Failing _that_, the sparse binary polynomial hash with Markovian weights
(SBPH/Markov) matching system kicks in, and tries to figure out
whether the mail is good or not. SBPH/Markov matching can occasionally
make mistakes, since it's statistical in nature. You actually have four
matchers available- the default is SBPH/Markov, but there's also
an OSB/Markov, an OSB/Winnow, and a full correlator.
- The mailfilter can be remotely commanded. Commands start in
column 1 and go like this (yes, command is just that- the letters
c o m m a n d, right at the start of the line. You mail a message
with the word command, the command password, and then a command word
with arguments, and the mailfilter does what you told it.
command yourmailfilterpassword whitelist string
- auto-accepts mail containing the whitelist string.
command yourmailfilterpassword blacklist string
- auto-rejects mail containing the blacklisted string
command yourmailfilterpassword spam
- "learns" all the text following this command line as spam, and will
reject anything it gets that is "like" it. It doesn't
"learn" from anything above this command, so your headers
(and any incoming headers) above the command are not considered
part of the text learned. It's up to your judgement what part
of that text you want to use or not.
command yourmailfilterpassword nonspam
- "learns" all the text following this line as NOT spam, and will
accept any mail that it gets that is "like" it. Like
learning spam, it excludes anything above it in the file
from learning.
The included five files (priolist.mfp, whitelist.mfp, blacklist.mfp,
spam.css and nonspam.css) are meant for example, mostly.
- rewrites.mfp is a set of rewrites to be applied to the
incoming mail to put it in "canonical" form.
You don't _need_ to edit this file to match your
local system names, but your out-of-the-box
accuracy will be improved greatly if you do.
- priolist.mfp is a set of very specific regexes, prefixed by +
or -. These are done first, as highest priority.
- whitelist.mfp is mailfilterpatterns that are "good". No line-spans
allowed- the pattern must match on one line.
- blacklist.mfp is mailfilterpatterns that are "bad". Likewise,
linespanning is not allowed (by default). Entries in
this file are all people who spam me so much I started to
recognize their addresses... so I've black-holed them.
If you like them, you might want to unblackhole them.
- spam.css and nonspam.css: These are large files and as of
2003-09-20, are included only in the .css kits. CRM
.css files are "Sparse Spectra" files and they
contain "fingerprints" of phrases commonly seen in
spam and nonspam mail. The "fingerprint pipeline" is
currently configured at five words, so a little spam
matches a whole lot of other spam. It is difficult but
not impossible to reverse-engineer the spam and nonspam
phrases in these two files if you really want to know.
To understand the sparse spectrum algorithm, read the
source code (or the file "classify_details.txt");
the basic principle is that each word is
hashed, words are conglomerated into phrases, and
the hash values of these phrases are stored in the
css file. Matching a hash means a word or phrase under
consideration is "similar to" a message that has been
previously hashed. It's usually quite accurate, though
not infallable.
The filter also keeps three logs: one is "alltext.txt", containing a
complete transcript of all incoming mail, the others are spamtext.txt
and nonspamtext.txt; these contain all of the text learned as spam
and as nonspam, respectively (quite handy if you ever migrate between
versions, let me assure you).
Some users have asked why I don't distribute my learning text, just
the derivative .css files: it's because I don't own the copyright on
them! They're all real mail messages, and the sender (whoever that
is) owns the copyright, not me (the recipient). So, I can't publish
them. But never fear, if you don't trust my .css files to be clean,
you can build your own with just a few day's spam and nonspam traffic.
Your .css files will be slightly different than mine, but they will
_precisely_ match your incoming message profile, and probably be
more accurate for you too.
A few words on accuracy: there is no warranty- but I'm seeing typical
accuracies > 99% with only 12 hours worth of incoming mail as example
text. With the old (weak, buggy, only 4 terms) polynomials, I got a
best case of 99.87% accuracy over a one-week timespan. I now see
quality averaging > 99.9% accuracy (that is, in a week of ~ 3000 messages,
I will have 1 or 2 errors, usually none of them significant.
Of course, this is tuned to MY spam and non-spam email mixes; your
mileage will almost certainly be lower until you teach the system what
your mail stream looks like.
============== stop here stop here stop here ==========
----- Old News -----
Jan 17, 2006 - BlameTheReavers
This is a big new functionality release- we include mailtrainer.crm as
well as changing the default mailfilter.crm from Markovian to OSB.
This new mailtrainer program is fed directories of example texts (one
example per file), and produces optimized satistics files matched to
your particular mailfilter.cf setup (each 1meg of example takes about
a minute of CPU). It even does N-fold validation. Default training
is 5-pass DSTTTTR (a Fidelis-inspired improvement of TUNE) with a
thick threshold of 5.0 pR units. Worst-offender DSTTTTR training as a
(very slow) option. There are also speedups and bugfixes throughout
the code. Unless you really like Markovian, now is a good time to
think about saving your old .css files and switching over to the new
default mailfilter.crm config that uses OSB unique microgroom. Then
run mailtrainer.crm on your saved spam and good mail files, and see
how your accuracy jumps. I'm seeing a four-fold increase in accuracy
on the TREC SA corpus; this is hot stuff indeed.
Version CRM114-20050511.BlameMercury
This version is a documentation primary release - CRM114 Revealed
(.pdf.gz, 240 pages) is now available for download. BlameMercury has lots of
bugfixes and only three extensions - you can now demonize minions onto
pipes, you can re-execute failing commands from a TRAP finish, and you
can now use a regex directly as a var-restriction subscripting action,
so [ :my_var: /abc.*xyz/ ] gets you the matching substring in the :my_var:
variable. Var-restriction matches do NOT change the "previous MATCH" data on
a variable.
Version 20050415.BlameTheIRS (TRE 0.7.2)
Math expressions can now set whether they are algebraic or RPN by a
leading A or N as the first character of the string. Listings are now
controllable with -l N, from simple prettyprinting to full JIT parse.
A bug in the microgroomer that causes problems when both microgrooming
and unique were used in the same learning scenario was squashed in
Markovian, OSB, and Winnow learners (it remains in OSBF). Dependency
on formail (part of procmail) in the default mailfilter.crm has been
removed. A cleaner method of call-by-name, the :+: indirection-fetch
operator, has been activated. The var :_cd: give the call depth for
non tail recursive routines. Minor bugs have been fixed and minor
speedups added. "make uninstall" works. Documentation of regexes has
been improved. Cached .css mapping activated. Win32 mmap adapters
inserted.
Version 20041231.BlameSanAndreas (TRE 0.7.2)
Major topics: New highest-accuracy voodoo OSBF classifier (from
Fidelis Assis), CALL/RETURN now work less unintuitively, SYSCALL can
now fork the local process with very low overhead (excellent for
demons that spawn a new instance for each connection), and of course
bug fixes around the block. Floating point now accepts exponential
notation (i.e. 6.02E23 is now valid) and you can specify output
formatting. MICROGROOM is now much smarter (thanks again to Fidelis)
and you can now do windowing BYCHUNK.
This new revision has Fidelis Assis' new OSBF local confidence factor
generator; with the OSB front end and single-sided threshold training
with pR of roughly 10, it is more than 3 times more accurate and 6
times faster than straight SBPH Markovian and uses 1/10th the file
space.
The only downsides are that the OSBF file format is incompatible and
not interconvertable between .css files and OSBF .cfc files, and that
you _must_ use single-sided threshold training to achieve this
accuracy. Single-sided threshold training means that if a particular
text didn't score above a certain pR value, it gets trained even if it
was classified correctly. For the current formulation of OSBF,
training all nonspams with pR's less than 10, and all spams with pR's
greater than -10 yields a very impressive 17 errors on the SA torture
test, versus 42 errors with Winnow (doublesided threshold with a
threshold of 0.5) and straight Markovian (54 errors with Train Only
Errors training, equivalent to singlesided training with a threshold
of zero)
We also have several improvement in OSB, which gets down to 22 errors on
the same torture test, again with the same training regimen (train if
you aren't at least 10 pR units "sure", and _is_ upward compatible
with prior OSB and Markovian .css files, and with the same speed
as OSBF. It also doesn't use the "voodoo exponential confidence factor",
so it may be a more general solution (on parsimony grounds); it has
similar properties to OSB. (though there is a known bug that feature
counts are all == 1 for now, but this doesn't hurt anything)
CLASSIFY and LEARN both default to the obvious tokenize regex of
/[[:graph:]]+/.
CALL now takes three parameters:
CALL /:routine_to_call:/ [:downcall_concat:] (:return_var:)
The routine itself gets one parameter, which is the concatenation of
all downcall_concat args (use a MATCH to shred it any way you want).
RETURN now has one parameter:
RETURN /:return_concat:/
The values in the :return_concat: are concatenated, and returned to the
CALL statement as the new value of :return_var: ; they replace whatever was
in :return_var:
SYSCALL can now fork the local process and just keep running; this saves
new process invocation time, setup time, and time to run the first pass
of the microcompiler.
See the examples in call_return_test.crm for these new hoop-jumping tricks.
WINDOW now has a new capability flag - BYCHUNK mode, specifically for
users WINDOWing through large blocks of data. BYCHUNK reads
as large a block of incoming data in as is available (modulo limits of
the available buffer space), then applies the regex. BYCHUNK assumes
it's read all that will be available (and therefore sets EOF), so repeated
reads will need to use EOFRETRY as well.
Version 20041110.BlameFidelisMore (TRE 0.7.0)
This new revision has Fidelis Assis' new OSBF local confidence factor
generator; with the OSB front end and single-sided threshold training
with pR of roughly 10, it is more than 3 times more accurate and 6
times faster than straight SBPH Markovian and uses 1/10th the file
space.
The only downsides are that the OSBF file format is incompatible and
not interconvertable between .css files and OSBF .cfc files, and that
you _must_ use single-sided threshold training to achieve this
accuracy. Single-sided threshold training means that if a particular
text didn't score above a certain pR value, it gets trained even if it
was classified correctly. For the current formulation of OSBF,
training all nonspams with pR's less than 10, and all spams with pR's
greater than -10 yields a very impressive 17 errors on the SA torture
test, versus 42 errors with Winnow (doublesided threshold with a
threshold of 0.5) and straight Markovian (54 errors with Train Only
Errors training, equivalent to singlesided training with a threshold
of zero)
We also have several improvement in OSB, which gets down to 22 errors on
the same torture test, again with the same training regimen (train if
you aren't at least 10 pR units "sure", and _is_ upward compatible
with prior OSB and Markovian .css files, and with the same speed
as OSBF. It also doesn't use the "voodoo exponential confidence factor",
so it may be a more general solution (on parsimony grounds); it has
similar properties to OSB. (though there is a known bug that feature
counts are all == 1 for now, but this doesn't hurt anything)
CLASSIFY and LEARN both default to the obvious tokenize regex of
/[[:graph:]]+/.
Version 20040921.BlameResidentWeasel
Bugs stomped: several in the exit routine code, as well as fixes for
detecting minion creation failures. SYSCALL can now do forking of the
currently executing code; CALL is now implemented and provides
CORBA-like argument transfer. A missing INSERT file in a .crm program
is now a trappable error
Version 20040815.BlameClockworkOrange
Start/Length operators in match qualification are now working (same
syntax as seek/length operators in file I/O), -v (and :_crm_version:)
now also ID the regex engine type and version, and several bugs
(including two different reclaimer bugs) have now been stomped. Other
code cleanups and documentation corrections have been done.
Version 20040808.BlamePekingAcrobats
This is a bugfix/performance improvement release. The bugs are minor edge
cases, but better _is_ better. SYSCALL now has better code for
<async> and <keep> (async is now truly "fire and forget"; keep
keeps the process around without losing state, and default processes
will now not hang forever if they overrun the buffer.)
Documentation has been improved. Both OSB and Winnow are now both
faster and more accurate (as bugs were removed). A particularly nasty
bug that mashed isolated vars of zero length was quashed. -D and -R
(Dump and Restore) are available in cssutil for moving .css files between
different-endian architectures.
Version 20040723.BlameNashville
This is a major bugfix release with significant impact on accuracy,
especially for OSB users. There's now a working incremental
reclaimer, so there's no more ISOLATE-MATCH-repeat bug (feel free to
isolate and match without fear of memory leakage). The "exit 9"
bug has been fixed (at least I can no longer coerce it to appear)-
users of versions after 20040606-BlameTamar should upgrade to this
version.
Version CRM114-20040625-BlameSeifkes
Besides the usual minor bugfixes (thanks!) there are two big new features
in this revision:
1) We now test against ( and ship with ) TRE version 0.6.8 . Better,
faster, all that. :)
2) A fourth new classifier with very impressive statistics is now
available. This is the OSB-Winnow classifier, originally designed by
Christian Siefkes. It combines the OSB frontend with a balanced
Winnow backend. But it may well be twice as accurate as SBPH Markovian
and four times more accurate than Bayesian. Like correlative matching,
it does NOT produce a direct probability, but it does produce a pR,
and it's integrated into the CLASSIFY statement. You invoke it
with the <winnow> flag:
classify <winnow> (file1.cow | file2.cow) /token_regex/
and
learn <winnow> (file1.cow) /token_regex/
learn <winnow refute> (file2.cow) /token_regex/
Note that you MUST do two learns on a Winnow .cow files- one
"positive" one on the correct class, and a "refute" learn on
the incorrect class (actually, it's more complicated than that
and I'm still working out the details.)
Being experimental, the OSB-Winnow file format is NOT compatible with
Markovian, OSB, nor correlator matching, and there's no functional
checking mechanism to verify you haven't mixed up a .cow file with a
.css file. Cssutil, cssdiff, and cssmerge think they can handle the
new format- but they can't.
Further, you currently have to train it in a two-step
process, learning it into one file, and refuting it in all other
files:
LEARN <winnow> (file1.cow) /regex/
then
LEARN <winnow refute> (file2.cow) /regex/
which will do the right thing. If the OSB-winnow system works as well
as we hope, we may put the work into adding CLASSIFY-like multifile
syntax into the LEARN statement so you don't have to do this two-step
dance.
Version 20040601-BlameKyoto
1) the whitelist.mfp, blacklist.mfp, and priolist.mfp files shipped
are now "empty", the prior lists are now shipped as *list.mfp.example
files. Since people should be very careful setting up their black and
white lists, this is (hopefully!) an improvement and people won't get
stale .mfp's .
2) The CLASSIFY statement, running in Markovian mode, now uses Arne's
speedup, and thus runs about 2x faster. Note that this speedup is
currently incompatible with <microgroom>, and so you should use either
one or the other. Once a file has been <microgroom>ed, you should
continue to use <microgroom>. This is _not_ enforced yet in the
software; if you get it wrong you will get a slightly higher error
rate, but nothing apocalyptic will happen.
3) the CLASSIFY statement now supports Orthogonal Sparse Bigram <osb>
features. These are mostly up- and down-compatible with the standard
Markovian filter, but about 2x faster than Markovian even with Arne's
speedup. Even though there is up- and down-compatibility, you really
should pick one or the other and stick with it, to gain the real
speed improvement and best accuracy.
4) The CLASSIFY <correlate> (that is, full correlative) matcher has
been improved. It now gives less counterintuitive pR results and
doesn't barf if the test string is longer than the archetype texts (it
still isn't _right_, but at least it's not totally _wrong_. :) Using
<correlate> will approach maximal accuracy, but it's _slow_ (call it
1/100th the speed of Markovian). We're still working on the
information theoretic aspects of correlative matching, but it may be
that correlative matching may be even more powerful than Markovian or
OSB matching. However, it's so slow (and completely incompatible with
Markovian and OSB) that a statistically significant test has yet to
be done.
Note: this version (and prior versions) are NOT compatible with TRE
version 0.6.7. The top TRE person has been notified; so use TRE
version 0.6.8 (which is included in the source kit) or drop back to
TRE-0.6.6 as a fallback.
Documentation is (as usual) cleaned up yet further.
Work continues on the full neural recognizer. It's unlikely that
the neural recognizer will ue a compatible file format, so keep
around your training sets!
Version 20040418-BlameEasterBunny
This is the new bleeding edge release. It has several submitted
bugfixes (attachments, windowing), major speedups in data I/O,
and now allows random-access file I/O (detailed syntax can be found
on the QUICKREF text). For example, if you wanted
to read 16 bytes starting at byte 32 of an MP3 file (to grab one of
the ID3 tags), you could say
input [myfile.mp3 32 16] (:some_tag:)
Likewise, you can specify an fseek and count on output as well; to
overwrite the above ID3 tag, use:
output [myfile.mp3 32 16] /My New Tag Text/
As usual for a bleeding-edge release, this code -is- poorly tested
yet. Caution is advised.
There's still a known memory leak if you reassign (via MATCH) a
variable that was isolated; the short-term fix is to MATCH with
another var and then ALTER the isolated copy.
March 27, 2004 - BlameStPatrick
This is the new bleeding edge release. A complete rewrite of
the WINDOW code has been done (byline and eofends are gone, eofretry
and eofaccepts are in), we're integrating with TRE 0.6.6 now,
and a bunch of bugs have been stomped.
For those poor victims who have mailreader pipelines that alter headers,
you can now put "--force" on the BASH line or "force" on the
mailer command line, e.g. you can now say
command mysecretpassword spam force
to force learning when CRM114 thinks it doesn't need to learn.
However, this code -is- poorly tested yet. Caution is advised.
There's still a known memory leak if you reassign (via MATCH) a
variable that was isolated; the short-term fix is to MATCH with
another var and then ALTER the isolated copy.
February 2, 2004 - V1.000
This is the V1.0 release of CRM114 and Mailfilter. The last few known
bugs have been stomped (including a moderately good infinite loop detector
for string rewrites, and a "you-didn't-set-your-password" safety check),
the classifier algorithms have been tuned (default is full Markovian),
and it's been moderately well tested.
Accuracies over 99.95% are documented on real-time mail streams,
and the overall speed is 3 to 4x faster than SpamAssassin.
My thanks to all of you whose contributions of brain-cycles made this
code as good as it is.
20040118 (final tweaks?)
It turns out that CAMRAM needs (as in is a virtual showstopper) the
ability to specify which user directory all of the files are to be
found in. Since #insert _cannot_ do this (it's compile time, not
run time), mailfilter.crm (and classifymail.crm) now have a new
--fileprefix=/somewhere/ option.
To use it, put all of the files (the .css's, the .mfp's etc) that are
on a per-user basis in one directory, then specify
mailfilter.crm --fileprefix=/where/the/files/are/
Note that this is a true prefix- you must put a trailing slash
on to specify a directory by that name. On the other hand, you can
specify a particular prefix on a per-user basis, e.g.:
mailfilter.crm --fileprefix=/var/spool/mail/crm.conf/joe-
so that user "joe" will use mailfilter.crm with these files:
/var/spool/mail/crm.conf/joe-mailfilter.cf
/var/spool/mail/crm.conf/joe-rewrites.mfp
/var/spool/mail/crm.conf/joe-spam.css
/var/spool/mail/crm.conf/joe-nonspam.css
and so on. Note that this does NOT override --spamcss and --nonspamcss
options; rather, the actual .css filenames are the concatenation of
the fileprefix and spamcss (or nonspamcss) names.
Version 20040105 (recheck)
Version 1.00, at last! The only fixes here are to make the Makefile
a little more bulletproof and lets you know how to fix a messed-up
/etc/ld.so.conf, and of course this document has been updated.
Otherwise this version should be the same as
the December 27 2003 (SanityCheck) version, which has no reported
reproducible bugs higher than a P4 (documentation and feature request).
For the last two weeks, I had _one_ outright error and two that I
myself found borderline out of about 5000 messages. That's 2x
better than a human at the same task.
My thanks to all of you whose contributions of brain-cycles made this
code as good as it is.
-Bill Yerazunis
Version 20031227 (SanityCheck)
This is (hopefully) the last test version before V1.0, and bug
fixes are minimal. This is really a sanity check release for V1.0 .
It is now time to triage what needs to be fixed
versus what doesn't, and very few things NEED to be fixed.
Things that changed (or not) are:
1) BUGS ACTUALLY FIXED:
removed the arglist feature from mailfilter.crm; there's a
poorly understood bug in NetBSD versus Linux that breaks things.
allmail.txt flag control wasn't being done correctly. That's
fixed.
a couple of misleading comments in the code are fixed.
2) THINGS THAT ARE NOT CHANGED IN THIS VERSION BUT ARE V1.1 CANDIDATES:
the install location fix is NOT in V1.0. This will move
the location of the actual binary (/usr/bin/crm versus
/usr/local/bin/crm-<version> and then add a symlink
/usr/bin/crm --> /usr/local/bin/crm-<favored version> )
the --mydir feature of mailfilter.crm is not yet implemented
and won't be in V1.0 . Expect it in V1.1
Other than that and a few documentation fixes, this version is identical
to 20031217. It's just the final sanity check before we do V1.0
Version 20031215-RC11
Minor bugs smashed. Math evaluation now works decently (but be nice
to it). Mailfilter accuracy is up past 99.9% (less than 1 error per
thousand, usually when a spammer joins a well-credentialed list and
spams the list, or a seldom-heard-from friend sends a one-line message
with a URL wrapped in HTML). Command line features for CAMRAM added
("--spamcss" and "--nonspamcss"; these will probably become unified to
a --mydir). Lots of documentation updates; if it says something in
the documentation, there's actually a good chance it works as described.
Version 20031111-RC7
More bugs smashed- there are still a few outstanding bugs here and
there, but you aren't likely to find them unless you're really pushing
the limits. Improvements are everywhere; You can now embed the
classical C escape chars in a var-expanded string (e.g. \n for a
newline) as well as hex and octal characters like \xFF and \o132.)
EVAL now can do string length and some RPN arithmetic/comparisons;
approximate regexing is now available by default, and the command line
input is improved.
Version 20031101-RC4 (November 1, 2003)
The only changes this release are some edge-condition bugfixes (thanks
to Paolo and JSkud, among others) and the inclusion of Ville
Laurikari's new TRE 0.6.0-PRE3 regex module. This regex module is
tres-cool because it actually has a useful approximate matcher built
right in, dovetailed into the REGEX syntax for #-of-matches.
Consider the regex /aaa(foo){1,3}zzz/ . This matches "foo", "foofoo", or
"foofoofoo". Cognitively anything in a regex's {} doesn't say what
to match, just how to match it.
The cognitive jump you hve to take here is /foo{bar}/ can have a {bar}
that says _how accurately_ to match foo. For instance:
foo{~}
finds the _closest_ match to "foo" (and it always succeeds).
The full details of approximate matching are in the quickref.
Read and Enjoy.
(for your convenience, we also include the well-proven 0.5.3 TRE library,
so you should install ONE and ONLY one of these. Realize that
0.6.0-PRE3 is still a fairly moderately tested library; install
whichever one meets your need to bleed. :-) )
Oct 23, 2003 ( version 20031023-RC3 )
Yes, we're now at RC3. Changes are that EVAL now works right, lots
of bugfixes, and the latent code for RFC-compliant inoculation is
now in the shipped mailfilter.crm (but turned off in mailfilter.cf)
All big changes are being deferred to V1.1 now; this is bugfix city.
Make it bleed, folks, make it _bleed_.
-Bill Yerazunis
October 15, 2003
It's been a long road, but here it is - RC1, as in Release Candidate 1.
WINDOW and HASH have been made symmetrical, the polynomials have been
optimized, and it's ready. Accuracy is steady at around 3 nines.
Because of all the bugfixes, upgrading to this version (compatible with the
BETA series) is recommended.
-Bill Yerazunis
This is the September 25th 2003 BETA-2
What's new: a few dozen bugs stomped, and new functionality
everywhere. Command line args can now be restricted to acceptable
sets; <microgroom> will keep your .css files nicely trimmed; ISOLATE
will copy preexisting captures, --learnspam and --learnnonspam in
mailfilter.crm will perform exactly the same configured mucking as
filtering would, and then learn; --stats_only will generate ONLY the
'pR' value (this is mostly for CAMRAM users), positional args will be
assigned :_posN: variables, the kit has been split so you don't have
to download 8 megs of .css if you are building your .css locally, and
it's working well enough that this is a full BETA release.
'August 07, 2003 bugfix release.
Changes: lots and lots of bugfixes. Really. The only new code is
experimental code in mailfilter (to add 'append verbosity as
attachment') and getting WINDOW to work on any variable, everything
else is bugstomping or enhanced testing (megatest.sh runs a lot of tests
automatically now).
There's still a bug or dozen out there, so keep sending me bug reports!
(and has anyone else done the cssutil --> cssmerge to build small .css files
for fast running?)
This is the July 23, 2003 alpha release.
This release is a bugfix release for the July 20 secret release.
Fixes include: configuration toggles for allmail.txt and rejected_mail.txt,
execution time profiling works, (-p generates an execution time profile,
-P now limits number of statements in program),
Good news: the new .css file format seems to be working very well;
although we spend a little more time in .css evaluation, the accuracy
increase is well worth it (I've had _one_ error since 07-20, a false
accept to a mailing list that came back as "marginally nonspam" because
the mailing list is usually squeaky clean).
Merging works well; you can now make your .css files as big (or small)
as you dare (within reason; you'll need to throw away features if
you want to compress the heck out of it and you'll use lots of memory or
page like crazy if you make them too big). If experiment shows that
this memory usage is excessive, let me know and I'll see if I can do
a less-space-for-more-time tradeoff.
Profiling indicates that we spend more time in blacklist processing
than in the whole SBPH/BCR evaluator, (which isn't that surprising,
when you get down to it), so maybe trimming the blacklist to people
who spam _you_ would be a good performance improvement.
Anyway, here you go; this is a _recommended_ release. Grab it and
have fun. :)
As usual, prior news and updates are at the end of this file.
---------
This is the July 19, 2003 SECRET alpha release. It won't be linked on
the webpage- the only people who will know about it are the ones who
get this email. Y'all are special, you know that? :-)
Since this is a SECRET release, you all have a "need to know". That
need is simple: I'd like to get a little more intense testing on this
new setup before I put it out for general release.
Enough has changed that you _need_ to read ALL the news before
you go off and install this version. Be AFRAID. :)
LOTS of changes have occurred - the biggest being that the new,
totally incompatible but far better .css format has been implemented.
The new version has everything you all wanted- both for people who
want huge .css files, and for people who want _smaller_ .css files.
This new stuff has necessitated scouring cssutil and cssdiff
so don't use the old versions for the new format files.
Lastly, because the old bucket max was 255 and the new is 4 gigs, the
renormalization math changed a little. Expect pRs to be closer to 0
until you train some more. Accuracy should be better, even _before_
training, so overall it's a net win.
There's also string rewriting in the pre-classification stage (who
wanted that? Somebody did....) and since term rewriting is so darn
useful, I'm releasing an expurgated version of the string rewriter I
use to scrub my spam and nonspam text of words that should not be
learned. This scrubber automatically gets used if you "make cssfiles".
Here's the details:
1) The format of the .css files has changed drastically. What used to
be a collisionful (and error-accepting) hash is now a 64-bit hash
that is (probably) nearly error free, as it's also tagged with the
full 64-bit feature value; if two values clash as to what bucket
they would like to use, proper overflow techniques keep them from
both using the same bucket. Bucket values were maxxed at 255 (they
were bytes) now they're 32-bit longs, so you are _highly_ _unlikely_ to max
out a bucket. These two changes make things significantly more
robust.
These changes also make it possible (in fact, trivial) to
resize (both upward and downward!), compress, optimize, and
do other very useful things to .css files. Right now, the only
supported operation is to _merge_ one .css file onto another... but
the good news is that now these files can be of different sizes!
So, the VERY good news is that you can look at your .css files with
cssutil, decide if (or where) you want to zero out less significant
data, and then use dd to create a blank, new outfile.css file that will
be about half to 2/3 full, then use cssmerge outfile.css infile.css to
merge your infile.css into the outfile.css.
This will be a real help for people who have (or need) very large OR very
small .css files. :)
You can create the blank .css file with the command 'dd' as in:
dd bs=12 count=<number of feature buckets desired> if=/dev/zero of=mynew.css
(the bs=12 is because the new feature buckets are 12 bytes long)
Because chain overflowing is done "in table, in sequence" you can't
have more features than your table has feature buckets. You'll get a
trappable error if you try to exceed it.
Minor nit- right now, feature bucket 0 is reserved for version
info- but it's never used (left as all 0's). That's no major
hassle, but just-so-you-know... :)
2) A major error in error trapping has been corrected. TRAPs can now
nest at least vaguely correctly; a nonfatal trap that is bounced does
not turn into a fatal. Also, the :_fault: variable is gone, each
TRAP now specifies it's own fault code.
This isn't to say that error trapping is now perfect, but it's a
darn sight better than it was before.
3) term rewriting on the matched or learned text is now supported; this
will mean significant gains in out-of-the-box accuracy as well as keeping
your mail gateway name from becoming a spam word. :) Far more fancy
rewritings can be implemented, if you should choose.
The rewriting rules are in rewrites.mfp - YOU must edit this to match
your local and network mailer service configuration, so that your
email address, email name, local email router, and local mail router
IP all get mapped to the same strings as the ones I built the
distribution .css files with.
4) Minor bugs - a minor bug (inaccurate edge on matching) for
the polynomial; annoying segfault on insert files that ended with
'#' that were immeidately followed by a { in the main program was fixed;
5) a new utility is provided - rewriteutil.crm. This utility can do
string rewriting for whatever purpose you need. I personally use it
to "scrub" the spam and nonspam text files; the file
rewrites.mfp contains an (expurgated) set of rewrite
rules that I use. You will need to edit rewrites.mfp
to put your account name and server nodes in, otherwise you'll be using
mine (and losing accuracy)
For examples on the term rewriting, both in the mailfilter and in
the standalone utility rewriteutil.crm, just look at the example/test
code in rewritetest.crm (which uses the rewrite rules in test_rewrites.mfp)
This is the July 1, 2003 alpha release.
This is a further major bugstomping release. The .css files are
expanded to 8 megabytes to decrease the massive hash-clashing that has
occurred. UNION and INTERSECTION now work as described in the
(updated) quickref.txt, with the (:out:) var in parens and the [:in1:
:in2: ...] vars in boxes. A major bug in LEARN and CLASSIFY has been
stomped; however this is a "sorta incompatible" change and you are
encouraged to rebuild your .css files with a hundred Kbytees or so of
prime-grade spam and nonspam (which has been stored for you in
spamtext.txt and nonspamtext.txt). The included spam.css and
nonspam.css files are already rebuilt for the corrected bug in LEARN
and CLASSIFY. These .css files are also completely fresh and new;
I restarted learning about a week ago and they're well into the 99.5%
accuracy range.
This is the June 23, 2003 alpha release.
This is a major bugstomping release. <fromstart> <fromcurrent> and
<backwards> now seem to work more like they are described to work.
The backslash escapes now are cleaner; you may find yuor programs work
"differnently" but it _should_ be backward_compatible. The
preprocessor no longer inserts random carriage returns. A '\' at the
end of a line is a continuation onto the next line. Mailfilter now
can be configured for separate exit codes on "nonspam", "spam" and
"problem with the program". Exit codes on CRM114 itself have been
made more appropriate; compiler errors and untrapped fatal faults now
give an error exit code. Additionally, FAULT and TRAP are scrubbed,
and the documentation made more accurate.
June 10 news:
This new version implements the new FAULT / TRAP semantics,
so user programs can now do their own error catching and
hopefully error fixups. Incomplete statements are now flagged
a (little bit) better.
Texts are now Base64-expanded and decommented before being learned
There's a bunch of other bugfixes as well.
Default window size is dropped to 8 megs, for compatiblity
with HPUX (change this in crm114_config.h).
June 01, 2003 news:
the ALIUS statement - provides if/then/else and switch/case
capabilities to CRM114 programmers. See the example code in
aliustest.crm to get some understaning of the ALIUS statement.
the ISOLATE statement - now takes a /:*:initial: value / for the
freshly isolated variable.
Mailfilter.crm is now MUCH more configurable, including inserting
X-CRM114-Status: headers and passthru modes for Procmail, configurable
verbosity on statistics and expansions, inserting trigger 'ADV:' tags
into the subject line, and other good integration stuff.
Overall speed has improved significantly - mailfilter is now about
four times FASTER than SpamAssassin with no loss of accuracy.
bugfix - we now include Ville Laurikari's TRE regexlib version 0.5.3
with CRM114; using it is still optional ("make experimental")
but it's the recommended system if your inputs include NULL bytes.
bugfix - OUTPUT to non-local files now goes where it claims,
it should no longer be necessary to pad with a bunch of spaces.
yet more additions to the .css files
April 7th version:
0) We're now up to "beta test quality"... no more "alpha"
quality level. This is good. :-)
1) As always, lots of bugfixes. And LOTS of thanks from all of you
poor victims out there. We've reached critical mass to the point now
where I'm even getting bug _fix_ suggestions; this is great!
If you do make a bug report or a bugfix suggestion, please include not
only the version of CRM114 you're running, but also the OS and version
of that OS you're running. I've seen people porting CRM114 to Debian,
to BSD, to Solaris, and even to VMS... sp please let me know what
you're running when you make a bug report. PLEASE PUT AT LEAST THE
CRM114 VERSION IN THE SUBJECT LINE.
2) We now have an even better 'mailfilter.crm' . Even with the highly
evolved spam in the last couple of, we're still solidly above 99%
(averaging around 99.5%). (it's clear that the evolution is due
to the pressures brought by Bayesian filters like CRM114)... some
of these new spams are very, VERY good. But we chomp 'em anyway. :-)
3) The new metaflag "--" in a CRM1114 command line flags the
demarcation between "flags for CRM114" and "flags for the user program
to see as :_argN:". Command line arguments before the "--" are seen
only by CRM114; arguments after the "--" are seen only by the user
program.
4) EXPERIMENTAL DEPARTMENT: We now have better support for the
8-bit-clean, approximate-capable TRE regex engine. It's still
experimental, but we now include TRE 0.5.1 directory in this kit; you
can just go into that subdirectory, do a .configure, a make, and a
make install there, and you'll have the TRE regex engine installed
onto your machine (you need to be root to do this). Then go back up
to the main install directory, and do a "make experimental" to compile
and install the experimental version as /usr/bin/crma (the 'a' is for
'approximate regex support'.
Using the experimental version 'crma' WILL NOT AFFECT the
main-line version 'crm'; both can coexist without any problems.
To use the approximate regex support (only in version 'crma') just add a
second slashed string to the MATCH command. This string should contain
four numbers, in the order SIMD (which every computer hacker should
remember easily enough). The four integers are the:
Substitution cost,
Insertion cost
Maximum cost
Deletion cost
in an approximate regex match. If you don't add the second slash-delimited
string, you get ordinary matching.
Example:
match /foobar/ /1 1 1 1/
means match for the string "foobar" with at most one substitution,
insertion, or deletion.
This syntax will eventually improve- like the makefile says, this
is an experimental option. DO NOT ASSUME that this syntax will not
change TOTALLY in the near future.
DO NOT USE THIS for production code.
4) Yet futher improvements to the debugger.
5) Further improvements to the classifier and the shipped .css files.
6) The "stats" variable in a CLASSIFY statement now gives you an extra
value- the pR number. It's pR for the same reason pH is pH - it gives an
easy way to express very large numeric ratios conveniently.
The pR number is the log base 10 of the .css matchfile signal strength
ratios; it typically ranges from +350 or so to -350 or so. If you're
writing a system that uses CRM114 as a classifier, you should use pR
as your decision criterion ( as used by mailfilter.crm and
classifymail.crm, pR values > 0 indicate nonspam, <0 indicates spam )
If you want to add a third classification, say "SPAM/UNSURE/NONSPAM",
use something like pR > 100 for nonspam, between +100 and -100 for
unsure, and < -100 for spam. CAMRAM users, take note. :)
6) The functionality of 'procmailfilter.crm' has been merged back into
mailfilter.crm, classifymail.crm, learnspam.crm and learnnonspam.crm.
Do NOT use the old "procmailfilter.crm" any more - it's buggy,
booger-filled, and unsupported from now on. PLEASE PLEASE
PLEASE don't use it, and if you have been using it, please stop now!
Jan 28th release news
Many thanks to all of you who sent in fixes, and taught me some nice
programming tricks on the side.
0) INCOMPATIBLE CHANGES:
a) INCOMPATIBLE (but regularizing) change: Input took from the file
[this-file.txt] but output went to (that-file.txt); this was a
wart and is now fixed; INPUT and OUTPUT both now use the form of
INPUT [the-file-in-boxes.txt]
and
OUTPUT [the-file-in-boxes.txt]
b) INCOMPATIBLE (but often-requested) change: You don't need to say
"#insert" any more. Now it's just ' insert ', with no '#' . Too many
people were saying that #insert was bogus, and it was too easy to
get it wrong. Now, insert looks like all other statements;
insert yourfilenamehere.crm
c) The gzip file no longer unpacks into "installdir", but into
a directory named crm114-<versionnumber> .
1) BUGFIXES: bugs stomped all over the place - debugger bugs (now the
debugger doesn't go into lalaland if an error occurs in a batch file),
infinite loop on bogus statements fixed, debugger "n" not doing the
right thing), window statement cleaned and now works better, '\' now
works correctly even in /match patterns/, default buffer length is now 16
megabytes (!), the program source file is now opened readonly.
2) 8-BIT-CLEAN: code cleanups and reorganizations to make CRM114
8-bit-cleaner; There may be bugs in this (may? MAY?) but it's a
start. (note- you won't get much use of this unless you also turn on
the TRE engine, see next item.)
3) REGEX ENGINES: the default regex engine is still GNU REGEX (which
is not 8-bit-clean) but we include the TRE regex engine as well (which
is not only 8-bit-clean, but also does approximate regexes. TRE is
still experimental, you will need to edit crm114_config.h to turn it
on and then rebuild from sources. Do searches of www.freshmeat.net to
see when the next rev of TRE comes out.
4) SUBPROCESSES: Spawned minion buffers now set as a fraction of the
data window size, so programs don't die on overlength buffers if they
copy a full minion output buffer into a non-empty main data window.
The current default size is scaled to the size of the main data buffers,
currently 1/8th of a data buffer, with the new default of a 16-meg
allocate-on-the-fly data buffer that means your subprocesses can
spout up to 2 megs of data before you need to think about using
asynchronous processes.
5) The debugger now talks to your tty even if you've redirected stdin
to come from a data file. EOF on the controlling tty exits the
program, so -d nnnn sets an upper limit on the number of cycles an
unattended batch process will run before it exits. (this added because
I totally hosed my mailserver with an infinite loop. Quite the
"learning experience", but I advise against it. )
6) An improved tokenizer for mail filtering. You can pick any of
7) Option for exit codes for easy ProcMail integration, so the old
"procmailfilter.crm" file goes away, it's no longer necessary to
have that code fork.,
8) For those of you who want eaiser integration with your local mail
delivery software, without all the hassle of configuring
mailfilter.crm, there's three new very bare-bones programs, meant to
be called from Procmail. These do NOT use the blacklist or whitelist
files, nor can they be remotely commanded like the full
mailfilter.crm:
learnspam.crm
learnnonspam.crm
classifymail.crm
* learnspam.crm < some-spam.txt
will learn that spam into your current spam.css database. Old
spam stays there, so this is an "incremental" learn.
* learnnonspam.crm < some-non-spam.txt
will learn that nonspam into your current nonspam.css database. Old
nonspam stays there, so this is an "incremental" learn.
* classifymail.crm < mail-message.txt
will do basic classification of text. This code doesn't
do all the advanced things like base-64 armor-piercing nor
html comment removal that mailfilter.crm does, and so it
isn't as accurate, but it's easier to understand how to set
it up and use it. Classifymail.crm returns a 0 exit code on
nonspam, and a 1 exit code on spam. Simple, eh? Classifymail
does NOT return the full text of the message, you need to
get that another way (or modify classifymail.crm to
output it- just put an "accept" statement right before
the two "output ..." statements and you'll get the full
incoming text, unaltered.
November 26, 2002:
NEW Built-in Debugger - the "-d" flag at the end of the command line
puts you into a line-oriented high-level debugger for CRM114 programs.
Improved Classifier - the new classifier math is giving me > 99.92%
accuracy (N+1 scaling). In other words, once the classifier is
trained to your errors, you should see less than one spam per
thousand sneak through.
Bug fixes - the code base now should compile more cleanly on newer
systems that have IEEE float.h defined.
Security fix- a non-exploitable buffer overflow fixed
Documentation fixes - Serious doc errors were fixed
Nov 8th, 2002 version
*) Procmail users: a version of mailfilter.crm specifically
set up for calling from inside procmail is included-
see the file "procmailfilter.crm" for the filter, and
"procmailrc.recipe" for an example recipe of how to call it.
(courtesy Craig Hagan)
*) Bayesian Chain Rule implemented - scoring is now done
in a much more mathematically well-founded way.
Because of this, you may see some retraining required,
but it shouldn't be a lot. Users that couldn't
use my pre-supplied .css files should delete the supplied
.css files and retrain from their own spamtext.txt and
nonspamtext.txt files.
*) classifier polynomial calculation has been improved but is
compatible with previous .css files.
*) -s will let you change the default size for creating new
.css files (needed only if you have HUGE training sets.)
Rule of thumb: the .css files should be at least 4x the
size of the training set.
*) Multiple .css files will now combine correctly - that is,
if you have categorized your mail into more than "spam" and
"nonspam", it now works correctly. Ex: You might create categories
"beer", "flames", "rants", "kernel", "parties",
and "spam", and all of these categories will plug-and-play
together in a reasonable way,
*) speed and correctness improvements - some previously fatal
errors can now be corrected automagically.
Oct 31, 2002:
Bayesian Chain Rule implemented - scoring is now done in a much more
mathematically well-founded way. Because of this, you may see some
retraining required, but it shouldn't be a lot. Users that couldn't
use my pre-supplied .css files should delete the supplied .css files
and retrain from their own spamtext.txt and . nonspamtext.txt files.
Classifier polynomial calculation has been improved but is compatible
with previous .css files. -s will let you change the default size for
creating new .css files (needed only if you have HUGE training sets.)
Rule of thumb: the .css files should be at least 4x the size of the
training set. Multiple .css files will now combine correctly - that
is, if you have categorized your mail into more than "spam" and
"nonspam", it now works correctly. Ex: You might create categories
"beer", "flames", "rants", "kernel", "parties", and "spam", and all of
these categories will plug-and-play together in a reasonable way, e.g.
classify (flames.css rants.css spam.css | beer.css parties.css kernel.css)
will split out flames, rants, and spam from beer, parties, and
linux-kernel email. (I don't supply .css files for anything but spam
and nonspam, though.) Lastly, there are some new speed and correctness
improvements - some previously fatal errors can now be corrected
automagically.
Oct 21: Improvements everywhere - a new symmetric declensional parser,
a much more powerful and accurate sparse binary polynomial hash
system ( sadly, incompatible; - if you LEARNed new data into the
.css files, you must use learntest.crm to LEARN the new data into
the new .css files as the old file used a less effective polynomial.)
Also, many bugfixes including buffer overflows fixed, -u to change
user, -e to ignore environment variables, optional [:domain:]
restrictions allowed on LEARN and CLASSIFY, status output on
CLASSIFY, and exit return codes. Grotty code has
been removed, the Remote LEARN invocation now cleaned up, and CSSUTIL
has been scrubbed up.
Oct 5: Craig Rowland points out a possible buffer exploit- it's
been fixed. In the process, the -w flag now boosts all intermediate
calculation text buffers as well, so you can do some big big things
without blowiing the gaskets. :)