spamassassin horribly low scores? - SpamAssassin

This is a discussion on spamassassin horribly low scores? - SpamAssassin ; I put SA on my server and have had it running for a while now (couple months). I have been training it with ham and spam this whold time time and am probably up to a couple hundred messages of ...

+ Reply to Thread
Results 1 to 2 of 2

Thread: spamassassin horribly low scores?

  1. spamassassin horribly low scores?


    I put SA on my server and have had it running for a while now (couple
    months). I have been training it with ham and spam this whold time time and
    am probably up to a couple hundred messages of ham and a couple thousand
    messages of spam.

    What I am seeing is a TON of email that is obvious spam (to me) get scored
    and fail several checks, but the scores are so insanely low that it still
    gets through. One message in particular might fail 4 or 5 spam checks, but
    each only adds .1 or .2 to the score for a total of .8 or something.

    Each of these checks are obvious spam to me, like enhancement and drugs and
    the like. I've been adjusting the scores to straight up 10 for the checks
    as I see them, but so far I'm up to 20 or so checks that I've modified and I
    just see this to be a never ending battle.

    What would be GREAT is a global switch for things like AM_DOCTOR, and
    MEDS_OK. By setting those two things to no, then if the system would bump
    up every single check that relates to medicine or medical things to like a
    4.0 score then that would solve 99% of my issues.

    Why do these checks carry such low scores? I mean I understand being
    cautious, but for an erectile fail to score .2??? On what planet does that
    make sense? The system would have to fail on 20 levels as well as having a
    very low total threshold to cause issues with that low of a score.


    --
    View this message in context: http://www.nabble.com/spamassassin-h...p17830923.html
    Sent from the SpamAssassin - Users mailing list archive at Nabble.com.


  2. Re: spamassassin horribly low scores?

    archaic0 wrote:
    > I put SA on my server and have had it running for a while now (couple
    > months). I have been training it with ham and spam this whold time time and
    > am probably up to a couple hundred messages of ham and a couple thousand
    > messages of spam.
    >
    > What I am seeing is a TON of email that is obvious spam (to me) get scored
    > and fail several checks, but the scores are so insanely low that it still
    > gets through. One message in particular might fail 4 or 5 spam checks, but
    > each only adds .1 or .2 to the score for a total of .8 or something.
    >
    > Each of these checks are obvious spam to me, like enhancement and drugs and
    > the like. I've been adjusting the scores to straight up 10 for the checks
    > as I see them, but so far I'm up to 20 or so checks that I've modified and I
    > just see this to be a never ending battle.
    >
    > What would be GREAT is a global switch for things like AM_DOCTOR, and
    > MEDS_OK. By setting those two things to no, then if the system would bump
    > up every single check that relates to medicine or medical things to like a
    > 4.0 score then that would solve 99% of my issues.
    >
    > Why do these checks carry such low scores? I mean I understand being
    > cautious, but for an erectile fail to score .2??? On what planet does that
    > make sense?

    Erm, the human one?

    Actually, that is a real, valid answer here, if you'll allow me to
    explain a moment.

    The first thing to realize about spamassassin is that the rules aren't
    scored individually. They aren't. You can't look at one rule, and
    determine a good score for it, alone, by itself, and expect it to work
    well with hundreds of other rules that were each scored individually.
    You need to consider how the rules interact with each other.

    I don't have the exact data in front of me. But usually when you see a
    "really good" spam rule with a low score, it's low because in the
    mass-check it nearly always fired coincidentally with another rule, but
    that rule fired off on less of the nonspam email. So, SA picked the
    better of the two to throw its weight behind.

    In the case of DRUGS_ERECTILE, it's got a noticable non-zero false
    positive rate, actualy 0.7% of email it hit was nonspam. This happens
    because some people have personal email accounts, which may contain
    jokes, even a short ribbing from a friend about you needing it, or
    medical discussions which may mention any of these drugs in a non-spam
    context. And in the SpamAssassin world, 1 false positive is as bad as
    100 false negatives. Your threshold of pain may be different, but that's
    how the ruleset is tuned.

    Also consider SpamAssassin has to be designed with a broad userbase in
    mind, from the guy swapping off-color jokes with his friends, to a rigid
    business environment. It's not perfect for every situation, but does
    surprisingly well.

    Regardless it would be interesting to see some samples of some
    troublesome spam that's not being hit. We might be able to offer some
    suggestions for how to handle them that is less risky than jacking
    scores up.


    > The system would have to fail on 20 levels as well as having a
    > very low total threshold to cause issues with that low of a score.



+ Reply to Thread