Fighting ham
Our bayes was apparently giving negative scores incorrectly and I re-built it since it was not effective and letting through a lot of spam. I didn't realize, but it seems those negative scores were keeping SA from applying other tests? Since fixing bayes, we are blocking so much ham it is not funny. These are the rules that I have basically had to disable them below. We run Rules Du Jour, but only zero level rules, those are the only updates besides bayes, plus KAM.cf and Botnet.cf. Given Botnet.cf blocks quite a few, but I understand why. I don't know if any of these rules are part of RDJ, but why so much ham is being hit with only these rules. Does SA with updates and these rules hit so much ham for others? We are constantly getting complaints of our over aggressive spam filters. score PART_CID_STOCK 0 score PART_CID_STOCK_LESS 0 score TVD_FW_GRAPHIC_ID1 0 score TVD_FW_GRAPHIC_ID3 0 score TVD_FW_GRAPHIC_ID3_2 0 score MY_CID_AND_STYLE 0 -- Robert
Re: Fighting ham
Robert: It sounds like your problem rests with your bayes database. Some SA rules will fire on almost all mail, but a properly trained bayes filter should be able to reduce your scores to under your spam threshold. None of these scores rate out very aggressively so I am surprised that these are pushing you over your spam threshold. How have you trained bayes with you spam and ham mail? Also I think that the default SA setting of 200 spam and 200 ham is a little low and do not regard bayes as truly effective until about 1000 message of each kind are learned. That being said I would, and have, reduced the default score for Botnet from 5.0 to 3.0. Also, if your run the 00_ version of Fred's rules note that many of them are very aggressively scored. I personally do not let any rule score at over 3.0, except some network test, to allow bayes to recover the mail from listing as a FP. Best Robert Fitzpatrick wrote: > Our bayes was apparently giving negative scores incorrectly and I > re-built it since it was not effective and letting through a lot of > spam. I didn't realize, but it seems those negative scores were keeping > SA from applying other tests? Since fixing bayes, we are blocking so > much ham it is not funny. These are the rules that I have basically had > to disable them below. We run Rules Du Jour, but only zero level rules, > those are the only updates besides bayes, plus KAM.cf and Botnet.cf. > Given Botnet.cf blocks quite a few, but I understand why. I don't know > if any of these rules are part of RDJ, but why so much ham is being hit > with only these rules. Does SA with updates and these rules hit so much > ham for others? We are constantly getting complaints of our over > aggressive spam filters. > > score PART_CID_STOCK 0 > score PART_CID_STOCK_LESS 0 > score TVD_FW_GRAPHIC_ID1 0 > score TVD_FW_GRAPHIC_ID3 0 > score TVD_FW_GRAPHIC_ID3_2 0 > score MY_CID_AND_STYLE 0 > >
Re: Fighting ham
On Wed, 2007-04-18 at 10:23 -0500, Craig Carriere wrote: > Robert: > > It sounds like your problem rests with your bayes database. Some SA > rules will fire on almost all mail, but a properly trained bayes filter > should be able to reduce your scores to under your spam threshold. None > of these scores rate out very aggressively so I am surprised that these > are pushing you over your spam threshold. How have you trained bayes > with you spam and ham mail? Also I think that the default SA setting of > 200 spam and 200 ham is a little low and do not regard bayes as truly > effective until about 1000 message of each kind are learned. That being > said I would, and have, reduced the default score for Botnet from 5.0 to > 3.0. Also, if your run the 00_ version of Fred's rules note that many > of them are very aggressively scored. I personally do not let any rule > score at over 3.0, except some network test, to allow bayes to recover > the mail from listing as a FP. > Thanks, we are rebuilding bayes and now have in SQL with auto learn on, is that good? Now has over 25K spam, but just 180 ham. I have plenty of ham on my own, is it going to effect it all coming from just a few different addresses if I learn all my own ham? -- Robert
Re: Fighting ham
On Apr 18, 2007, at 4:26 PM, Robert Fitzpatrick wrote: Thanks, we are rebuilding bayes and now have in SQL with auto learn on, is that good? Now has over 25K spam, but just 180 ham. You *really* want to train with more ham than spam. -faisal
Re: Fighting ham
On Wed, 18 Apr 2007, Faisal N Jawdat wrote: On Apr 18, 2007, at 4:26 PM, Robert Fitzpatrick wrote: Thanks, we are rebuilding bayes and now have in SQL with auto learn on, is that good? Now has over 25K spam, but just 180 ham. You *really* want to train with more ham than spam. I have a hard time believing auto learn could be so off-balance. I had auto learn turned on here once and the two were usually within 200-300 messages. Before I turned auto learn off, the bayes_token table had over 85 million records in just over three weeks. We ended up letting our customers choose whether they wanted to use auto learn or not through using the sasql plugin for SquirrelMail.
Re: Fighting ham
Does this really mean that auto-learn is "out of balance"? My first guess is that this site probably relies only on SA to combat spam and does little at the MTA level to reject UBE mail. They may even run a catch-all account which would markedly increase his spam count if he is not rejecting for non-existent users. At my small mail server even with MTA restrictions, conservative ones, in place our spam hits out number ham by probably 4-5 to 1. It is just the nature of the beast. I do agree that he needs to manually train his bayes bases and probably keep feeding ham into the bayes engine. after it starts to fire. As an aside do you use any MTA restrictions and/or greylisting? Best Duane Hill wrote: > On Wed, 18 Apr 2007, Faisal N Jawdat wrote: > >> On Apr 18, 2007, at 4:26 PM, Robert Fitzpatrick wrote: >>> Thanks, we are rebuilding bayes and now have in SQL with auto learn >>> on, is that good? Now has over 25K spam, but just 180 ham. >> >> You *really* want to train with more ham than spam. > > I have a hard time believing auto learn could be so off-balance. I had > auto learn turned on here once and the two were usually within 200-300 > messages. Before I turned auto learn off, the bayes_token table had > over 85 million records in just over three weeks. We ended up letting > our customers choose whether they wanted to use auto learn or not > through using the sasql plugin for SquirrelMail. >
Re: Fighting ham
On Thu, 19 Apr 2007, Craig Carriere wrote: Does this really mean that auto-learn is "out of balance"? My first guess is that this site probably relies only on SA to combat spam and does little at the MTA level to reject UBE mail. They may even run a catch-all account which would markedly increase his spam count if he is not rejecting for non-existent users. At my small mail server even with MTA restrictions, conservative ones, in place our spam hits out number ham by probably 4-5 to 1. It is just the nature of the beast. I do agree that he needs to manually train his bayes bases and probably keep feeding ham into the bayes engine. after it starts to fire. As an aside do you use any MTA restrictions and/or greylisting? I'm using Postfix+ClamAV+SA on our two border filter servers. Roughly 95% of all inbound is messages are weeded out before getting to our internal server our customers use. I have a couple internal blacklists used and greylisting. And, I had set the following values within the local.cf: bayes_auto_learn_threshold_nonspam 0.01 bayes_auto_learn_threshold_spam 18.0 Their normal defaults are 0.1 and 12.0 respectively. I had set a higher value for auto learn as you don't have hardly any control over what messages get learned in either direction. Some others on this list have the auto learn values set even higher. As far as the numbers mentioned by the OP, 25,000 spam to 180 ham? That is a lot more than your ~5 to 1. I would not have suspected auto learn to be that far off.