Re: Tips on training bayes?

W dniu 18.09.2024 o 16:30, Reindl Harald (privat) pisze:
>
>
> Am 18.09.24 um 16:19 schrieb natan:
>> Hi
>> I was very disappointed with spamassassin 4.x because it started to 
>> grow /var/lib/amavis/tmp/
>>
>> With SA 3.4.X - on average 100MB and it deletes on the fly
>> With SA 4.X - on average 2-6GB and I had to do a quick fix:
>> 59 23 * * * root find /var/lib/amavis/tmp/ -mtime +0 -delete;
>
> don't blame SA when a blind man can see that your problem is on the 
> Amavis side - why do one need Amavis tu begin with when there is SA 
> and spamass-milter
yes yes .... everyone knows better why I use amavis ? Because each of my 
users has their own whitelist and blacklist and score rules....

Amavis bad... uga buga ..... but this bad amavis works fine with every 
SA version except 4.X
Of course, if I didn't need it, I wouldn't use it.

I'll tell you more I have amavis and rspamd (NTG) for testing and this 
bad amavis also works correctly ........
>
>> W dniu 18.09.2024 o 16:09, Matus UHLAR - fantomas pisze:
>>> On 18.09.24 13:42, Grega via users wrote:
>>>> Right now in SA 4.0.1 bayes at least for me is really challenging 
>>>> to train and set up.
>>>>
>>>> I had good trained DB from past V3 install, and it behaved really odd.
>>>>
>>>> I trained it on new set of mails 3000 spam and 3000 ham (HAND 
>>>> PICKED mail it was PAIN) and I cant get either BAYES_00 or BAYES_99 :)
>>>>
>>>> I mean I get them occasionally, but not even close to what it was 
>>>> in V3.
>>>>
>>>>
>>>> In V3 SA bayes was decisive, when well trained it was awesome.
>>>>
>>>> Nov in V4.0.1 bayes is NON decisive, and in 90% of cases it gives 
>>>> me BAYES_40 or _50 even after I mark those mails as SPAM OR HAM.
>>>>
>>>>
>>>> What is even more weird is, that some mails aren`t even bayes 
>>>> scored at all. BAYES_XX is missing from headers entirely and I
>>>>
>>>> don`t know why...
>>>>
>>>>
>>>> I`m kind of sorry that I upgraded to 4.0.1...
>>>
>>>
>>> looking at your first mail, it seems that you only have tokens for a 
>>> few days:
>>>
>>> dbg: bayes: corpus size: nspam = 1190, nham = 12441 dbg: bayes: DB 
>>> expiry: tokens in \
>>> DB: 979401, Expiry max size: 1500000, Oldest atime: 1725361640, 
>>> Newest atime: \
>>> 1725888528, Last expire: 0, Current time: 1725888537
>>>
>>> % date -d @1725361640
>>> Tue Sep  3 13:07:20 CEST 2024
>>>
>>> % date -d @1725888528
>>> Mon Sep  9 15:28:48 CEST 2024
>>>
>>>
>>> How do you call spamassassin, directly, via spamass-milter, amavis 
>>> or other way?
>>> Did you tune any bayes settings?
>>> Do you have your trusted_networks and internal_networks set up 
>>> properly?

--