Enable Word Training
Enabling word training allows Mailshell Anti-Spam
to analyze the presence of words in your spam and legitimate
messages. This helps Mailshell Anti-Spam learn to better
identify future messages as spam. Enabling word training
does require more system memory.
Maximum Number of Word
Word training involves tracking the frequency
of words in spam and legitimate messages. Thus, increasing
the number of words may increase accuracy but also requires
For word analysis to be effective in identifying
spam or legitimate messages, Mailshell Anti-Spam training
on sets of spam and legitimate messages is required. Messages
are trained automatically based on the messages from your
approved and block lists and based on the auto-training
threshold option. You can modify the value of minimum
training to control when the word analysis data will be
If the number is too low then the accuracy
could be poor due to insufficient data. If the number
is too high, then the training data will not be fully
taken advantage of.
Training weight determines how much Mailshell
Anti-Spam relies on word training to score a message.
Selecting "low" will set a lower weight for
local training data compared to global Mailshell-provided
weight, while selecting "high" will set a heavier
weight for local training data compared to global Mailshell-provided
A "high" setting may adversely
affect the score if there is insufficient or improper
training. The recommended default setting is "low".
Mailshell Anti-Spam has built in rules that
are used to identify spam messages. Enabling them improves
accuracy but requires more time.
Enable Extended Rules
Mailshell Anti-Spam can download and use
additional rules to help identify spam messages. This
can help catch more spam, but startup time will increase
and about 20 MB of memory is required.
Spam Scoring Timeout
You can configure Mailshell Anti-Spam to stop
spam scoring if it takes too long. Although message scoring
occurs in milliseconds, under extreme server load conditions
scoring may take longer. If you wish to place a maximum
bound on the time used to score your messages, specify it
here. Timeouts happen on a message-by-message basis. Those
messages that timeout will have a X-Spamcatcher-Score header
based solely on the spam rules.
RBL Expert Options
Enable RBL MultiHit
If you have chosen to use multiple public
blacklists, then each message is checked against multiple
blacklist servers. If you want your messages to be scored
based on the results of all queries, then you can enable
this option. If you are confident in identifying a spam
message based on results from just one of the blacklist
servers, then you should disable this option.
Disabling the option may improve performance
because you will not have to wait for all blacklists servers
This allows you to set the maximum timeout
in seconds before finishing all queries against public
blacklists. The default timeout of 5 seconds is normally
sufficient to allow for instances where the blacklist
server may be slow to respond.
If you wish to always wait for a response,
then enter "0".
Maximum Number of IPs to
Query Against RBLs
Each email message can pass through multiple
mail servers with different IP addresses. If you have
enabled checks against public blacklists, each of the
IP addresses would be checked. You can specify a maximum
number to be checked to control the amount of time taken
for queries. This can affect performance.
Since checks against blacklists can introduce
latency and a decrease in performance, this option allows
you to do checks conditionally based on the score prior
to performing the RBL checks.
Auto Training Threshold
This allows you to set a threshold for auto-training.
If a message is scored at or above the high threshold,
it is considered as spam and is then used to train all
the enabled Bayesian modules (rules and/or word) except
for sender and fingerprint of the message. If a message
is scored at or below the low threshold, it is considered
as legit and is then used to train all the enabled Bayesian
modules (rules and/or word) except for sender or fingerprint
of the message.
This allows running netchecks conditionally
based on the score. Network is only queried if score is
at or between the 'low' and 'high' range specified via
this option. Networks can introduce latency and decrease
performance, hence an option for conditional checks.