How long to train? - michaeltsai

It generally takes about 500-1000 messages in the corpus to reach 99+% accuracy, but it should get to 90-95% with only about 200 messages. Since you're only getting 50% accuracy, I'm guessing that something might not be set up properly. Please contact spamsieve@c-command.com if you're still having problems, and I'll help you check the setup.

I used to include some seed spam in the corpus, but it turned out that this decreased the accuracy, so I removed it. Plus, with some changes to make it learn more quickly, most everyone should have enough spam to start out with decent accuracy.

Reply to This

Wednesday, November 01 2006 @ 07:29 PM PST

How long to train? - Felix01

Thanks for the reply, Michael. I went through your help files (moved a spam message back to the inbox and then clicked on "Apply Rules") and determined that it was set up properly.

But I think I figured out the problem. I initially only had about a dozen spam messages for training since I delete them several times/day. But I had several hundred "good" messages which I had used for training.

Apparently, all those "good" messages overwhelmed the corpus and it kept sending spam to my inbox.

So I "Reset Corpus" and started over...this time with the correct percentage of "good" messages and spam specified in your help files.

And today I haven't had a single spam slip in to my inbox.

Looks like this could be a winner and I'll soon be a paying customer.

As 'they' say, RTFM...Read The F***ing Manual!!

Reply to This

Thursday, November 02 2006 @ 06:12 AM PST

How long to train? - Tortoise

Try going to the developers site and post your experience on his forum, Michael monitors the site and replies promptly with help for people experiencing trouble with the application. He will assist you in sorting out why it is not working as designed. I have used it for several years now and it was losing accuracy gradually, I posted on his forum and he replied the same day and assisted me over two days getting things sorted out and it is once again nearly 100% accurate and working like a charm.

Reply to This

Wednesday, November 29 2006 @ 03:49 PM PST