Dan Barker wrote: > I've heard to try for equal numbers of spam training and ham training. > Optimally yes. Realistically, wild deviations from this ratio don't hurt very much. > I've used the defaults for autolearn, and manually relearned all the false > positives. It seems that learning the false negatives would be a good thing > too, but dump magic is already way over 10:1 spam. > > Do I need to do something different? > No.. I'd go ahead and train your FN's. Mine's wildly off as well. I think I was at 20:1 last time I checked.
Do what you can to keep your ham training up, and if you can find ways to train more, do it, but don't kill yourself over it. > Dan > > 3 0 non-token data: bayes db version > 1402564 0 non-token data: nspam > 119267 0 non-token data: nham > 151248 0 non-token data: ntokens > 1179379647 0 non-token data: oldest atime > 1179466091 0 non-token data: newest atime > 0 0 non-token data: last journal sync atime > 1179466101 0 non-token data: last expiry atime > 86400 0 non-token data: last expire atime delta > 44795 0 non-token data: last expire reduction count > > >
