Understanding tuning results
Boris 'pi' Piwinger
3.14 at logic.univie.ac.at
Fri Jun 6 15:12:55 CEST 2003
David Relson wrote:
>> > 06/05 14:59:51 1 0.475 fpos...0 at cutoff 0.503000, run0... 62 run1... 66 run2... 78 206
>> > 06/05 15:13:54 3.2e-1 0.475 fpos...0 at cutoff 0.503000, run0... 62 run1... 66 run2... 78 206
>> > 06/05 15:28:38 1e-1 0.475 fpos...0 at cutoff 0.503000, run0... 62 run1... 66 run2... 78 206
>> > 06/05 15:43:10 3.2e-2 0.475 fpos...0 at cutoff 0.503000, run0... 62 run1... 66 run2... 78 206
>> > 06/05 15:55:58 1e-2 0.475 fpos...0 at cutoff 0.503000, run0... 62 run1... 66 run2... 78 206
>>
>>That suggests that robs is without any meaning. I doubt it.
>>There is something really broken.
>
> I beg to differ. The results say that the robs value isn't a significant
> contributor to the result. Based on other tests, I'd suggest using 1e-2
> for the value.
So with the new version .13.6.2 I did a new run:
> 06/06 14:00:17 1 0.450 fpos.. 0 at cutoff 0.978100, run0.. 160 run1.. 160 run2.. 143 463
> 06/06 14:13:06 3.2e-1 0.450 fpos.. 0 at cutoff 0.978100, run0.. 160 run1.. 160 run2.. 143 463
> 06/06 14:26:57 1e-1 0.450 fpos.. 0 at cutoff 0.978100, run0.. 160 run1.. 160 run2.. 143 463
> 06/06 14:40:36 3.2e-2 0.450 fpos.. 0 at cutoff 0.978100, run0.. 160 run1.. 160 run2.. 143 463
> 06/06 14:55:05 1e-2 0.450 fpos.. 0 at cutoff 0.978100, run0.. 160 run1.. 160 run2.. 143 463
That is strange. Again no significance of robs. dev_min
slightly lower, *but* spam_cutoff much greater and so ist
the number of false negatives.
The number of mails to work with only changed
insignificantly from yesterday, i.e. the results are not
really stable. Why?
pi
More information about the Bogofilter
mailing list