dijital, sorry if this sounds too obvious, but marking the legit results as "not junk" is supposed to help such situations you describe, where the filter (baynesian, IIRC), can learn what you think is junk or not.

If you have the technical ability to provide a test case for the algorithm not reponding appropriately (or better yet, a patch), the devs are fairly appreciative. Posting on LimeWire Forums will probably attract their attention.

If you can read java code, check out https://www.limewire.org/jira/browse/LWC

cheers

btw, you may want to check out Credence. LW worked with Credence for a bit Credence: Thwarting P2P Pollution but I can't recall the reason for not implementing it.