Token leak

#1
by Varkoyote - opened

Hello! Was about to test this model, but upon generating a first message, I'm instantly met with it ending with "user" πŸ˜…

I regenerated, now it's either "assistant" or my "name:", like it has no stop string.

Thanks for letting me know! I made the error of specifiying Mag-Mell as the tokenizer source, instead of creating a union. I had trouble getting the tokenizer working so I messed with the merge config a lot.
I'll remerge this model and redrix/matricide-12B-Unslop-Unleashed to use a union tokenizer and then reupload them.
I apologize for the mistake, I'm quite busy right now and just doing this on the side. I also recently switched my computer's operating system (windows -> fedora linux) and am working on a lot of assignments for school, so something was bound to slip through the cracks.
If you use ST, you can specify custom stopping strings to band-aid fix this. I've no access to my PC at the moment so I can share those later. Something like ["\n{{user}}:", "\nAssistant:"] might help.

redrix changed discussion status to closed

Sign up or log in to comment