Can't stop this from rambling and repeating
#5
by
FlareRebellion
- opened
Both of these problems are typical for Mixtral models, but this merge seems super prone to them. It's very easy to get into a situation where it can't be stopped from rambling or repeating at nauseam, no matter the sampler settings. BagelMisteryTour, BondBurger, Fish & co. seem much more usable in that regard.
Maybe there's something that I'm missing with this model, I like the prose and 'character' of it, but it seems to lose its shit at no more than 4k ctx.
merges tend to-do that, its pretty unpredictable - again that's on dampf - i just provided the compute