Image-Text-to-Text
Transformers
Safetensors
llava
image-to-text
conversational

Endless loop

#8
by alcalde - opened

I asked a question of your model (comparing two press releases) and it went into a ten minute endless thinking loop. This was on the official Huggingface demo page, so it's not anything I configured wrong. With the few other questions I asked it also seemed to output more thoughts than content for the final answer. When I asked a question about horse race wagering, it spent several paragraphs debating with itself whether this violated some rule about not providing gambling instructions or not. It repeated itself a few times here too but eventually managed to decide to produce output.

You never noticed your model doing this before you released it? It can't just be me, could it? Maybe the bigger the question, the more likely the endless loop? I would have named this model Azathoth instead of Apriel, because in Lovecraft's stories the god Azathoth lies dreaming at the center of chaos, kept asleep by the endless playing of maddening flutes. In a similar way, Apriel slumbers forever without answering, babbling madly and endlessly to itself. :-)

I have also encountered endless generation (though, using koboldcpp/llamacpp + sillytavern frontend), however, it seems to be tied to the sampler settings? Perhaps there's a problematic range of parameters that make it fall into a loop. Regrettably, I haven't recorded what was causing the problem, as I've been switching various sampler presets, but it did eventually went away. Perhaps the demo page chat is misconfigured too?

Sign up or log in to comment