Model overthinks
#1
by elevendr - opened
You appear to be using a specific Heretic model variant of sorts for this test, unless I'm mistaken? The base StyleTune version, when I benchmarked it, showed no signs of overthinking.
You appear to be using a specific Heretic model variant of sorts for this test, unless I'm mistaken? The base StyleTune version, when I benchmarked it, showed no signs of overthinking.
It seems it also happening on the normal styletune as well, it tends to overthink on some prompts and think fine for other prompts. I wonder what values you recommend for repetition penalty or perseverance penalty since LM Studio dose not have any DRY samplers.


