Text Completion

#3
by AnotherOddity - opened

I know it's not recommended but I was impressed by the quality of the prose and character progression using ChatML in Text Completion, though I did have to manually trim "< im_start>assistant" and "< im_end>" messaging from almost every (but not all) replies, even with NoAss installed. I ran through about 24k context of a single character card, and 10k context of a five character group chat, both in Sillytavern using some Sphiratrioth presets (Temp 1.3, MinP 0.02), and in both cases it held up really well. I even tried it out in Chat Completion proper and ran into more issues there, but I'm pretty sure that was due to both {{char}} and myself using first person perspective and muddling things up.

If you ever released a model tuned for Text Completion more than Chat with less assistant messaging I would gladly give that a go, but regardless, nice work and I look forward to your future models.

Sign up or log in to comment