Whether its a new model(looks so) or just general improvements, its much better now! It follows instructions more closely and is capable of writing some really good stuff.

One old issue that is left from previous update is repetative generation. Sometimes model gets stuck in a loop where its incapable of writing anything but one certain sentence, e.g. it often makes character say exactly one thing or send you into exactly one place, like cafe or canteen. You can regen 30 times and there wont be any significant differences between answers.

I heard its regulater by "temperature" stat on dev's side or something like that? It shouldnt be hardcoded inside the model, right? Imo fixing that, if possible, would solve a lot of issues.

Thanks for your work!

you are viewing a single comment's thread
view the rest of the comments
[–] 1 point 4 months ago* (last edited 4 months ago)

Thanks very much, I'll give it a try and see if it helps. Edit to say thanks but it hasn't worked for me. All it's done is expose it's deep deep flaws which I now wonder if are being hidden by relying on empty blunt verbs instead of being vividly descriptive about what characters are doing. Spatial awareness has always been a problem like it is for all LLM's but it's utterly unusuable in the way it is now with it. If it can tell me a character has 'clung' then it doesn't expose how forcing it to describe how they're clinging makes it fall into hands clinging the waist of someone standing, while their head is also on the floor and also somehow their nose nuzzles hair. Completely nonsensical at something so basic which was not the case before this change or whatever is going on. A few weeks ago it could describe what a character was touching and the position they were in, it wasn't perfect of course, it made some mistakes with it but not to the extreme level it is now.

  • source
  • parent