C and D
So, what's the deal with C and D?
I've given both a try. Both seem to follow instructions better than A and B, as in follow patterns (I'm using AI a lot nowadays for moist text adventures, where I get some narration and possible actions, I prefer that to pure conversations).
If I had to say, D seems a bit more logical?
In general, I think v5 seems to behave better than v4. Going back to my text adventure format, I have a template & character card that tells the AI to format it like this
Several (moist) status stuff
Narration + dialog
Choices for player:
A)
B)
C)
D)
With same parameters and template (Alpaca), v4 skips this, and it's too verbose. It never seems to end its turn and give me the choices, despite having a template in the context for it to follow. It just keeps writing and writing and writing. V5 is a lot more well behaved and follows the patterns a lot better. v5c and v5d so far seem somewhat smarter or more logical. One thing I appreciate from any AI is the updating of the status. Let us say that status says:
- Your clothes: jeans and t-shirt
And in the next choice, I take them off. If after that the new status shows this, it's a pretty good model. Just in case that helps with testing, but I think a model that keeps always prefixing a status and it can update the status logically as a result of the previous exchange is quite logical. So far v5d has done this coherently.
I was also surprised with how it handles rewriting text. I gave it something from a fanfic I'm writing, and the rewrite was decent, and especially it kept the meaning of it. v5d in particular seems to incorporate previous context & lorebooks better, but since I haven't done that much testing, can't say for sure but loosk like my favorite so far.
Thank you @Culturedniichan ! Could you try v5d-e4? It's moistier and I think the intelligence is still there. https://huggingface.co/BeaverAI/Moistral-11B-v5d-e4-GGUF
Hmm e4 didn't perform as well as v5d