DavidAU

/

Fimbulvetr-Grande-V2-19B-GGUF

like 3

[using IQ4_XS, 1.12 rep.pen, 1.25 temp]

I can already recognize the flavorful and creative writing from the 11b model. There's no doubt this is a fimbulvetr-based model!

Problem 1
It really likes long responses! Just like the 11b version, it likes generating extremely long responses with no end in sight. Despite prompts telling it to limit its length in various ways, it just keeps generating and generating and generating. The only thing that seemed to work for 11b was to stop and shorten its responses manually until it catches on.

I see this is mentioned in the description, so maybe it's intentional? (i.e. it's by design?)

I found that simply using "Alpaca Roleplay" context template in SillyTavern, no instruct mode, no system prompt, seems to produce the shortest answers, most of the time. However, its writing style becomes more erratic/worse.

Problem 2
Also, like in 11b, it keeps making decisions and performing actions for my character. Despite heavily prompting against this behaviour, it still writes what I'm doing, what I'm saying and what choices I make, giving me little chance to write them myself. In 11b it was manageable with various prompt strategies. The struggle persists in this model.

Actually a good thing if you don't want to think, and just want to go with the flow.

These problems make it hard to test longer chats, because with my expectations/preferences I keep having to restart or regenerate the response while tweaking settings, which gets tedious. Maybe I have the wrong settings, or this isn't the right model for me. But following the temperature and rep.pen suggestions do seem to improve the general writing.

/2cents