SillyTavern or similar - storyteller

HoodedSilence

That which passes is forgotten.
Game Developer
Jan 18, 2023
238
144
A bit of an odd question but does anyone here use an offline chat client like with SillyTavern with Kobold (or similar) to help you write characters conversations or for storyline brainstorming?

If so which models have you found the most useful?
 

osanaiko

Engaged Member
Modder
Jul 4, 2017
2,420
4,273
Yes, I do. I use Oobabooga as a backend, and then two extensions by "Fartypants" as writing environments: "twinbook" and "playground".

For models, I have a 3090 w/ 24GB VRAM which means I can use the largest home-hobbyist models.

"LoneStrikerNoromaid-v0.1-mixtral-8x7b-Instruct-v3-3.5bpw-h6-exl2" just fits into 22GB free VRAM.
It's very strong but as with many of these models tends to go into "happily ever after" mode toward the end of every generation. By that, I mean it will complete a paragraph or two of hardcore sex and then finish with something like "and so Karla was swept into a daydream about her future life of shared laughter and love with her perfect partner."

Some others that i've kept around include. I can't particularly remember which of these are good or not, but the 13B param models have the advantage of fitting into 12GB VRAM:
  • "TheBloke_HornyEchidna-13B-v0.1-GPTQ"
  • "TheBloke_Wizard-Vicuna-13B-Uncensored-GPTQ"
  • "TheBloke_Wizard-Vicuna-30B-Uncensored-GPTQ_gptq-4bit-128g-actorder_True"
I haven't downloaded or used any newer models much recently as I've been focusing on catching up the progress of the artwork and game code implementations to the story parts that I already have created.

Regards the two extenstions I mentioned above, they are VERY useful for keeping the AI on track. Twinbook is probably the best - you have two panes of text. Left is the instructions, right is the previous output. The idea is that you add background info, and the prompt for the next section in the left, and can also freely edit the previously generated text on the right to remove or improve misfired sentences etc. By keeping up a process of adding "scene facts" to the left (i.e. current state of clothing), and manipulating the instruction prompt, you can get it to do a good job progressing the story on the right.

Playground is a bit more of a creative workspace where you setup various background info blocks and then work across 3 or 4 input boxes. You can end up with a similar result as twinbook but it's a bit more... "indirect", I guess.

There's a thread somewhere in "Dev Help" subforum that had quite a bit of discusison about the tools. Actually, that's how I got started using them. Ahh, here it is: https://f95zone.to/threads/ai-uncensored-text-generation-via-oobabooga.162713/ (warning, it is more than 1 year old so the info is potentially out of date)

As I said above, I've not used Oobabooga much for the past 3-4 months so the "state of the art" might have moved on a bit.
 
  • Like
Reactions: HoodedSilence

B0n3r4tacos

Newbie
May 9, 2018
53
29
Yes, I do. I use Oobabooga as a backend, and then two extensions by "Fartypants" as writing environments: "twinbook" and "playground".

For models, I have a 3090 w/ 24GB VRAM which means I can use the largest home-hobbyist models.

"LoneStrikerNoromaid-v0.1-mixtral-8x7b-Instruct-v3-3.5bpw-h6-exl2" just fits into 22GB free VRAM.
It's very strong but as with many of these models tends to go into "happily ever after" mode toward the end of every generation. By that, I mean it will complete a paragraph or two of hardcore sex and then finish with something like "and so Karla was swept into a daydream about her future life of shared laughter and love with her perfect partner."
How much context were you able to get away with with the larger models you tried? I have been running WizardLM-30B uncensored on the same GPU and couldn't load the model via oobabooga with more than 2536 context.
 

no_more_name

Newbie
Mar 3, 2024
91
26


Didn't break on a (real) 65k context (3090), 32k for general use I guess (or die waiting).
Quite good.
 

osanaiko

Engaged Member
Modder
Jul 4, 2017
2,420
4,273
How much context were you able to get away with with the larger models you tried? I have been running WizardLM-30B uncensored on the same GPU and couldn't load the model via oobabooga with more than 2536 context.
It's been a while, but I recall the context was at least 8K.
I did have to make sure I had no VRAM hogs running (no paused games, no chrome/chromium based browsers) and minimal Firefox tabs open.
 
  • Like
Reactions: B0n3r4tacos

B0n3r4tacos

Newbie
May 9, 2018
53
29
It's been a while, but I recall the context was at least 8K.
I did have to make sure I had no VRAM hogs running (no paused games, no chrome/chromium based browsers) and minimal Firefox tabs open.
I was able to achieve similar results recently. Really do have to have everything closed to be able to use it lol
 
  • Like
Reactions: osanaiko