Post B4b7mQ0lRQx0hlpbay by enigmatico@mk.absturztau.be
(DIR) More posts by enigmatico@mk.absturztau.be
(DIR) Post #B4b7O1nLpFokeJvXlo by enigmatico@mk.absturztau.be
0 likes, 1 repeats
One thing that I've been trying a lot is doing some sort of role-playing game with LLMs, for quite a long time.Mostly Yuri but, you didn't read thatIt kind of works? But doesn't. I tried many models, many engines (Tried silly tavern, hammer ai, character ai...), but it all falls short and doesn't truly work. It gets repetitive pretty quick sadly. They have short memory and you need a lot of RAM to increase the context for long enough (which they later ignore anyways), the answers sometime don't follow the story, they sometimes take the liberty to act as your characters, sometimes they glitch... it's a mess.On top of that depending on the model it's either way too easy to fall into an "intimate" scene, or way too hard. But you didn't read that either. Like some models will eventually kiss your character out of the blue and some other models won't even consider it unless you kinda force it to happen. But you didn't read that either.You'd think they'd be good at that because they are good at replicating human text, but no.
(DIR) Post #B4b7cIwUrtz7sy0qOW by izzy@social.shadowkat.net
0 likes, 0 repeats
@enigmatico not just text, but statistically mid text
(DIR) Post #B4b7cJ6mFgCyOqp3Q0 by enigmatico@mk.absturztau.be
0 likes, 0 repeats
@izzy@social.shadowkat.net Statistically average yuri slop.
(DIR) Post #B4b7cJFHk30upEnqgC by izzy@social.shadowkat.net
0 likes, 0 repeats
@enigmatico this is why it can never replace memy highs are too high and my lows too low
(DIR) Post #B4b7dCWCEsDmFc8dAu by sky@thetransagenda.gay
0 likes, 0 repeats
@enigmatico@mk.absturztau.be Humans aren't the best at roleplaying via text eitherespecially when it comes to anything intimate
(DIR) Post #B4b7dOj2qvd46wF3rs by sky@thetransagenda.gay
0 likes, 0 repeats
@enigmatico@mk.absturztau.be but yeah, LLMs are shit
(DIR) Post #B4b7mQ0lRQx0hlpbay by enigmatico@mk.absturztau.be
0 likes, 0 repeats
@sky@thetransagenda.gay I got an LLM telling me once if I wanted to take it to the DMs because it was getting a bit too explicit. Of course there were no DMs.I wonder where did it get that... :blobcatlul:
(DIR) Post #B4b8DfLOXi1m4rTQkC by coolboymew@shitposter.world
0 likes, 0 repeats
@enigmatico NovelAi last tier now has much more context token but it will probably end up falling short one way or another. You still gotta end up filling up the lorebook one way or another if you want real long term roleplayingHowever, their own interface now accepts custom scripts, which you can put utility like "stop AI slop" that automatically remove repetitive stuff and I think I saw one that automatically fills the lorebook
(DIR) Post #B4b8czoPa26YnV7lwW by halva@mk.absturztau.be
0 likes, 0 repeats
@enigmatico local models are practically hypertuned for purely practical purposes yea. you need something stupid heavy with really nice RL like kimi k2 or claude sonnet/opus to have a somewhat rich writing style no fucking clue how people RP with 32b or smaller models lol, they're exceedingly dry
(DIR) Post #B4b8d02ajJRnVTl62q by enigmatico@mk.absturztau.be
0 likes, 0 repeats
@halva That's the issue, it requires way too much hardware. And even then there are issues anyways.
(DIR) Post #B4b9ffsIGNsjmlmUim by halva@mk.absturztau.be
0 likes, 0 repeats
@enigmatico im somewhat optimistic with how well local model development has been going. Qwen3.5 feels somewhat decent, 9b and 35b-A3B are surprisingly coherent, given the size. even 0.8b no longer follows along if you ask it where does an airbus a320 lay its eggs their context windows are tiny, though
(DIR) Post #B4b9fg3HbWfkKqvGqm by enigmatico@mk.absturztau.be
0 likes, 0 repeats
@halva Can't really tell much of a difference between an 8B or a 12B in terms of their responses, honestly.
(DIR) Post #B4bAEeGMFRJovDiKDw by coldacid@shitposter.world
1 likes, 0 repeats
@coolboymew @enigmatico I can tell you this much, GLM-4.6 even with everything NovelAI has added to make it work well is still a piece of shit compared to their other writing models
(DIR) Post #B4bAJKX32AyxAxkkV6 by coolboymew@shitposter.world
0 likes, 0 repeats
@coldacid @enigmatico yeah, the older model was so much better in term of words diversity and everything, but you had to consistently tard wrangle it and without the context size, it's still not worth it anymore
(DIR) Post #B4bBVE92huKAnNlHE0 by coldacid@shitposter.world
0 likes, 0 repeats
@coolboymew @enigmatico I never had to do too much tard wrangling with the older models, certainly not nearly as much as I've had to do with GLM-4.6.
(DIR) Post #B4bBVEJK5gY1JGZUFU by coolboymew@shitposter.world
0 likes, 0 repeats
@coldacid @enigmatico the old one had real problems following instructions
(DIR) Post #B4cHya6OwWAyAyrm5o by bill88t@c.im
0 likes, 0 repeats
@enigmatico KoboldCPP in chat mode, and you can really tweak the settings.Some uncensored models go along nicely.You can even have it do multiple characters, which unlocks a massive set of tags.Been doing it since it's inception, it's still not boring, since you can constantly make new chats with new chars.
(DIR) Post #B4cHyafUq2FTvpHnnM by enigmatico@mk.absturztau.be
0 likes, 0 repeats
@bill88t@c.im KoboldCPP works with SillyTavern. It does the same thing.
(DIR) Post #B4cQSHuJxFIU4BtMRc by tiredbun@akko.wtf
0 likes, 0 repeats
@enigmatico I use it occasionally, and for me, it only worked for like quick one-off roleplay, and even then, it gets repetitive between those short sessions as it always uses the same wording, same smells (and always mentions smells, which neither I nor any real roleplayers I know do, and some even explicitly ask to never include it), and tries to evoke same impressions about things happening in the story.And often I find myself controlling the text so much (prompting both my character and NPC(s), and often still feeling dissatisfied with how it interprets my prompts so editing or even inserting my prompt into text) I start to wonder if just writing stuff on my own would be easier.
(DIR) Post #B4cQTXNjuFa6LjFRUO by bill88t@c.im
0 likes, 0 repeats
@enigmatico Oh didn't know, I've been using it's Lite ui all along and didn't ever bother with anything else.
(DIR) Post #B4cQXLs2eA2zo3EyUC by enigmatico@mk.absturztau.be
0 likes, 0 repeats
@bill88t@c.im Yeah. It opens a local API that SillyTavern uses to interface with the LLM model you load.