LocalLLaMA

2884 readers

6 users here now

Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.

Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.

As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.

founded 2 years ago

MODERATORS

pax@sh.itjust.works

SkySyrup@sh.itjust.works

noneabove1182@sh.itjust.works

Smokeydope@lemmy.world

MonsterBug@sh.itjust.works

Less positive model (lemmy.nz)

submitted 2 days ago by absGeekNZ@lemmy.nz to c/localllama@sh.itjust.works

8 comments fedilink hide all child comments

I'm currently running Gemma3, it is really good overall, but one thing that is frustrating is the relentless positivity.

It there a way to make it more critical?

I'm not looking for it to say "that is a shit" idea; but less of the "that is a great observation" or "You've made a really insightful point" etc...

If a human was talking like that, I'd be suspicious of their motives. Since it is a machine, I don't think it is trying to manipulate me, I think the programming is set too positive.

It may also be cultural, at a rule New Zealanders are less emotive in our communication, the LLM (to me) feels like are overly positive American.

you are viewing a single comment's thread
view the rest of the comments

[–] Sims@lemmy.ml 11 points 2 days ago* (last edited 1 day ago) (1 children)

Can't you just give it a system message that says: "User prefers a professional conversational style", "User prefers a neutral attitude", "Shut the F*** up!" or similar ? You probably have to experiment a little, but usually it works okay.

Edit. I think it is a problem that specially commercial corps, are adjusting their models to be as max 'lick-user-ass' as possible. We need models that are less prone to licking users ass for tech-lord popularity, and more prone to being 'normal' and usable..

[–] mutual_ayed@sh.itjust.works 1 points 1 day ago

I've had success restricting responses to 120 words or less.