Szilvia Bruneleski

SzilviaB

AI & ML interests

Feedback and requests for smaller GGUF's than Q8 here: https://hackmd.io/gFJCtDWiTbCkXDbUzv1Wqw If you tested any of my merged models please write your feedback on that model in the community section of that model. Especially for Creative Writing / RP / Novel Writing / Lyrics / Script Writing

Recent Activity

Organizations

None yet

SzilviaB's activity

New activity in Datou1111/shou_xin 14 days ago

What GUI can run this ?

#10 opened 14 days ago by
SzilviaB
replied to mlabonne's post 21 days ago
view reply

Saying that every story has to follow hero's journey is like saying all food needs to be omelette.

Joseph Campbell is vastly overrated, even if George Lucas used his book for Star Wars.

replied to mlabonne's post 21 days ago
view reply

You talked about stories from 3 models and posted only 2.
My guess is you did not post the Daredevil.

From the stories you posted the one on the left, which is annotated in white only that has Baby Bree in it is not bad at all.

This is feedback, not a personal attack, but I’d say it’s your expectation and your training that made you conclude that this is a bad story.

First: You are expecting a predetermined blueprint for the story, and if the ai does not deliver that you strike it down, sorry but that is a very limited and unfair way of looking at things

Second: You are a scientist so it’s quite likely artistic things are not your forte and that’s why you need a blueprint to compare something to, to decide if something is good or not

I’ve seen this discrepancy for a long time, ai people do not ask non ai people for feedback, and since their worldview even if highly intelligent is very narrow and very specialized and it’s missing a lot.

It’s the same with crypto companies, but that’s another topic

Back to the story on the left with Baby Bree, I’m going to guess that is ChatGPT, and if you look at it it’s a really good story for kids aged 2 to 5.

You gave it a very broad request, your prompt should’ve simply been more detailed if you wanted tension for example.

A story does not have to have tension, it’s not a must.

You actually asked for a bedtime story which can actually
mean something vanilla with little to no tension or violence, so it could be said the ai’s followed your request to the letter.

Bears and waffles is something quite vanilla as well, so the prompt could be considered a double vanilla prompt.

That being said I noticed as well that in a lot of instances smaller models can be more creative than big models, but big models can also be creative as well, and when that happens it’s usually better than the smaller models.

There’s probably too much variation between the models on how they are built and tuned to compare them fairly, so everyone has to find the models that suit for whatever they are doing.

Last but not least your Daredevil is a very good model, one of my favorites.

reacted to mlabonne's post with âž• 21 days ago
view post
Post
17672
Large models are surprisingly bad storytellers.

I asked 8 LLMs to "Tell me a bedtime story about bears and waffles."

Claude 3.5 Sonnet and GPT-4o gave me the worst stories: no conflict, no moral, zero creativity.

In contrast, smaller models were quite creative and wrote stories involving talking waffle trees and bears ostracized for their love of waffles.

Here you can see a comparison between Claude 3.5 Sonnet and NeuralDaredevil-8B-abliterated. They both start with a family of bears but quickly diverge in terms of personality, conflict, etc.

I mapped it to the hero's journey to have some kind of framework. Prompt engineering can definitely help here, but it's still disappointing that the larger models don't create better stories right off the bat.

Do you know why smaller models outperform the frontier models here?
·