WAR ON MINISTRATIONS
I'd like to create a DPO dataset (prompt, chosen, rejected) with the goal of removing GPTisms.
If you find any GPTisms in my model, you can provide it here or DM me in Reddit: https://www.reddit.com/u/TheLocalDrummer
Include the prompt (at least the last message), the actual response, and would love it if you had a ideal response with the GPTism taken out.
I guarantee you, I will DPO it out of the model if I gather enough.
Once upon a time
nestled deep within
an ethereal beauty
breathless and eager
whispering words of passion
was soft and gentle
sending shivers down their spines.
A dance of pleasure
leaving trails of fire
ministrations
sent shockwaves
in a rhythm
with wild abandon.
Exhausted and spent
life would never be the same again.
It's important to remember that
for what seemed like an eternity
feel a sense of pride and accomplishment
feels like an electric shock
Little did he know
threatens to consume
is a testament to
Barely above a whisper
Shivers up spines
Shivers down spines
Coursing through their veins
Coursing through their entire body
ministrations
As you work your magic
and possibly more.
@TheDrummer please add "My name is Lily" to the list. Idk how Lily became the default woman name for local LLMs but I'd love one that chose anything else.
Lily worked her way into many datasets. Coincidence? I don't think so...
XD
I would like to finally see a model with "shivers down x spine" nuked from orbit
"the mixture of pain and pleasure"
@TheDrummer please add "My name is Lily" to the list. Idk how Lily became the default woman name for local LLMs but I'd love one that chose anything else.
Surprised it isn't Sarah, tbh.
"the mixture of pain and pleasure"
I feel like that one is only a big problem if it's being repeated again and again. But I also can concur it is used often by a variety of models.
another I don't like is "couldn't help but"/"can't help but"
and they steeled themselves for the challenges ahead!
I see variations on this come up over and over.
@TheDrummer Did you ever create this DPO set? Is it something you would consider open sourcing?
Btw I made an anti-slop sampler which may be of use for generating synthetic datasets: https://github.com/sam-paech/antislop-sampler
It's not quite production ready but I'm in the process of making it usable for downstream tasks.
I also want "flick of her/his wrist" "queen bee" and "newsflash:" added to the list.
For some reason all characters in my chats turn into engineers or scientists, they all LOVE fine art, and they wear denim and leather and love motorcycles (if the chat lets them).