@sierra bongwater turned freakywater

@sierra all ts effort for 2 likes and 1 boost vro 😭✌️

@sierra is this two balls and a chud

i asked grok to rewrite all of akkoma in rust tmrw kill-corp will be 67x faster 🙏 trust

@sierra idk how but ur finetuned model is somehow constantly using the chain-of-thought data even when it isn't enabled which is fucking funny

fine-tuning a 1B was such a bad idea

@sierra "grok can u make a readme.md out of what u just told me" ahh

who is this diddy blud on the LLM fine-tuning script

@sierra you're making it think?? are you also fine-tuning the thinking part????

@sierra all LLMs are just transformers so dw abt that! :3

@sierra i lied on accident there is no Ministral 1B and even 3B is commercial-only. fuck. anyways

i don’t think you will be able to avoid Python sadly as the two biggest open-source frameworks to train AI models are PyTorch and TensorFlow, which are both Python…

as for how, I’m looking at it right now, as I haven’t done it since GPT-2, but i’m very curious! :)

@sierra there's ministral (Mistral, european), gemma (da goog), maybe da Qwen's (Alibaba)? idk how good Qwen is anymore within the world of SLMs... but anyways yea these are the 3 big ones :3

Ministral and Gemma is available in 1B so u shouldnt have much problem, idk abt Qwen tho. hope that helps!

@sierra pingas

@sierra why is there a small PINGAS for the next dialog arrow

@sierra the normal evolution of burrs.s3

@sierra sierra.simplesecret.se... :3

outbound federation is broken on oomfie city :3

this is yet another federation test // ignore this thanku :)

@sierra green goop nitro badge