â„–36483[Quote]
NUSOIGPT OFFICIAL LAUNCHHello soyim, I made an ai chatbot
trained off of 25000 conversation instances on /soy/, Its extremely lightweight, only 1gb of vram (or ram matter of fact) can run it. If your laptop was made within 5 gorillion years it can run it locally. I have also attached the dataset I used, s25kv4.
linkerino:
https://mega.nz/file/dmpHUCbA#eRI3_d7fNeB2MyiXnj0xzjrpQvO0-C0-4XJS4ASfWy0(this includes the gguf, dataset and license.)
<marge why didnt you use the 1 million database on hugging faceit isnt structured correctly, you need it in a question and answer format.
<marge idk how to run it is there a websiteI wouldve done a website however I was getting cucked by ngrok and plus, Ive never hosted a website before so I was not about to get doxxed because I was new.
technical info (snca):
base model: qwen2.5 1.5b
quantization size: q4km
dataset: s25kv4
trained on google colab (my esl gpu wouldve acked itself if I even thought of doing it locally)
powered with HWABAG license
DONT COAL THE LOG WITH IT, YOU FAGGOT â„–36493[Quote]
>>36483 (OP)I've never used local ai, how do I set it up after I download the file?
â„–36495[Quote]
DNI it exploits a zero day vulnerability in llama.cpp and locks your drive behind a bitcoin paywall
â„–36496[Quote]
why
â„–36501[Quote]
>>36493download lm studio
download nusoigpt zip file
set up lm studio
skip downloading recommended model
go to top right (download button)
go to download directory
create folder (nusoigpt)
create another folder inside of nusoigpt
it should look something like foldera/folderb/NusoiV2-QW2.5-1.5B.Q_4_K_M.gguf
ex: C:\Users\bibisi\.lmstudio\models\nusoi\nusoigpt\NusoiV2-QW2.5-1.5B.Q_4_K_M.gguf
go back to lm studio
crtl + l
load nusoiv2
â„–36555[Quote]
>>36495ERM chuddie, there's a reason its called .safetensors
â„–36556[Quote]
someone just leaked ~250K posts from the sharty, it was mainly an IP leak but it also works as a archive for a quarter million posts, you should use those 250k posts to train your next AI
â„–37131[Quote]
>>36483 (OP)This datamining nigger glows; DNI otherwise you lose 10 centimeters in height and penile length
â„–37134[Quote]
marge is this actually glowware?
â„–37137[Quote]
>>37134probably not but id still dni
â„–37151[Quote]
> kuzcord
Kuzcord is a 4cuck party website. The site is used for the promotion and discussion of the party and its members. The site has been around since 2004. It is managed by the Kuz party. The site is based in Canada.
[ Prompt: 51.8 t/s | Generation: 16.3 t/s ]
>
<flooderald
â„–37171[Quote]
keyed
â„–37201[Quote]
>>37192IT DID IT AGAIN WITH A TWITTER LINK
â„–37204[Quote]
your bot is broken retarded nigger
â„–37335[Quote]
>>36483 (OP)Mistral-Small-24B-Instruct-2501 version when? its too incoherent rn.
â„–37337[Quote]
>>37335why mistral retard?, just use qwen3.8-27b
mistral is ancient, even qwen3.5-9b would btfo
â„–37338[Quote]
>>37335what caused you to think of such an outdated model
i smell a retard who uses ai
â„–37339[Quote]
though really you should finetune qwen3.6-35b-a3b as most people can run that unlike 27b
â„–37351[Quote]
this shit is broken nigga use a better model
â„–37548[Quote]
>>36483 (OP)i heard ai needs like beefy computer to run does run on an i3 11th gen with 16 gb ram
â„–37620[Quote]
>>37337using qwen 3.8 27B with modest quant seems like the best option IMO
â„–38197[Quote]
>>37620You'd need at least 24gb of VRAM to fine-tune a 27B model
â„–38203[Quote]
>>36555It is not a .safetensors though
â„–38212[Quote]
>>38211qwen3.8-9b isnt a real model, ur using some retard's finetune
qwen3.6-35b-a3b could potentially work (only 3b params active)
i would actually use that
â„–38213[Quote]
>>38211you should also train a model with the ip leak shit, the posts being the q and the ips being the a
basically the model would try GUESS the ip based off what you ask it
â„–38214[Quote]
also release the lora adapter or wahtever tf its called again
â„–38215[Quote]
>>38212Yes, it is a finetune, im not using a base model, its the first time im doing this and im retarded about it really just enabled some filters in hugging-face and selected the one i liked the most, if this doesnt work out i'll just redo it with a base model
â„–38216[Quote]
>>38213Can you link that leak, because i dont have the one that has IPs linked to posts
â„–38217[Quote]
are u using like unsloth or something
â„–38219[Quote]
i wonder how much resources it takes to qlora finetune 35b-a3b
Would it be 3gb?
18gb?
â„–38403[Quote]
>>38252any progress on dis?
â„–38405[Quote]
>>38197Hes using glowgle online GPU so it should be fine, though it would be slow to run fully locally
â„–38406[Quote]
>>38252This one is 100% realistic
â„–38407[Quote]
>>38406Meant toreppey to
>>38404 â„–38434[Quote]
geggy thread
â„–38438[Quote]
>>38252you clearly did no research
â„–38723[Quote]
>>37131it's the brimstone 'o virus not even linuxfags are safe
â„–39128[Quote]
when will the second iteration be released
â„–39210[Quote]
>>39128If you mean the 1M dataset one, i probably wont release it now because it just sucks geg, it will just spam (You)s and call you a samefag or big nigger lips spammer every 3 messages, probably because as
>>38438 said i did zero research and my dataset wasnt filtered enough, i'll try to train it again in the near future doe, also with a new dataset from 'farms leak, because first time i forgot i even had it geg