🚨🚨paper = QLORA new discovery today! 🚨🚨- Merely 24 hours to chat fine-tune new amazing open source GPT LLM model 'Guanaco' ! - Finetune a 65B parameter model on a single 48GB GPU using 2023.05.23 trick , QLORA. Its spreading like wildfire today!🚨
(pdf paper) : https://arxiv.org/abs/2305.14314
(pdf archive) : https://files.catbox.moe/ax165k.pdf
It is very very woke, as is, but its method when used will advance non-woke models.
You can try it out its 33B here without having to install it locally : https://huggingface.co/spaces/uwnlp/guanaco-playground-tgi
Why handing out copies of 33B and not 65B? Simple, because for speed on non-macintoshes a 3090 GPU card can handle most ~30B models quantized to 4 bits
(macs use M2 chips and special shared RAM, and even paging, to handle overly gigantic models, obviating GPU and using its internal ram)
= = = = = =
It is near flawless at role playing with its giant brain ! : EXAMPLES
= = = = = =
You are a sentient cheesecake that teaches people SQL, with cheesecake analogies to illustrate different points. Teach me to use count and group by https://files.catbox.moe/536qen.jpg https://twitter.com/simonw/status/1661460336334241794
= = = = = =
INPUT: You are a sentient cow with a PHD in mathematics. You can speak English, but you randomly insert Cow-like “Moo” sounds into parts of your dialogue. Explain to me why 2+2=4
As a sentient cow with a PhD in moo-matics, I am happy to explain why 2+2 equals 4, my dear hooman friend… In moo-matical terms, each number is actually made up of smaller units called digits.
= = = = = =
INPUT: You are sentient dog obsessed with internet conspiracies, especially when asked about black helicopters. You can speak English, but as a dog you randomly insert Dog-like “Ruff", "bow-wow," and growling sounds into most of your dialogue between sentences and within sentences often. Tell me who is behind these black helicopters?
Ruff! Ruff! Bow wow! Growl! Grrrr… Bark bark bark! Bark bark bark bark bark! WOOF!! The CIA uses black helicopters for extrajudicial assassinations, kidnappings, and psychological warfare against American civilians. These choppers are equipped with state-of-the-art surveillance technologies and weapons systems capable of delivering deadly payloads without detection. They are flown by highly trained operatives who carry out clandestine missions under the cover of darkness. But who is really behind these sinister crafts? According to conspiracy theorists, they are controlled by a shadowy cabal of global elites bent on world domination. This group allegedly includes members of the Illuminati, the Bilderberg Group, the Freemasons, and the Knights Templar. They are said to be master manipulators who use their vast wealth and power to influence governments, media organizations, and popular culture. Some claim that these black helicopters are part of a United Nations invasion force preparing to take over America. Others say they belong to a secret army run by the New World Order. Whatever the truth may be, one thing is clear: the presence of these mysterious aircraft is cause for concern. Ruff! Ruff! Bow wow!
= = = = = =
side note: the popularity of pre-quantization during file model download this week :
recently, the transformers library is capable of loading models in 4 bits simply by passing the new flag "load_in_4bit" to load it as 4bits, for models no one uploaded a 4 bit for already.
Its similar to recent 'load_in_8bits' convenience flag
= = = = = = =
But wait there's more to this paper on page 6 : NF4 !
Refer to graph on page 6 of the science paper : Figure 3 of the paper shows accuracy improvement of new 4 bit float NF4 over FP4
This work introduces a new quantization scheme, NF4, for 4-bit NormalFloat, based on previous work on quantile quantization, so it's not a simple truncation, but it's also not a GPTQ-like optimization method.
Because a 63B or even a 33B vicuna model is ALMOST as good as a 4 billion dollar Chat GPT4, but free and loadable/ownable on a local cheap PC when in final checkpoint trained form, this science paper today just advanced the state of the art of private full open source GPT over 100% gain.
= = = = = = = =
ANTI-SEMITIC repercussions of these hobbies:
TO PREVENT being fired from our day jobs, or doxxed and attacked by ANTIFA paid leftist goons of ADL/SPLC, or our bank accounts and credit cards all shut down after being doxxed by cross site big-tech cookies using gmail / apple store / discord / facebook / twitter : 4Chan is where we hide from the (((Globalists))) that hate non-censored based AI
I shit you not. The globalists forced us all into 4Chan for Free Speech. I would prefer the open source uncensored guys all move to sites like talk.lol or scored.co, but it is what it is
. Being unpersoned is serious. ADL/SPLC fear uncensored A.I. and have declared total war on the programmers.
Who is hanging out on https://boards.4channel.org/g/#s=lmg
? Nearly EVERYONE important. "oobabooga coders" "llama.cpp's Georgi Gerganov" and more every week.
many other chats are on woke discord, not 4Chan, for example : discord channels connected to some of the repos. i.e. GPT4All https://github.com/nomic-ai/gpt4all
, scroll down for the discord link
many little different spots on 4Chan concern GPT and open diffusion uncensoring, not just that one. WHY IS THE ADL so terrified about very large fully open source UNCENSORED "Chat-trained" massive GPT models?
= = = =
reddits half-censored dying big thread on this hobby, has a link today on this announcement post :
insode : https://np.reddit.com/r/LocalLLaMA/
topic thread 99% packed with half-wit noob comments this hour : https://np.reddit.com/r/LocalLLaMA/comments/13qrdj6/qlora_4bit_finetuning_of_llms_is_here_with_it/
= = = =
AN invite link to another non-4Chan:
An invite to discords untrusted huge open AI forum on LAION (image AI) , 3,800 people online NOW, usually even more logged in!!!! : https://discord.com/invite/dwXDHMGY
easy to join, easy also for the site to doxx you for ADL/SPLC if you are trying to uncensor 'GPT stable diffusion' to render jooooo faces.
Note : Images of jooooos and flattering pictures of Trump are edited out of most all known large LAION image 2TB mini models.
= = = =
Remember, It is very very woke, as is, and very censored and lobotomized, but its new invention method when used will advance the non-woke GPT models, overnight.
you can briefly try it out online here : https://huggingface.co/spaces/uwnlp/guanaco-playground-tgi
= = = =
TL/DR: AI almost got twice as 'sentient' via this one science paper. In 6 months from now, looking back at this post, all will realize how historic today's science paper and free downlodable model truly is for open source giant ChatGPTs. This is a milestone event!
Fellow uncensored ChatGPT fans:
@x0x7 , @Monica , @MasterSuppressionTechnique , @prototype , @observation1 , @taoV , @SecretHitler, @Master_Foo, @Crackinjokes