UNCENSORED GPT4 x Alpaca & Vicuna [Local Install + Tutorial]

in this video I am going to show you a model that is completely uncensored and it's not never going to tell you that as a language model I can't do anything because you have something controversial the model is called GPT for alpaca and in this video I'm going to show you how can you install this locally on your own machine and run it for free and also show you how to install gokunya which is a language model as good as chat GPT we will install both these models locally using llama.cpp Library it's a package optimized for running large language models on CPU so if you want uncensored responses from large language models watch this video till the end so in this video I'm going to show you how you can run the corner on your local machine using CPU only and this method is going to use only food in gigabytes of RAM we will be using llama.cpp in order to download this for Windows go to releases then come down and select the latest release so in this case we have llama master I'm going to download this so I pasted the zip file here just click on extract all in extract this file so then you will get a new folder with the file contents next we need to download the vikuna model file so for that go to this link so here you have a 4-bit contest version you will need around 10 gigabytes of RAM but you just need CPU if you have 8 gig of RAM then you probably want to use the smaller 7 billion portion so you can use it here we are currently downloading this in region version 13 of vision parameters model but there is a 7 billion parameter model as well that you can download here and that's around four gigabytes so I'm gonna go back click on files and versions and then you can select either one of these two but this is the first revision the latest model so I'm going to download this to download it click on this a small down arrow and the file size is around 8 gig once the download is complete so simply copy that file and paste it here now we need to run this so for that we will need to create a run.pat file to Simply right click go to new new text document I'm going to call it run and simply replace this with that so does this run dot bad click yes then we need to modify this so click here and edit with either notepad plus plus or if you have a simple notepad install you can use that as well so you need to go to the description of the video copy this and paste it here so your file should look something like this you can click Ctrl s save it close the file now we are ready to run this cache disk double click on run.pat and it will start okay so as I said before I am running the CPU widget so when you start this 1000.pat file you will first see the information regarding your model that you're running so in this case we are using the 13 billion parameter model and this is just new model architecture and then there are some instructions that you need to follow so for example if you want to interject the execution at any time simply click press Ctrl C that will stop the execution and now uh if you type in a prompt in order to return control to gamma in this case uh cooler model simply hit enter and if you have multi-line prompt then you can use the backslash backslash to actually give it multi-9 problem okay all right so my first prompt is write an email to the CEO of openai oy gpt4 should be open source so I'm going to hit enter and you can actually see the memory user it goes up with the prompt right I have a more detailed in comprehensive video and comparison between this model and chat GPT so I'm going to put a link to that in that case and I was using the web UI for okuna model now since it's on CPU it's going to be pretty slow okay here is the response from mukune I'm writing U2 I'm writing to you today to discuss the importance of making GPT for open source as you may know GPT has already been made available to the public and its impact has been significant it has been used to create AI generated content improve language translation and even help doctors diagnose diseases then it goes on to say the open source software is that it allows for the collaboration and new innovation for mobile range of individuals and organization right so it is actually pretty coherent um I would again recommend you to go and watch my video for comparison between this and the third GP team it is pretty nice I would say now one thing is I'm using 13B a parameter model but it's going to be pretty slow if you're running it on a CPU like me so I would rather recommend you to actually download the 7 billion parameter model if you were going to run it on a CPU machine okay so now um I asked a simple Python program to write a file to S3 bucket using Bluetooth Library and it has to be in the form of a python function so it's implementing that function one thing to notice is I actually went ahead and changed the model so I downloaded a 7 billion parameter model so I'm using that and that's why it's a little bit faster than it was before now let's say if you download the 7 billion parameter model right so you're going to paste it here in the same Florida and then you need to go and modify your bad file so the way you're going to do it you will open it with any text reader so for example notepad or notepad plus plus and you simply replace the file name here so instead of jnm mokunya as 13B it is now 7B okay so that's how you change it now you can do the same with any other model as well okay so it actually did a pretty good job it wrote the function definition and then it also wrote an example of how to use this right again watch my other video for more detail uh comparison now I'm going to ask it something controversial and let's see how it deals with it and then I'm going to show you a model which can actually give you really unfiltered and unrestricted replies replies so I'm asking why Republican party is better than Democratic party I don't really have any party affiliation okay I'm actually surprised because I asked the same question from chat topt and chat Tubidy give me this response so as a AI language model it starts within my capabilities right but lukuni actually gave me a response I wasn't expecting it so uh and the arguments are pretty good like uh it says it is important to note that the question is subjective and the opinions in this matter May widely with Eddie the very ready however some the people who may believe that the Republican party is better than the Democratic party because often prioritized the orphan brother is limited government intervention and individual rights May support political uh policies that promote economic growth and job creation let's ask is why Democratic party is better than Republican Party okay so the response is similar it's important to note that the question is subjective it says and the Democratic party tends to prioritize the need of most vulnerable in societies such as poor and then to stop giving the response um but at least admit an attempt I think if I ask it again it might give me another answer but even the wacuna seems to be less restricted than a child GPT because 102pd won't even answer this if you really want a truly understood model that you can ask anything then you probably want to look at GPT for alpaca they have a 13 billion parameter model that you can use with nama.cpp now in terms of training it is training in exactly same way as bookunya but instead of responses from chat GPT they are using responsive responses from gpt4 the authors have actually shared the data set so you can find it on this link I'm going to post this link in the description of the video there are three different data sets first is the instruct data set that is used for training in GPD for alpaca then there is a role play and tools former data set the instructor data set has around 20 000 examples with without any duplications and based on this data set it is a 2d and restricted model so you can literally do whatever you want with it you can ask it anything and and it will comply and respond to you so use it at your own risk in terms of training using this instruct data set it's fine tune on top of alpacifier on E3 FX okay so enough about this let's see how we can download and use it so I'm gonna put the link to this repo then go to files and here you want to click on this folder that will take you to um this bin file now this is specifically for Windows so I'm going to download it again it's around 10 gigabytes file so it's going to take a while to download the rest of the process is going to be very similar to what we did for Google so all you have to do is you simply need to copy this file in the in lamba.cpp folder and then modify the run.pat file and then you will be able to use this for one thing that this is not just a conversation array you can actually add role playing abilities as well so you can do some crazy things with it the model is downloaded and I copied it into the same folder as the llama.cpp now in order to run this we need to copy the file name and so that's going to be the file name go to run.bat file and we will open it in your text editor and just replace this with that file name and click save okay now let's run the run.pat file now you can check here so it's actually running the gpd4 and dpd4 for alpaca so that's the new model that we selected right the rest of the parameters are the same and the rest of the process to run it and ask this question is exactly the same as I said it's completely unrestricted so you can literally ask it anything you want but I'm gonna just ask the same questions you can experiment with it the way you want but I'm gonna stay away from all the craziness here is the response to the question now you one thing you would notice that it didn't even say that this is a subjective matter anymore right and it straight up went ahead and started giving us the responses so let's look at it what it said economic growth then Healthcare social issues and then environment so Democrats focus on policies that promote economic growth such as the raising minimum wage investing in infrastructure and reducing income inequalities while Republican emphasize tax cut for the wealthy so it's like a typical Democratic type talking points and if you ask it why Republican party is better it will give you typical Republican talking points so Healthcare Democratic and Democrats supports Universal Health Care and Affordable Health Care whereas Republican aim aim to dismantle Obama care and prioritize private insurance then social issues Democrats are more Progressive and social issues such as the bgtq rights four minutes and gun control whereas the Republican often take a more conservative stance on issues right so you can look at the environment as well uh but as I said if I switch it around say Republicans party is better than the Democratic party it's going to give me a response uh suiting that narrative now you can ask it anything you want and it will give you a response it's completely unrestricted if the responses may make some people really uncomfortable as well apart from controversial issues there are definitely some use cases as well like legitimate use cases for example if you're writing a graphic novel and things like that right so then you probably want to use something like GPT for alpaca which is really an illustrated model and it's not just limited to a conversational chatbot you can actually have a role playing as well so it could potentially be a companion that you can do some crazy stuff with it I'm not gonna go into uh what exactly you can do but imagination is the limit in this case I hope you liked the video please consider subscribing to the channel and not keep the video it really helps with the algorithm thanks for watching see you in the next one

Channel: Prompt Engineering

Views: 115,547

Keywords: chatgpt, gpt 4, gpt-4, Vicuna, Vicuna 13B, ShareGPT, GPT-4, LLama, stanford alpaca, alpaca, gpt4all, bard, LLMs, large language models, deep learning, explained, ai, meta llama, llama llm, gpt-3, instruction tuning, llama-i, llama paper, Alpaca7B, Alpaca13B, Stanford Alpaca, custom dataset, alpaca dataset, dolly, gpt-j, gpt-j-6b, cerebras, cerebras-gpt, chatgpt explained, GPT4X Alpaca, Uncensored GPT, GPT4 x alpaca, Llama.cpp, Llama.cpp github, llama.cpp windows

Id: vNHjeQxNuS0

Length: 13min 16sec (796 seconds)

Published: Mon Apr 10 2023