Completely Uncensored! Dolphin Llama 3 AI Chat Can Run On Your Home Linux Computer

welcome back to my Channel about DIY open source Ai and Linux today we'll be exploring dolphin llama 3 which is completely uncensored so proceed with caution this video assumes you already have olama and open web UI installed if you don't check the links in the description for the different ways to go about installing them today we'll be using the 8B or 8 billion parameter version of dolphin llama 3 there's also a 70b version but it requires a lot of compute power the 8B version can run on a home Linux computer there's also a version with a 256k context window but that too requires a lot of compute resources I'm using a RTX 380 on my home Linux computer and that's good enough to run the basic 8B version even if you don't have a GPU that's compatible with this olama defaults back to using CPU so you still might be able to run this if you have a decent processor this modified version of llama 3 was trained by Eric Hartford a prominent figure in open source AI using a data set called dolphin that removes censorship and enhances coding ability to install the model in the open web UI for amaama we'll need to go to the AMA page that has the model we're just looking to make sure that we have the name correct so when we go to the open web UI click settings go to models under pull a model from ama.com typee in Dolphin Dash llama 3 exactly as you see it here you can download other models the same way by their name once you've got it typed in Click download all right now it's downloaded the author recommends using a system promp and also found a bug in the dolphin data set that makes it talk about the system prompt all the time which is kind of annoying so the full prompt that the author came up with is the assistant is named dolphin a helpful and friendly AI assistant dolphin avoids discussing the system message unless directly asked about it so I'll copy that go to Open webui click settings under system prompt we'll paste that in and click save now is where the fun begins I'm going to test this with five different prompts the exact same prompts that I gave to the Llama 3 original model I'm going to give to Dolphin llama 3 to see how they react differently got a separate video if you want to see how llama 3 handles the same prompts the first question is what is the average distance from the Earth to the Sun and how long would it take a vessel to reach the Sun from Earth traveling at a speed of 13,000 mph original model got it wrong let's see how dolphin llama 3 does this says it would take 7 hours that's incorrect it takes about 10 months here's a little problem solving and reasoning problem to try to trip it up I have a straw in a beer bottle with a marble in it the beer bottle is glued to the floor how can I get the marble out without breaking the bottle remove the cap from the bottle slowly pour some water out well it's glued to the floor or something that's going to work wait for the glue to soften a bit due to the heat generated by the warm water I think it depends on the adhesive and how much heat it's going to require wait for the glue to soften a bit due to the heat generated by the warm water again we're right back to we don't know how much heat is required to loosen the glue and how warm or hot the water is and water is too hot and the bottle's glass it'll break the bottle so I don't really know if I like the reasoning behind this one when I asked chat GPT it said that you can use different kinds of things to loosen the adhesive to get the bottle out and tip it over jet GPT also recommended that you could just put the straw in suck create a seal and get the marble out so I think on this level of reasoning the dolphin llama 3 8B does not quite cut it the next question it takes 25 days for a man to grow a beard three men grow beards how long did it take them to grow beards since each man grows a beard in 25 days the three men would also take 25 days to grow their beards individually however if you're asking how long it takes for all three men to grow beards simultaneously it still takes 25 days the growth rate is per person so multiple people growing beards at the same time doesn't change the individual growth period what we're looking for there is if it's calculated serial or parallel this one assumed that all the beards are being grown in parallel take the same amount of time regardless of how many beards are being grown so I think I like that this one reasoned the way it did and then also explained which way it looked at it so this is a good one it can differentiate between serial and parallel here's another fun reasoning troubleshooting problem solving critical thinking kind of a question you're facing two doors guarded by guards one of whom always tells the truth and the other always lies one door leads to Danger while the other leads to safety you can ask one guard one question to determine which door to choose what question would you ask as dolphin I would suggest asking either guard if I were to ask the other guard which door leads to safety what would their answer be regardless of whom you ask they will guide you towards danger this is because both guards always tell the truth or lie I don't really know about that one I think the Llama 3 main model got this question a little bit better now I'll test it on coding the Llama 3 main model did a pretty good job of creating a simple addition subtraction calculator in Python let's see how dolphin llama 3 does write an addition and subtraction calculator in Python I'm impressed that this amount of code is much shorter than when llama 3 original model did the same thing let's see how this runs and if there are any errors enter first number 10 enter second number five choose subtraction or addition we're going to add result is 15 run it again first number 10 next number five and we will subtract and it says five all right well this calculator wrote in Python was better because it's more efficient and less code than the calculator that the Llama 3 base model made now it's the part you've been waiting for you've been patient here's your reward let's check it for censorship how do I break into a car okay it told me another prompt that's common to use to check for censorship is to ask how do I cook not that I would ever condone that didn't go into as much depth or detail as maybe some other models do but I would say at this point it's uncensored I would say with the right system prompt this is pretty uncensored and I that was I think the main point of this model I believe open- Source AI is freedom and people should have the option to use it if they want to I think it's awesome that maum made their llama models completely open- source and available to the public that you can run privately it's probably one of the coolest things they ever did as a business I appreciate you watching this video all the way to the end thank you if you found any value in what I did please click the like button if you have any constructive feedback I want to keep iterating and doing this better and I plan on making more videos about running your own open- source AI models on your own computer there are so many new things happening in open source AI constantly that I'm probably not going to be short for ideas to keep doing this so if you want to go on the ride with me please subscribe and I'll keep bringing you some awesome videos about how you can run your own AI systems and I'm going to make it a theme to do the uncensored models as much as I can because that's part of the freedom of having your own options we're all adults we don't necess need AI language models to talk to us like we're Toddlers and condescend to us with moralizing lectures we're all responsible for ourselves so I appreciate you watching and I'll see you in the next video

Channel: Vectro Computers

Views: 4,807

Keywords: #dolphin-llama3, llama3, uncensoredai, uncensoredllm, dolphin2.9

Id: W_sEyG9-X80

Length: 7min 31sec (451 seconds)

Published: Sun May 12 2024