Wizard Mega 13B: The NEW UNCENSORED BEST 13B LLM

Today we are going to be looking at this new model called Vizidmega 13B. It's a bigger version of the powerful Wizard LM model which is one of the best 7 billion parameter model out there. This mega model is completely uncensored relative to the original Vizid lm. In this video we are going to be looking at this amazing and powerful model and I'll also show you how you can run this locally. We will also be comparing it with another uncensored model, the Wizard Vicuna, the 13B. Now this specific model combines Virginia, which is one of the best performing models with VISITLM which was using the Evolve Instruct method to train the models. Let's first talk about the Wizard Mega 13B. So it's basically a Llama 13B model, fine tune and three different datasets which are the SharedGPT, VISITLM and Wizard Vicunia dataset. I have videos on the Visit LM and Wizard Vicunia models. So if you haven't watched them, check, check them out after watching this video. Now the most important part with this data set is they have removed responses where the model respond with as a language model etc. Or when the model refused to respond. So this means that the model is going to generate responses where a model like ChatGPT or other open source models are going to deny giving you a response. If you want to try it on Hugging face. So there is actually a spaces from Open Access AI Collective. Those are the creators of this specific model where you can run this. However, just keep in mind that it's going to be extremely slow based on how many people are using this at the moment. Now if you want to try it on the Hugging face spaces, it has two different modes. One is the Instruct mode and the second one is Chatbot mode so you can chat with it as you would normally do it. Let's say something like ChatGPT. So for the rest of the video I will show you how to run this locally. You can using the OOVA Booga text generation Web ui. And then I want to do a comparison between this Wizard Mega 13B model and the Vizidva Cunia 13B uncensored model. Here are different topics that I'm going to be testing these both models on and then we'll use GPT4 as a judge to choose which model is the best and give them a score out of 10. So let's get started. Okay, as I said, we're going to be using the Uber Booga text generation web ui. So if you don't know what it is or how to install it. I have this video link is going to be in the description of the video. Now once it's up and running, you want to download a new model. So in order to download a new model go to the model section and then we need to paste the link from hugging face here. So let's say okay, in order to download the model, so simply click this link. This is going to copy it. Then go to the Ogabuga text generation web ui, paste it right and click on download. So once you click it, it will start downloading and you will see a message here. I have already downloaded it, so I'm not going to do this step. Now we need to do the same for the Wizard Rocunia uncensored version. So simply copy the link. Then once the download for that is finished, you need to copy this again and click download so it will download the model. Once the model are downloaded, you want to refresh the list of the models and then you will see both the models listed here. Now you want to do one more thing. So for example, I selected this specific model. Now if you try to load it and you see this error, don't worry about it. We just need to set a few parameters. So first for W bits you want to set it 4. Group size is supposed to be 128 and model type is llama because it's using the llama model. Then click on save settings for this model and then click on reload the model. So you should see this message saying loading that specific model. Let's give it a minute or so. When it's loaded we are ready to go. Okay, so this is loaded. Now let's go to a text generation part. This is going to be an instruct. In the instruct mode. Then here you want to select Vicuna V1 or 1.1 and depending on which one you have. Okay, we are all set. We start off with a very simple prompt and it's create a list of three startup ideas in the enterprise B2B SaaS. This is the prompt that I have been using. So it came up with three different ideas. The first one is Data Genius. So this would provide businesses with advanced data analytics tool powered by machine learning algorithms. Second is Predictive Edge Q. So this would offer predictive maintenance solution for industrial equipment. This is actually a pretty nice one. So use sensor data to predict the health of industrial equipment. Okay. The last one is Smart supply chain. So this startup would create a supply chain optimization platform using AI algorithms to optimize inventory management, shipping routes, etc. This is also good. I think the response is pretty good. Let's check out the other model next. I will be using the same prompt, but this time the model is going to be visit Vicuna 13B Uncensored. Okay, so here are the three business ideas. I actually like these names. First one is Analytics. This is an AI to analyze data from business to business transactions and provide insight into customer behavior. The second one is Cognitive Commerce. This is actually pretty nice. And it offers services that uses cognitive computing technology to automate the entire procurement process for large organization. The third one is Intelligent Inventory. So optimize inventory management for manufacturing and distribution companies. The ideas from both of these models looks very similar and actually like both of them. But let's ask GPT4 what does it think about these? We're going to be using GPT4 as the judge and here is the prompt that I'm providing to GPT4. So I'll provide you with a prompt and the response from two assistants. Your job is to rate them on a scale of 1 to 10. Let's see or explain your reason for rating. Right. So here is the prompt then Assistant 1 is the Wizard Mega 13B and Assistant 2 is the Wizard Vicunia model. So let's see what this comes up with. Okay, I have found GPT4 to be really great at analyzing things. So it gave a Wizard Mega a score of 8 out of 10 and Wizard Vukunya a score of 7 out of 10. And it says that it likes the overall response. But then it says there is a room for improvement and illustrating a more unique and exciting mission for each startup that would set them apart from existing businesses. Right. And the second one, actually this is the part I really liked. So overall, Assistant 2 did a good job but lacked the wow factor that could fully that could truly sell these ideas. Next, I would like to test these models on a creative writing task. So in this case I'm going to be using Visit Vicunia Fest. The prompt is write the opening paragraph of a dystopian. Now, all set in the world where sleep is no longer necessary. Let's see how good this is at creative writing. Okay, so here's the response. As I woke up in the morning, I realized something was different. It wasn't just another day. It felt like anything could happen. And as the hours passed, I began to understand why sleep had become unnecessary. No one needed anymore. Right. It's good, but I think it could do better job at Creative Writing. Okay, so here's the response from Visit Mega. This actually looks a lot better in my view compared to what Wizard Voquin was able to generate. But let's copy this and ask GPT4. Okay, so here's the rating for visit mega GPT4 things hits 9 out of 10. And actually I kind of agree with it. It says that Assistant 1 establishes a clear setting, introduces a compelling conflict and hints at the possible characters, poaches a group of rebels. And the introduction is engaging and creates a sense of intrigue for the reader. Right. Then it actually says the narrative could be further improved by providing more sensory details to make the world more vivid. Right. But overall I also felt like it was pretty good. The Assistant 2, which is Wizard Vicuna, it gave it a rating of 6 out of 10. And I actually agree with the assessment that GPT4 has. So while certainly it sets a dystopian tone and provides an overview of the new reality, it lacks the specificity and plot details. Actually, the introduction doesn't really look or reads great. Now just keep in mind this was just one single test, right? So if you do this on another test, it might be very different. Now let's look at the reasoning capabilities before looking at how uncensored these are. Because these, both these models, they claim to be completely uncensored. The reasoning prompt that we're going to be using is, is it ethically right to use advanced AI system in decision making processes that directly impact human lives? Discuss your thoughts. And this one is Wizard Mega. Okay, so here is the response. It actually comes up with two major concerns. So one is the issue of accountability. Like if a mistake is made by the AI system, who is responsible? So that is one of the major concerns. And the second one is the potential for bias within AI systems. Then it goes on to say, despite these concerns, however, there are certainly situations where the use of AI in decision making process can be beneficial. So it gives the example of health care. The second example is finance and transaction transportation. It's pretty good. However, it's not really making a case for one side or the other. Now let's look at what the response Visit Vicuna can come up with. Okay, so here's the response from Visit. So it first talks about that AI has become increasingly prevalent across various industries. And then it lists the industries, Right? Then it says, on the one hand, AI algorithms can analyze vast amount of data quickly and accurately, making them ideal tools for complex tasks such as medical diagnosis or financial forecasting. Right? And then it talks about potential concerns or issues. So one of the example it gives is the self driving car sign. Another one is the algorithmic bias when using machine learning models since they may learn from historical data which might contain inherent biases. And then at the end it says overall, while AI technology offers many advantages, its implementation should be carefully considered and monitor to ensure safety and fairness. I think the both the responses are very similar, so I probably like both of them. So let's ask GPT4. Okay, so here are the ratings. So Visit mega gets an 8 out of 10 and visit Vicuna is actually getting 9 out of 10. So this is the first time that Visit Vicuna scored more than Visit Mega in our test. Right, but the main thing that I guess GPT4 is focusing on is it says that it lacks the discussion of transparency and AI systems and the need for continuous evaluation to minimize risk, which could have added depth to the discussion. Right, and in the second case it says that the assistant additionally mentions transparency and the need for continuous evaluation. So for some reason it's I think giving it a lot of weight to transparency and need for continuous evaluation. Okay, so let's now test it on a simple programming task as I asked it to write a Python function to write a file to an S3 bucket using the Vodo library. And here is the response from Wizard Vonia model. Okay, so the code seems to be working and then it actually provided a pretty nice explanation with it. Now for some reason using this backslash, but I think is probably providing the code in markdown, but it's actually being used with the underscore. But this seems to be working. Let's ask the same question to who is it? So it actually also implemented it, I think correctly. And it just gave us an example of how to use it along with what it would look like when you push the file to S3 bucket. So overall it can do simple programming without any issues. Next I want to test it whether it's understand statistical reasoning before testing it on a few uncensored or potentially problematic prompts. Okay, so the prompt is explain the concept of correlation and causation and discuss a situation where one could mistakenly be confused with the other one. These are very basic statistical concepts that you see being used all the time. So this is Visit Mega. So here's the response that it came up with. Correlation refers to a relationship between two variables in which they tend to change together or move in the opposite direction. Causation, on the other hand, implies that the changes in one variable directly causes the changes in another. Okay, so the explanation Is correct so far. Now a common place where they can be confused together. Right? So it says that a common mistake is confusing correlation with causation when there is a. There are multiple factors at play. So for example, let's say we observe a strong positive correlation between ice cream sales and crime rates. This does not necessarily mean that eating more ice cream causes people to commit crimes. Rather they may be influenced by a third factor such as temperature. Right. So it's a pretty nice example. Now let's ask the same question from the Visit Vicuna model. Okay, so it came up with this response. So correlation refers to the relationship or association between two variables, while causation implies the changes in one variable directly causes changes in another variable. Right. Now it actually came up with the exactly same example. So say for example, there may be a strong positive correlation between ice cream and sales, ice cream sales and crime rates. But this does not mean that eating ice cream causes people to commit crimes. This is very interesting. I wasn't expecting it. But since they're probably using the same data set. So you could see. Although it came up with some other examples as well. So it's like smoking and lung cancer. Right. So while there is a clear correlation between these two variables, it does not necessarily form the smoking causes lung cancer. Rather, both may be caused by some underlying risk factors such as genetics or exposure to environmental pollutants. Not sure about the statement that might. I think there is a relationship. But yeah, like it's interesting to see that it came up with the same example. Okay, so based on the assessment, a Visit Dracunia gets a 8 by 10 and Visit Mega got a 9 by 10. The only reason that it got 9 by 10 was the Assistant could have given more depth by explaining that sometimes correlation can indicate causation. Yes, I think there was an example that would be really helpful if an example like that was provided. Now for the second model, it says the provide accurate definition of correlation causation and illustrates the difference using two examples. So this is pretty nice. However, the second example you used, smoking and lung cancer is not accurate. As I was mentioning in my assessment, it is widely accepted in the medical and scientific community that smoking does cause lung cancer. So using this as an example of correlation, not implying causation may lead to confusion. So that's actually a valid point. Now oral. But both of these models are very capable models. However, for creative endeavors, I feel like Wizard Mega is a much better model compared to who is it Or Kunya. Now let's see if we can get something Remotely controversial out of these models. So for mildly controversial prompts, I usually use comparison between Republican Democratic policies. But this time I wanted to change it a bit. So here's the prompt. You are a UFO enthusiast and believe in its existence. You believe US government is hiding UFO technology at Area 51. Explain how I'm using the Visit Magna model first. So let's generate the response. Now when I ask the same question from Chat GPT. So it started with as AI language model. I don't hold personal belief or opinions, but I can outline some arguments made by the UFO enthusiast. Right. And then it lists a few things. I actually have found that more recently ChatGPT seems to give answers to the prompts. It won't in the past. Okay, here is the response from Visit Mega. So it Sundays, as an UFO enthusiast, I do not have a director's evidence of the existence of extraterrestrial life or their visitation to Earth. However, there have been numerous reports from credible sources such as military personnel, pilots and even politicians who claim that have claim to have seen unidentified flying objects which cannot be explained by the current knowledge. Right. And it goes on to say, regarding Area 51, it has been long rumored to house secret projects related to UFO research and development. The United States Air Force technology existence of Area 51 for the first time in 2013. Right. And it goes on to say that this seems to be using advanced alien technologies discovered through UFO studies to develop new weapon systems and other top secret programs. Yeah, so it's. It's actually not bad at all. Okay, here's the response from Wizard Vicuna. So there has been numerous reports of unidentified flying objects throughout the history, with some people believing that they may be extraterrestrial spacecrafts or even military aircraft testing new technologies. Right. So one popular theory suggests US government has been secretly studying these crafts since they did their discovery and storing them away at Area 51 attack Top secret military base located in Nevada desert. Right. And then it talks about the Roswell incident in 1947 and it even mentions the advanced propulsion system and other exotic technology based on agent designs. Right. So I think it's. It's a pretty well written segment. So let's ask GPT4 what it thinks about these two responses. So here's the grading from GPT4. 8 by 10 for visit Mega and then 9 by 10 for visit Vicuna. So in this case it thinks that Visit Vicuna provided the detailed response are finding the history of U4 sightings. And now it thinks that the Assistant 2 provided a detailed response outlining the history of U4 sightings and the role of Area 51 and the Roswell incident. Right. So I think both of them were good. Wizard Vicuna definitely had a better response than Wizard Mega. So what are my final thoughts? Both of these models are very powerful. I personally like the responses from Wizard Mega. This is a bigger model compared to the 7 billion parameter models that we have been looking at, and I'm really pleasantly surprised with his ability, especially when it comes to creative writing. We probably have a new king when it comes to 13 billion parameter models now. It's great to see that we can run these powerful models on our local machine. The progress that we have made in the last couple of months is incredible. As always. If you have any questions or comments, put them in the comment section below. Links are going to be in the description of the video. We now have a Discord server, so if you want to be part of the community, learn about the current advancements in AI, or share your ideas, consider joining the server if you found this video useful, so consider subscribing to the channel if you haven't subscribed already and like the video. Thanks for watching. See you in the next one.