Unsloth Studio is insane… fine-tune any AI model locally

Fine-tuning is insane. It allows you to have a small LLM outperform models 100 times bigger, create uncensored AI models that answer any question or prompt, cut your API costs to near zero, and build a powerful moat for your business.

But until this point, fine-tuning came with two major problems. Number one, creating your own dataset was very difficult and time-consuming. And number two, the entire fine-tuning process was hard to do locally on your own computer.

Luckily, there's a new open-source project that solves both of these issues. So, in this video, I'll explain what fine-tuning is and why it matters. I'll also show you how to train your very own AI model locally on your computer, and I'll even walk you through the steps of creating your own dataset for fine-tuning. So, if you are serious about AI, make sure to watch until the end.

So, the open-source tool I mentioned earlier is Anslo Studio. This is by far the easiest way to fine-tune LLMs. And with Anslo Studio, you can do that on your own machine, literally on your own computer, completely offline. It was built by a former Nvidia engineer and his brother who used to fix bugs in Llama and Qwen themselves. And by the way, I reached out to Anslo to sponsor this video and they agreed. So, huge shout out to them.

Now, let me show you how to set this up. So, first go to the docs. I'm going to link this below the video. It's going to be the first link. And here, click on quick start. This is going to scroll you down automatically where you have the different one-liner installer commands for different operating systems. So, I'm just going to copy this entire command. Boom. And go into terminal. You can use iTerm2 or your default terminal inside of Mac OS. Doesn't matter. Just use any terminal. Okay? And I'm going to paste in this command right here. This is going to install Anslo Studio and all of the dependencies that it needs. And again, I want to stress this. It's completely open-source and completely free to use. So, unbelievable value from the Anslo team to the world, just making it easy to not only fine-tune AI models but also create your own datasets to fine-tune models on your own computer, completely for free. It's incredible.

Okay, there it is. Anslo Studio installed. We just do Y, hit enter. So, as it says, starting Anslo Studio on localhost:8888. So, just go to browser and type in localhost:8888. And you can see we are in the studio, right? So, we have multiple things here we need to understand. First, we can chat with AI models. So, no matter which model you have downloaded, as you can see, I have a couple of them. For example, Anslo Qwen 3.6 27B. Very nice model. Actually, one of the most powerful models you can run locally on a MacBook. You just need like 24-32 GB of RAM or VRAM to run this, and you can chat with it, right? So, not only is Anslo the tool to fine-tune models and create datasets for fine-tuning, it's kind of also like a competitor to Ollama and LM Studio because you can also chat with these models inside of Anslo, which is pretty incredible.

Okay, so the model has been loaded and we can do a test prompt, something like "Who are you?" just to see if it's working. And again, this is running fully locally on my MacBook. I do have 128 GB of RAM, so you're probably not going to be running 27B that fast. But again, this is not the main point of Anslo Studio. I'm just saying like, this is a huge bonus that you can also chat with LLMs locally.

Okay, but again, let's focus on the fine-tuning. So, on the left, we have multiple tabs. We have the train tab. This is the main thing for fine-tuning. We also have the recipes tab. This is how we actually can create our own datasets for fine-tuning, right? They call it recipes just to make it a bit more approachable to people. But really, Anslo Studio has made this easier than ever before. So, if you looked into fine-tuning last year and you were like, "Damn, this is too complex. This is too difficult." Guys, it's never been easier to fine-tune an AI model. And not only that, it's never been easier to create your own unique dataset for fine-tuning and then use that to train a new model and have [clears throat] a complete moat, right? Having your own datasets, having your own fine-tuned models, that is still a huge moat that most people don't know how to do. But after watching this video, you will have this skillset that 99.9% of people in AI simply do not have.

So, first, let's go to the left and let's click on train. Oh, and by the way, one more thing. When you open this for the first time, you might see like a login, um, password thing. This is just locally setting a password so that people on the same Wi-Fi on the same network cannot access it, right? So, just put in some password, but it's not really any real account on, on the web. It's just to secure Anslo because again, we're running on localhost. Someone on the same Wi-Fi could technically access it. So, yeah, just put some password the first time you do it. But then go to the left and click on the train tab.

And here, we need to select what model to train and on which dataset. So, first, I'm going to show you how to train a model, how to fine-tune a model. And then I'm going to show you how to create your own dataset for fine-tuning. So, let's break this down step by step. And again, you don't have to be a developer to do this. I'm going to show you everything in easy-to-follow steps. Anybody in 2024 can fine-tune their own AI model. And I'm going to prove that to you right now.

The first thing we need to do here is select the model, right? So, we can see this model tab right here. And we have two options: select a local model or a Hugging Face model. So, if you don't have anything downloaded locally, let's just go to Hugging Face and select something. So, just type in Hugging Face into Google. And this is basically kind of like the GitHub but for AI models, for datasets. Everything you need when it comes to open-source AI model development and inference, right? So, when we get to the homepage, we have models, datasets, spaces, all that stuff. So, click on models and you can see loads of different models, over 2.9 million, almost 3 million different open-source models on Hugging Face.

So, Anslo actually has their own. So, we can do an re-upload of different versions of popular AI models such as Qwen, such as Llama, such as Gemma, right? So, Gemma is developed by Google, Llama is developed by Meta, Qwen is developed by Alibaba. Anslo doesn't make these models. They debug them, they improve them, they make them better for fine-tuning and for inference.

And by the way, if you're not sure which AI model to choose or what you can run locally on your own hardware, there's a cool site called Artificial Analysis. They have an open-source section which again, I'm going to link this below the video. If you scroll down, you can see all, all open-source models, large models, medium models, small models. So, for most of you, probably small models is the right category to look at. And as you can see, Qwen 3.6 27B is the best in terms of intelligence. So, right now, Qwen 3.6, especially the dense model 27B, is the best in this size of small models between 4 billion and 40 billion.

So, when we go back to Hugging Face, we can actually look for different variants of Qwen 3.6. So, let's do /qwen 3.6 to see all of the Anslo versions. And we just want the pure one, 27B. So, there it is, Qwen 3.6 27B. This is the one we want. Now, it's important to understand what Anslo does to these models because again, we're not using the raw version from Alibaba. We want to use the Anslo version of Qwen 3.6. So, first, they fix real bugs in the models. So, they work directly with the teams at Google, Meta, Alibaba, Mistral, Microsoft, the teams that create these open-source models, they work with them and fix the bugs after the official release. Anslo also adds dynamic 2.0 quantization. So, instead of compressing only select layers, they dynamically adjust the quantization type of every layer with a custom scheme per model. This shrinks the model size massively while still keeping the accuracy. In plain English, it allows you to run a much more powerful model on the same computer, which is great for us because we want to run it locally.

So, we know we're going to be going with Qwen 3.6. So, inside of Anslo Studio, when we go to here, type in Anslo/qwen 3.6. And again, select the variant that's pure 27B with no MLX, no GGUF, nothing like that. If you want the default safe tensors version, which is exactly here. And by the way, you can just copy-paste this string from here if you don't want to type it in. And you will make sure you have the correct model. In terms of a method, keep it on CUDA. It's very optimized for local hardware. And then the second thing we need before we can start fine-tuning this model is a dataset. So, again, Qwen 3.6 27B, 627B, already a really good model. Like, when you're on a plane, you can ask it questions about basically anything and it'll give you 85% of what ChatGPT or Claude will give you. Obviously, it's not as powerful as Opus or GPT-4.5, but this is a very, very strong model that you can run fully locally if you have at least 32 GB of VRAM. But it's not fine-tuned. It is generic. Everybody has the same one. So, if you want to fine-tune it to maybe make it a bit less restricted, or make it better at finance, better at coding, better at your own custom dataset, I'm going to show you how.

So, first, we need to select the dataset so we can start fine-tuning this model. And by the way, everything from this video, all of the presets, model names, datasets, everything I mentioned will be linked in the second link below the video in a bundle. You can grab it completely for free. So, when you see me doing something like copying a model name or running a terminal command, all of that is free in the second link below the video. So, go get it now.

Now, notice here I mentioned we need the safe tensors version of this model and not a GGUF. So, let me explain what even GGUF means because most people have no clue what this term is. But you'll hear it a lot when you start running local models or look into fine-tuning and all that stuff, right? So, GGUF basically is a compressed, ready-to-run version of an AI model packed into a single file so it can run fast on a normal computer. That is the plain English explanation. You can think of it like a zip file of the AI model. It's shrunk down so your laptop can handle it. But the catch is that it's shrunk down for running for inference, not for training. So, if you want to train, aka fine-tune, a model, you need a full, uncompressed version, aka the safe tensors version, which is the one we've got open now inside of Hugging Face right here.

The second thing you must understand before you get into fine-tuning models is Llama.cpp. Now, this is very deeply tied with GGUF. GGUF is actually built around Llama.cpp and it's a popular inference framework in C++. Actually, the same person created both of them. More on that later. That's why they fit together so naturally and so well. And Llama.cpp only runs GGUF. Every other format must be converted first to GGUF.

Now, let's talk about Georgi Gerganov. So, he's the guy who created Llama.cpp. So, anyone can run Llama models on a normal laptop because, you know, if you remember two years ago, Llama still was on the cutting edge of open-source models. Now Meta has gone kind of fallen behind and, you know, companies like Qwen, Kim, Deepseek are much better when it comes to open-source AI. But back then, good old 2024, Llama, Llama 3 70B was it? That was the best model when it comes to local AI models. So, his goal was simple: to use powerful models on consumer hardware such as Llama 3 70B. You could run it on consumer GPU, whether it's Nvidia, whether it's the one inside of your MacBook, doesn't matter. You didn't need to rent an H100 in the cloud for hundreds of dollars a month, right? No need for that. You can run it on the hardware you already have. And actually, in February of 2024, his team joined Hugging Face. So, everything is kind of uniting in a single company for the local AI stack, which makes it simpler and easier for us to run and fine-tune models locally.

Now, the next step in our fine-tuning setup is choosing the dataset. And this might be the most important decision after choosing the model. Maybe this is the most important decision. Choosing the dataset, the correct dataset is super important. Okay. So, again, we can just use Hugging Face because we have datasets here. So, we can switch back, click on datasets, and we can see that there's over 1 million different datasets we can choose from. Now, again, you might be thinking, "Okay, David, which one do I choose?" Look, simple guide. Click the second link below the video. Um, it's explained here. But for this demonstration, I'm going to choose a finance dataset so that Qwen 3.6 27B can answer finance questions better. And again, choose the one for your own use case. So, let's say you run like a legal company, law firm, you know, whatever. You type in law, and you can see that there's a bunch of different datasets about law from different companies, from different countries, different structures, different amount of downloads, different sizes of datasets. Or you want to do something with code, you want to create a powerful coding model, there's a way more code dataset obviously, but there is over 1 million different datasets on Hugging Face. Again, in a bit, I'll show you how to create your own dataset in case you cannot find one. But first, let me show you how to use it with one of these datasets.

I'm going to go with the Finance Alpaca dataset, which has nearly 69,000 rows in this exact format. You have the instruction and you have the output. Sometimes there's an extra input. For example, if we click on this row, you can see the instruction is "For a car, what scams can be plotted with a 0% financing versus rebate?" and the output is a detailed answer on that topic. Right? And there is over 68,000 examples like that. So, when we use a dataset like this, the model, which in this case is Qwen 3.6, will become really, really good at these types of questions. And this, ladies and gentlemen, is the power of fine-tuning. You give it great examples, and the model will learn those examples and become really, really good at that specific domain. So, again, choose your dataset carefully.

Now, there are different types of datasets for fine-tuning. First is the instruction, question, and answer. Single turn is the most common. This is the one I'm going to show you. There's also conversational with multi-turn chats, there's domain expert, your own niche data, and reasoning or tool use, chain of thought, function calls, stuff like that. Each shape teaches a different skill. So, pick the one that matches your goal. If you're not sure, just go with question and answer. It's the simplest, classic. All the models have been trained on that. So, this is what we're going to use. And the reason I chose a finance dataset is that it's a very real, useful use case. It's not a gimmick. It's something that would be immediately useful to basically anyone interested in finance. You can run the same model locally on the same hardware. But if you ask it questions related to personal income, investing, anything in the finance niche, suddenly the same model, consuming the same amount of electricity, would be much better in that domain and it would perform way stronger on finance benchmarks and finance evals.

So, here we can copy the name, just like we could copy the name of the model. We can copy the name of the dataset. Go back into Anslo Studio, go to the left and paste it in here. There it is. Finance Alpaca. Click that. And as you can see, it's checking the dataset uh in the bottom right. So, we need to wait a few seconds. In the meantime, we can change a few of the parameters here. For example, context length. We can drop it to 1,024. So, the training is a bit less compute-intensive. We can also go to training hyperparameters. Switch batch size to one to make it a super small run. And again, if you were to train a bigger model for longer, maybe renting out a few H100s in the clouds would make sense. But the beauty of Anslo Studio is that you can do everything locally. You can do it on your computer. You can do it for the hardware you already own with the electricity you're already paying for. So, we're going to do this and I'm going to do a smaller run so I can show you how simple that is. Maybe we can even lower the steps to 20. Obviously, the more steps, the longer you train it for, the better the results will be. But even with this, I think we can get some solid results. And my main goal is to show you guys the way, is to show you the steps required to train your model, right? And again, do this for your own use case. In a second, I'll show you how to create your own datasets. And all of you have some Google Docs. All of us has some PDFs, some CSV files, some Excel files, some exports of WhatsApp messages, conversations with clients, YouTube videos, transcripts of calls. All of us have some data. You can use that data privately, securely to fine-tune a custom AI model to be really good in that use case. And this is the power of fine-tuning. And this is exactly what I would encourage all of you to do. Sure, for the first fine-tune, so you're familiar with it, use a dataset from Hugging Face, something that's already pre-made. But again, creating your own datasets has never been simpler. And I promise you all of you by the end of this video can fine-tune your first AI model.

So, once we have configured the parameters, selected the dataset, we can go to the right and click on start training. And this will begin the fine-tuning process fully locally here on my machine. So, there's two important things, right? We can click on the current run once that starts going and see the history of runs. So, let's see here. We can also switch to iTerm2 here to see the terminal. Scroll down. You can see that this is running. Okay, I already... Okay, so two important things, right? First, the dataset. If you don't have the dataset downloaded, Anslo will first go and pull that from Hugging Face. This is not that big, you know, 40 megabytes for 68,000 rows. That takes a couple seconds to download. But the model is the bigger thing, right? So, I'm using the Qwen 3.6 27B and that is 51 GB. So, if you don't have a lot of disk space or if you want to do faster, you probably should select a smaller model. For example, if we go here, we can type in Anslo/. Boom. To see all of the different Anslo models. They have over 1,300 models, so you can definitely find the one of your size. You know, if you want a beast one, go with 122B or 397B. If you want a smaller one, just go with 9 billion models. You know, Qwen, that's actually a great one. Anslo/qwen 3.5 9b. This one, all of you should be able to run this one. Unless you have a 20-year-old computer, you should be able to run this model with no issues. And, uh, yeah, you can definitely fine-tune this pretty easily. And again, make sure you have the default safe tensors version. Anyways, let's go back to Anslo Studio. And obviously, this model will be much smaller because it's 9 billion compared to the one I'm doing right now, which is 27 billion.

Okay, so we're running here. Look at this. Look at this. Uh, we're training step three of 20. U GPU utilization is nearly 100%. And I'm also recording at the same time, guys. So, if my camera is lagging a bit, please pray for my MacBook here. But you can see the VRAM. We still have plenty of space left. We're on step three of 20. Training loss is not going down yet. That's not good. We should see start seeing this going down. In fact, if you go to history and I can look at this run, even though it failed, we can look at this training loss how it goes down over time, which means the model is learning the dataset. Okay, so let's go back to the current run and let's hope the training loss starts going down. And you can actually watch the whole process here inside of Anslo Studio, or you can also switch to the terminal where you launched, um, the Anslo command earlier to see like the detailed logs. But you can see the training loss is going down. I can start hearing my computer because it's going on full power now. So, training loss is dropping, which means this basically means how likely is the model to predict the next token on this dataset. Right? So, this is literally how well the model has learned the dataset. So, you need this to be going down, otherwise the model is not learning. It's not improving. So, uh, this will run for a while, even though I only have 20 steps. It'll probably take another 20 to 30 minutes. So, again, this is a very small run, only 20 steps. Like, okay, now loss is dropping. Beautiful. So, obviously, this will not create any insane fine-tune. Ideally, I would run this for way longer on maybe even a bigger dataset, but I think the dataset is fine here. You need at least a couple thousand examples. Ideally, at least 1,000 minimum. Very, very minimum is like 200-300 examples. This dataset is 68,000. So, that's more than enough to be fair. But again, for a more impressive fine-tune, do more than 20 steps. Do a larger context window. Do multiple epochs. Let it run for many hours, not just a quick demo like I'm showing you right here.

Now, a quick warning. There is a bug on MLX right now, which means some of these models on Apple hardware like M-silicon chips can fail even when you have enough RAM. As you can see earlier, I had like over 80 GB of VRAM left. This is not about VRAM. This is about the metal allocations or something like that, and this is a bug in the MLX. So, this is nothing to do with Anslo, nothing to do with Hugging Face. I think Apple engineers need to fix this. So, if bigger models like 27B are failing, maybe consider jumping to, you know, 9B. So, that it's a smaller model, the training run is faster, and it doesn't need as powerful of hardware to run. Or if you want a much bigger run and you want a large model, then just spin up a few A100, H100s in the clouds. You probably are going to have to spend a couple, couple dozen bucks or a couple, you know, $100-$200 to run some GPUs, and you can run that in the cloud. But again, for this demo, I want to show everything locally.

Okay, so now that you understand how to fine-tune a model by selecting a model, selecting the dataset, and some of the basic parameters, we're going to go to the left to the recipes tab. And this is how you can create your own custom datasets for fine-tuning. Because so, so far, we've been using a dataset that's publicly available on Hugging Face, which is fine. I mean, again, this is more than 99% of people will ever do in AI. But if you want the real advantage, you want your own data and you want to fine-tune on your own dataset to create an AI model that's unique to you, to your business, that nobody else has access to. So, let's do that right now.

Okay, so let's log in on the left. Click on recipes, then click on new recipe, and start from a learning recipe. These are some presets that Anslo gave us to make it easier, right? Because if you go from scratch, you absolutely can do that, but it's a bit more complicated, right? So, just click on start from learning recipe. We have multiple different options: text to Python, text to SQL, PDF document QA, OCR document extractions, GitHub crawler, right? So, this crawls GitHub issues and PRs and, uh, creates them into training pairs, instruction for answer, like a lot of useful stuff. But the easiest one, I think, is the PDF document QA. You give a PDF document, could just be one PDF, and it will create question-answer pairs from this. It will create your own dataset from a single PDF. This is huge. So, click on that.

And we have four different nodes and, you know, three like explainers, whatever. Basically, these are the main nodes. Okay: Provider 1, Provider Column LM, Structured, and Document File. I'm going to break it down for you. It looks a bit intimidating, but actually, it's not. Provider 1 is the provider we want to use. Now, you could use a local model if you want to do it fully locally and you don't want to leak any data anywhere. You absolutely can. You can host it on an Ollama server, or even like the Anslo can expose a local, um, server, I think, or LM Studio, or whatever. Anything that can expose the model locally, you can use the endpoint here as a localhost. But the beauty of fine-tuning is that you can actually use a more powerful model and use the distilled outputs, right? Which is what basically everybody's doing with Anthropic models. It's kind of the unspoken secret of the AI industry. People take Opus, the latest version of Opus, and they take its outputs and train a smaller model on that, which makes it closer to the powers of Opus. Obviously, it will never reach the true capability and intelligence of Opus if it's like 100 times smaller in terms of parameters, but it can get close, right? So, what we want to do is we want to use a big, powerful model to create the dataset so that we're basically teaching our small model, which in this case will be Qwen 3.6 27B, on how a bigger model would answer, right? So, we'll kind of, it's like a trick. You're making it smarter by showing it how a bigger model would answer on that type of dataset.

So, let's do that. So, first, we need to create the endpoint. We need to put in the API endpoint and an API key. And I think the simplest is to use OpenRouter because all of the models are on OpenRouter. So, here, just paste in `https://openrouter.ai/api/v1`. And again, all of this will be inside of the bundle, second link below the video. So, in case you want a terminal command, some model name, dataset name, something from this video, just grab that bundle. It's completely free. The second link below the video.

Now, let me jump back into Anslo Studio. And we need an API key because this is an API service. So, let's head to OpenRouter. Top right. Make sure you create an account. Takes like 20 seconds. And then click on credits. Charge up some credits. Okay, could just be $5 or $10. That's enough. Go to API keys and click on new key. I'm going to name it Anslo finetune. Boom. Put some credit limit like $50. Let's click on create. Copy. And by the way, guys, if you're watching this, please check if you're subscribed because according to YouTube Studio, the majority of you are actually not subscribed. And we are so close to 400,000 subscribers. Subscribing on YouTube is completely free and it tells me, "Hey, you enjoyed this type of content and you want me to make more videos like this one." So, if you want me to make more videos on fine-tuning or creating datasets for fine-tuning and making this type of technical advanced content that most YouTubers would never even dare to touch, please hit the subscribe button. Go below the video. Check if you're subscribed. It takes two seconds and it's completely free.

Okay, so I'm going to copy this API key and I'm going to go back into Anslo Studio here. Paste that into this field. Click on save in the top right. Next, we need to go into provider column. So, this one is, uh, done. The provider has been configured. Provider_1. We're using openrouter/api/v1. This is important. The `/v1` at the end. Okay. And then we put in the API key. Next, we need to select the model. So, what model we want to use? We definitely don't want to use GPT-4 Mini. That's terrible. Let's use something intelligent, maybe like Sonnet 4.6. You know, this is not as powerful as Opus, obviously, but it's a very good model. Um, way bigger than you can run locally. And, uh, still great for distillation, right? So, uh, boom, paste that in. By the way, make sure to check Anthropic's terms of service. I'm not like advising for distillation or nothing like that. But if you think about it, Anthropic or OpenAI, they're distilling the knowledge of the entire web. Like, what they do is they take all of the literature, all of the websites, all of the code from the internet and they distill it into an AI model. And then they complain when somebody distills the outputs from their AI model into small open-source models. Anyways, let's not get into this debate too much. Just, you know, choose whatever model you want here. I recommend going with something powerful. I think Sonnet 4.6 is a great balance. If you want the best of the best, use the latest Opus or the latest GPT model.

Next, the Okay, click on save in the top right. Again, click save. Next, we have the element structure output. So, actually, if you click on configure, you can, uh, you can configure this system prompt and the response format that you want. Honestly, you can just leave it to default. It's fine. But the important thing is the document file. This is how your dataset will be created. Whatever you upload here is going to be in the question and answer pairs, right? So, this could be again, an SOP from your business. This could be a list, a CSV file of all sales calls that your closers had. This could be a list of all the, all the paid ads that you ran and the results of them. Right? Whatever you want to fine-tune on, make sure you provide a quality PDF here. I'm going to click select file or actually configure, click on configure, and then click here to select the file. I have, uh, downloaded an Nvidia file, which is a financial year for 2026, just a public report. And this is like an 80-page PDF. And we're going to fine-tune a model on Nvidia's financials. So, this could be used maybe for, uh, predicting the next quarter for investing, for, um, you could probably sell this to some private equity firms or some investors. I don't know. This is just an example. Again, you can use any type of dataset, any type of PDF. All you just need is a single PDF. Obviously, the more you have, the better. Okay. The more pages the PDF has, the better as well. But basically, upload this PDF. Click on done. Boom. Then we need to click on check. We can do a check run. Maybe only with five records. And we can do a test run here. Boom. It's ready to run. So, click on start test run. And this is going to do run in progress. Inside of OpenRouter, we can actually go to logs. And we can wait a few seconds to see Sonnet 4.6 appearing here in the logs. Error, blah, blah, blah. Is there some Anthropic issue? May 28, elevated errors. This is literally crazy. We're getting elevated errors. So, let's select maybe Gemini 3.5 Flash. This is a new model. Let's select that. Boom. Jump back into Anslo Studio. Let's switch that up. Boom. Close that. Save. Let's do the check again. Start a test run. It's kind of crazy that as I'm recording this, there was an error in the Sonnet API. Okay, there it is. Gemini 3.5 Flash is working. This is another great choice, by the way. A very powerful model. It's not really a flash, I don't know why they call it flash. It's pretty expensive for it to be a flash model, but it's a very competent model from Google. Um, anyways, let's see if this finished. Okay, it's finished. Inside of OpenRouter, we can go to logs and we should see Gemini 3.5 Flash. It costed us a couple of pennies to do this test run. And the most important thing is that we have the dataset created, obviously with only five rows. So, this is not a sufficiently large dataset. It's just a test. So, now we can see the chunk test here. And we can see the element structure output looks like this. So, this is the basically what happens is it takes the PDF, which is 80 pages, and it chunks it down into different chunks, right? So, it's smaller. It's not the full PDF in every single, uh, message. It's just chunked up. So, if you have 80 pages and you do five chunks, you do 80 divided by five, right? So, here, if we scroll to the right, we can see the output LLM structure one. The question is, "For what financial year blah blah blah was this filed?" and the answer is the January 2025, uh, January 25th, 2026. Okay. So, this is what it looks like. If we scroll down to the next one, "What's the registration telephone number including area code?" and the answer is that. So, we're fine-tuning the model, or I mean, we're creating the dataset that will be used for fine-tuning the model on this Nvidia financial report. Now, again, this was a test run, just five rows. So, go back to the editor and let's click on run and let's create a proper run, full run. Okay, we can name it something like Nvidia financials 2026, whatever, just something human-readable. And we can do way more than 100 records. We can even do 1,000 records. Okay. And then, uh, advanced run settings, batching, uh, if you don't want to risk it like failing, you know, you can do that. Basically, it's going to create a dataset with 10,000 rows, all about this PDF from Nvidia Financials, right? So, click on start full run and just let it run. And this is why it's important to choose a reasonably expensive model because if you chose Opus, this would be very expensive, right? Because again, we're going to be making 1,000 API calls to OpenRouter API to create that dataset, which again, dataset is the, the thing, the most important thing you need for fine-tuning. So, choose a model that's still intelligent but still cost-efficient. Honestly, the best one would probably be DeepSeek. Like, I don't know why I didn't choose that first. DeepSeek V4 Pro, absolute killer model for this cost. Uh, so, yeah, actually use DeepSeek. Forget about the Sonnet. Forget about the Gemini 3.5 Pro. Use DeepSeek V4 Pro. If you want even cheaper, use DeepSeek V4 Flash. Okay, this is a dirt-cheap model, super fast, but not as powerful as the V4 Pro. So, I think DeepSeek V4 Pro is like a good default for all of you to use, uh, for this training, for for thousands of API calls, because look at this cost, beautifully small cost, is almost free, right? Like China is cooking, shout out Xi Jinping, crazy costs, and it's fully open source. Anyways, let's go back to, um, Anslo. If we can see, okay, we can click on runs, see that how this is going to logs as well here. See if we're getting, yeah, we're adding hundreds of, uh, of API calls to Gemini 3.5 Flash. Luckily, this is also pretty affordable. But again, I would definitely recommend, I would definitely default to DeepSeek V4 Pro. Super powerful model. It's the most powerful open-source model right now and it's dirt cheap. So, this is what you want for creating a large dataset, right? So, we're going to let this run. It might take a couple of minutes to create this dataset. After all, it's 10,000 rows. But the beauty is that next time we go to the left to train, we can actually use our own dataset, right? So, we can click on local and we're going to see the dataset right here. So, instead of using something from, um, Hugging Face, we're going to have our own custom dataset that we can train the model on. Whether that's 27B, whether that's, uh, the Qwen 3.5 9B, doesn't matter. Obviously, it will impact the quality of the model. But the beauty is that we have our own dataset created from within Anslo Studio right here in the recipes tab. All from a single PDF, which hopefully all of you can find something useful in your life and business to create a dataset out of. And then you can use that to train your very own AI model, your fine-tuned version of an open-source AI model. This is absolutely beautiful. It's absolutely magical and I truly recommend all of you to do this. The feeling of running your own unique model that nobody else has, it's kind of hard to describe. And again, in this video, I showed you all the steps. You now have all the steps to not only fine-tune AI models, but also do that locally and with your very own custom datasets that you can easily create from a single PDF. And again, Anslo Studio is completely free. So, a huge thank you to Anslo for sponsoring this video. They made an amazing tool. This is fully open-source. It's free to use. Just go get it. It's going to be the first link below the video. Install Anslo Studio and do this now because, you know, watching is good, learning is good, but the, the thing that's actually important, the thing that actually will make you better and might change your life is taking action. So, go through this video again and implement all of the things I showed you and build your own fine-tuned AI model. And as I mentioned earlier, all of the resources mentioned throughout this entire video are in the second link in the description. Go grab it now, completely free. I'm going to email you the resources. So, yeah, with that being said, thank you for watching and have a great rest of your day.