Fine Tune Class 1 - What Is A Fine Tune?
===

Elizabeth: [00:00:00] Hi everyone. Welcome to the first class of the Fine Tunes. Class of the Fine Tunes extravaganza. We had been working with Fine Tunes in the Future Fiction Academy since December of 2023. We actually started researching on them in October and November, and it took us quite a while to kind of figure them out.

We shared that post on the Open AI forums and we were immediately contacted by John Allard, who's actually in one of the videos that I give you a link to, about our work in Fine Tunes because we were one of the first organizations to publicly talk about. Long form content, fine tunes. It was something that the people at Open AI didn't even know it was really possible at the time.

They had been focusing on short pieces for fine tunes. That's what most of the examples had been. And I'm happy to tell you now that a whole year later. , advances in fine tunes are something that authors should not sleep on. There is more and more content being published every single day with either [00:01:00] AI assistance, long passages with AI in it, or AI writing the entire book.

Yeah, that's happening. And I think there's like two ways to deal with this. You can either be like, I need to protest and boycott and try to stop it, which none of us really have the power to do. Or you can start to learn about the technology and figure out what does it mean for you being in this industry?

How do you remain competitive for the next five to 10 years? And I think fine tunes is actually the ticket for that because a fine tune. Is your special sauce? A fine tune is your special model that you've tweaked and made your own. And that's something that everybody else will not have access to.

It's something that only you will have access to. So even as production, speeds kind of require authors to at least think about AI for some part of their workflow. A fine tune is what allows you to still leverage ai, but. Bring that special, [00:02:00] unique voice that's yours to this amazing technology so that basically you can kind of clone yourself.

All right, let's take a look at your notes.

So we're gonna talk a little bit about fine tunes and like why this class. , like I said, models are becoming more advanced every single quarter. I don't know when you're taking this class. I'm recording this right now in, the end of February of 2025. But. , no matter when an author comes to learn about AI and think that maybe I should take a look at this, the biggest thing that I hear from authors is I can't get it to write like me.

A lot of them will work on prompt generations. They'll take prompts from people, they'll practice them, and they'll realize that they just get frustrated because they can't quite unlock making the model sound exactly like them. And truthfully, if your author voice is. More unique than not. You don't sound like the majority of the writing that the AI was trained on.

So there can be a quite a big mismatch. So what you can do is the answer is a fine tune. You can fine tune [00:03:00] select models, on open ai, minstrel, and Google. Minstrel's important minstrel means that you can fine tune models that will allow you to have not safe for work content. So if you are a romance writer and are sick of these pedestrian love making scenes, you can actually fine tune, mytral so that it writes like you write your love making scenes or your sex scenes.

You're not stuck with how AI often likes to just summarize it, you know, they start kissing, they start removing clothes, and all of a sudden everyone's happy. So right now the public cannot fine tune Claude models, however. I say the public because we do know that there are fine tuned capabilities for people on the enterprise level.

It's just not in public. By the time you're watching this, it might be public. That's the speed at which this thing moves. Um, but if so, we will always update this course to make sure that you have the latest information to fine tune any AI model that allows it. So if you love Claude though, you can fine tune a cheaper model to write, just like Claude by using outputs from Claude as your synthetic data [00:04:00] to feed into the other model to, to say, Hey, I want you to write like this.

And that's really what fine tunes do best. They change the style, the tone, or the output of the LLM response. So a fine tune can stop some of that nonsense of like little did she know her life would never be the same if you give it a bunch of chapters as examples where you never have those stupid paragraphs at the end of your chapters that are just bad writing, cliched writing.

Those examples are enough to make that fine tune model no longer give you those kinds of trite paragraphs at the end of your chapters that most of us just have to lo off fine tunes only change the how the AI responds. There's a couple of things that they do not do. They do not help the LLM memorize facts.

I'm gonna repeat that it does, and I'm gonna say it over and over and over again. A fine tune of an LLM does not help it memorize facts. , and it, they also don't really work very well with generic prompting. What I mean by that is that when we fine tune a model, we have a very [00:05:00] specific prompting style that we already have adopted as authors that have gotten us to like 70 to 75% there.

And those prompts are what we put into our fine tuned dataset. And so when we use that fine tuned dataset, we wanna use that same style of prompting. So a fine tune, your story information, your scene briefs, your beats, your instructions, however it is that you like to prompt the AI to say, Hey, write this. Write this chapter plus. Consistent prompting with a fine tune is going to equal outputs that need minimal editing.

And I think that's the dream, right? The dream is that the AI sits there inert, it doesn't do anything without us, but we come to it and we're able to say, okay, this is how I write. This is how I want you to write. This is my genius that I'm sharing with you. So that then you have, you are reflecting my genius, not some other rando generalized genius that's, out there.

And so that the outputs I get. They're really close to me. They already are very similar to something. I would already write my fine [00:06:00] tunes on my Jane Austen fan fiction writing. If someone was to run that fine tune, and I've tested this with outputs, I can't tell the difference between that output and like if I wrote it or not.

It's that good. It's that close to my own writing. And I've had my longtime editor even look at the AI writing, the fine tune stuff and she was like, oh girl, I can't tell that that's not you. And she's been editing me for over a decade. Additionally, as more and more authors turn to using AI to keep their readers happy, and it is to keep readers happy, that's why we use ai.

What's the number one thing any reader says? The second you publish a book, where's the next one? That's all they care about. And I get it. I'm a reader too. I don't care how my books get to me. I don't care if they use a ghost writer, if they have two editors, if they had to be kidnapped and stuck into a hotel by their acquisitions editor and forced to finish the book series.

I don't care how the book was written, really. As long as the book is good and it continues the story and I'm happy, it's okay that readers are selfish and they just want the next story. That's totally [00:07:00] okay. But this is your ticket to writing with ai, but not sounding like you write with AI sounding like you.

This is not something that you'll find in a lot of different AI writing groups or AI tutorials and stuff on YouTube. And the reason for that is because most industries have no use case for an AI that writes like them. They never had a use case for anything to sound like a consistent voice.

They're using AI in a, application usually of like processing data, sending form emails, crunching numbers, doing tasks. It's really only creatives or people who are writing nonfiction in like a creative way that they need to have AI sound like them. And that's why this is a use case that is not very often taught, or not very often shared because it's a very niche use case, but it is a use case that we have.

So what will this fine tuned course provide you? Well, the big thing is it's gonna walk you through everything you need to know about making a fine tuned dataset so you can make your own [00:08:00] here at the Future Fiction Academy. We're very, very big on, I'm not gonna fish for you, I will teach you how to fish.

, there's a myriad of reasons for that. We know, we know that a lot of you are like, can't you just make it easy for me and just like do it for me? I could. The problem is, especially when you're talking about a fine tune, if thousands of authors have the same fine tune, it's no different than, the main models writing , all generically the same.

You know, everybody having that phrase, if you write with oh one that says flattery will get you nowhere, because it's one of those catchphrases cliche phrases that was overtrained in the model. And so it has a tendency to just blurt that out all the time. So a fine tune is best. When it's yours, we're gonna provide you lots of great examples and we're gonna show you how to change those examples.

But at the end of the day, a fine tune should be as unique to you as your prompting is, if not more. So we're also going to give you access to, , easy software called Dyno Trainer, which will also be available for everybody else to use as well. Um, so [00:09:00] if you don't need full education on. fine tuned.

You already know how to make it. We're providing that tool, for the community. But it's gonna make it a snap for you to organize your data and format it properly into JSOL. Now, there's two file formats that we'll be talking about a lot in this class, and that's J-S-O-N-J-A-J-S-O-N, json l. What's the difference?

Js OL literally means, , json long and all it does is it allows you to have a js ON data set that allows for paragraph. , it is, Really ugly. It's really hard to do by hand. I mean, it can be done by hand, but you have to make sure you get every single comma and every single quotation mark.

It's, it's just very tedious to do. JSON is the files that the Ner trainer actually takes, and that JSON file is, we have extra, categorizations on there to make it easier for you, the human to manage your data sets. What I mean by that is each JSON file, and we're gonna get deeper into this, so if this is like Elizabeth, you're in the weeds, I have [00:10:00] no idea what you're talking about.

Hold on. A js ON file just takes like a bunch of categories and then like gives a value for it. So like for example, you could have like A-J-S-O-J-S-O category that was like, I color and it would say blip. For like a character profile. In this case we have JSON files that will say, , conversation header, and you could say chapter one with like what pen name you're training for.

Those headers and things like that won't be accepted by the LLMs. You can't take those headers into your J-S-O-N-L file. So we use JSON to make it easier for authors to work with their dataset. Dyno trainer takes that JSO and turns it into J-S-O-N-L that you can take to open AI in Mistral, or it turns it into the CSV file that you can take to Google.

Also the fine tune course will give you nine data sets to get started with so that you can modify them or test them as is. So here's what we're going to cover over the next 10 modules. Intro to fine tuning. You're, you're here. That's this class right now. [00:11:00] Hi. What is the dataset? Your first data set. I'll be back with you for that one, which is going to help you make better outlines.

'cause outlines are a really easy way to see immediately. Oh, that's not great. That is good. So we're going to start learning how to make data sets with data that's very simple to validate and say, Hey, that's good, or no, that's not so good. Consistent outputs, scene briefs. If you like to write with beats, you'll definitely like that module there because it's going to show you how to take that next step into scene briefs.

Right. Like me, Steph, PA Jonas will be here. Um, she, she was one of the first people to ever figure out how to make the fine tunes right, like her, because she's a very distinct voice. That is first person, present tense. I think I, I mean it, it was amazing. The, her, her fine tune is actually the example I used, when I shared it on the forum and people couldn't tell which one was human and which one was ai.

And that was back in December of 2023. New advances in fine tuning. I will be back, because these things are changing every couple of [00:12:00] months. Even in the span of us recording this course material and putting it all together, I highly expect we will suddenly get more models. We can fine tune or new possibilities.

For example, Google only allows you to fine tune right now on one model. I have a feeling that's gonna change in the very near future, conversational AI data set. So this is one of those new advances actually that we'll talk about. The old fine tunes were only prompt. And then what you would say, like your prompt, which was system and user, and then the assistant part is what you were, demonstrating you would like the AI to say to you.

So the first data sets were always just, here's how I wanna prompt you and here's how I want you to respond. Now we have the ability to be like, here's the system prompt. Here's my prompt, uh, here's the response I want from you. Here's my follow-up question. Here's the response I want from you. Here's my follow-up question.

This is a huge advance. That means if you're someone who routinely is using AI for brainstorming or other applications where you're going back and forth, you can actually [00:13:00] control how. Specific the AI is in its response, how much detail it's regularly giving you, you're demonstrating for it, the kinds of answers that you want it to do instead of perhaps some of those answers that you've been getting in the past with just the, the base model, that's what we call it when it's not fine tuned.

The base model, where sometimes it's a good response and sometimes you have to rerun it. Conversational AI dataset can prevent you from having to rerun it. Um, so you don't have to waste those tokens. , conversational AI is gonna be Stacy, by the way.

Genre specific, fine tuned data sets. Steph will be back to show you how to make a data set that is specific for, , one genre, so that then you have this fine tuned data set that's really good at writing that genre. I'll be back to show you. Direct preference optimization, , on how to make 4.0 right, like oh one, which I'll also be using to, to basically it's, it's kind of like combining a lot of different things.

The right, like me and everything like that. DPO is the newest form of data sets, in fine tunes. And that one is you give a prompt and you give a good [00:14:00] example and a bad example. You give a prompt, you give a good example, and you give a bad example. That is also going to be the way that we fine tune the , reasoning models.

So your oh one, your oh three, that's the, the new innovation, , for those kinds of models. They're not available to fine tune to the public yet, but I guarantee that's coming in the next six months. Then finally the final module will be the three of us taking the data sets that we've done in the previous modules and swapping them.

So we're each individually gonna show you how we would take a data set from some other author and change it for our needs, using it as a structure and then making our own. So you'll get three more data sets, , in that 10th one showing you how to modify, and make your dataset yours inside a diner trainer.

Okay, let's learn more about fine tunes. So this video here I'm sharing because this is the recommended viewing. It is older. This is from 2023, November of 2023. It was the very first time that OpenAI publicly started, , trying to do some education [00:15:00] about fine tunes. , the documentation is much better now, y'all, when we started the documentation was like horrible.

We were having to figure things out on the fly. This video, it is long, it's about, I think 45 to 50 minutes, but it's going to talk about rag retrieval, augmented generation. And fine tunes. It's about half and half on each. Now if you're like, but I'm not here to learn rag. I know it's fine. Rag was something that was very much needed before the models had the larger context windows.

This came out at a time where our biggest context window was 16,000 tokens. We are now more than 10 x that on most major models, most major models are taking 128,000 tokens or more. So why is this video still pertinent? Because I think it's actually a really good way to understand. Some of the base jargon technology and understand that when they use jargon in this, they're inventing the jargon on the spot.

Okay? This is a whole new field, so you are not behind. [00:16:00] If you hear something you don't understand. This is like Shakespeare's times where we're all just making words to describe what we're doing because it's a whole new field. Now, the reason I think that this thing is really good, and if you watch nothing else, just jump to like 30 minutes and John Allard will be on there.

So one of the first authors or first writers I should say, he wasn't an author, he was a writer. He was a writer of blog posts. They worked with on making a fine tune, and we still in the FFA laugh about this. He took like 140,000 slack messages to train on. And this is a good.

Learning example. Okay, now, first of all, if you asked us who run the FFA, if we would ever have trained our Slack messages to be how we write, we would've said, no, that's a bad sample set. I don't write in my Slack messages the way I write in my books. That's captain obvious, number one,

but anyway, this person trained on Slack messages thinking, oh, I'll just train on Slack messages and it'll be able to write me a blog post. Well, after he trained on the Slack messages, hundreds of thousands of [00:17:00] them made a data set, put it into the model. He asked this great model, all right, write me a blog post.

And the fine tuned model says, okay, I'll do that tomorrow, because that's how he was always responding in Slack. So it picks up the tone and the style in which you would talk. And then he was like, no, I need you to do this today. And the model was like, okay. And again, because how do we respond in Slack to our bosses?

It's TERs, right? It's very short. It's very like just the information, ma'am. And so basically he spent all this energy to teach at an AI model, how to be like a human, trying to avoid work inside of Slack. And it's quite comical. That's at the 31 minute mark. Now the biggest thing. About this, uh, whole methodologies, I think it still holds true to today.

This is still like a core crux of how I teach, , just writing with AI for authors. Open AI came up with a matrix, but I think it's, it's a fair matrix. , and it's just four quadrants. [00:18:00] So if you're thinking about this lower left hand corner being zero, and I'm gonna zoom in so you guys can see it. A little bit better.

Now this one's let's lit up. The other ones aren't lit up, but I'll tell you what they say. , the optimization flow. So everybody starts here at zero and they kind of learn how to prompt engineer. They learn, you know, better prompting. They learn to say avoid instead of no. They learn to be specific in their prompting and, and they're kind of like happy with their results, but they were like, I think that there's something more.

Well. In order to make the model no more. That's where we have RAG as an opportunity retrieval augmented generation. All that does is take a long piece of text, lots of text, and it starts to chunk it and try to summarize those chunks. Now, if you're trying to think for a second how you would chunk a book and understand that they're not chunking it at the chapter level, they're chunking it like the paragraph level, which works great for nonfiction, but you can imagine for fiction, that's not [00:19:00] really a great.

That's not really a great interval, I would probably chunk at the scene. So we have had advances now in RAG where we can control the rag and we can make the chunks bigger usually a thousand characters or whatever they usually would make it. But rag just takes a lot of information and tries to summarize it down and then inject that summary into the LLM.

You can imagine for fiction sometimes that has mixed results because some summarization is great. But sometimes fiction has metaphor and things like that that the LLM doesn't summarize as well. It's a lot better today, but it definitely wasn't that great of, a year ago. So to help a model no more, you can always inject rag, or you can inject information Today, I think most authors don't really need rag.

Just give it all your stuff and ask for a shortened plan of like what, what the AI needs. The AI can just extract the information from a long prompt and like, make you make a plan. Now, the other way that you can optimize is how the model needs to act. And so beyond prompt engineering, what's to the right here is fine tuning, [00:20:00] fine tuning.

Is what changes that tone so that the AI can write back in that that new voice or that unique voice. And so when an author they master prompt engineering, which is a requirement before you wanna fine tune and then fine tuning, and then they also are able to inject information about their book.

Then they're in this quadrant in the top right, which is all of the above, and AI creators who get to this top box. They're usually getting results that are 90 to 99%. Fine. I've gotten to the point where, I've had AI generate an entire book and I wouldn't have to make change like I read it, things in there, maybe about 10, that I was like, Ugh, I don't really like that.

But it wasn't a games stopper. There was a difference between, it was like the AI in previous years, would just write something that didn't even make sense plot wise. This made complete sense plot wise. I could sit there and read it. Maybe a few changes I would make, but it's completely publishable as is, and that's where we are [00:21:00] today.

Okay, so now the bad news, you do need to have prompt engineering BA before you try to fine tune. This is not a question of us. You know, we know this class is expensive, and it's not as simple as saying, well, I'm paying for this expensive class. I should just be able to skip. It's not about that. I can't teach you how you want to prompt the AI until you practice and you define that for you.

Every author prompts the AI differently, so it's important that you have tested. Played with these base models, open ai, et cetera, and you can voice why you don't like them, why they're not a hundred percent good enough, what are they getting wrong? If it's a question of information, like, I just got my character wrong, then your prompts probably just need to be injected with more information about the character.

If you're like, well, they got the plot right, but this isn't the way I would write it, that's a fine tune. That's a fine tune opportunity. So you need to know the [00:22:00] context, what a context window is, and be able to explain it to someone else. When we work with data sets, we are going to be making examples that have to fit inside a context window.

Context window is, you know how much the AI can take at one in one prompt or one, one chain of conversation. If you don't understand what that is, you're gonna have a hard time making a data set and staying within those limits if you don't understand what those limits are. You need to know the output limits of major foundational models.

Same thing. You'll need to know what system, user and assistant play in the open AI playground interface. Again, our free class on AI writing, at future fiction academy.com covers all of that information so that you can, know, exactly what those roles do. You need to know how to manipulate temperature and top P presence and frequency penalties.

Again, that's all in the free class. It is about being able to play with a fine tune and knowing those fine controls is going to make you more successful. You need to know the limitations and the opportunities, so like they don't think, but you can use that chain of [00:23:00] thought, which is that new conversational style.

You can craft specific prompts that give you decent results. So the results you wanted. That's the foundation of fine tuning. Fine tuning is the foundation of prompts that you know are pretty tried and true, are pretty, reasonable because you. You are going to give a whole bunch of examples, at least 20, minimum of 10, but you're going to give a whole bunch of examples and then you're going to continue to use those kinds of prompts, in the future with the fine tune.

, and you also know that when you don't get a result that you want, you know how to re-roll it, to tweak it, , or try both. That's also a core functionality of being able to fine tune, understand that at the end of the day, this is a roll of dice. It picks a token. It picks the next token. It picks the next word.

It picks the next word, it picks the next word if it chooses kind of like a really obscure word, somewhere in that sequence, you can just be off on a tangent. And then it doesn't mean that you did anything wrong in your prompting or anything's wrong with your fine tune. You just have to click that redo button.

Okay? So if you don't have [00:24:00] those skills, don't worry. You can get them before you try to do a fine tuned data set. Go watch the free classes. Go get the free intro to Raptor Write, play with Raptor Write, which is free. Put $5 on open router so that you can play with it. Take the fine, , the free classes and educate yourself about the basics of AI prompting.

If you don't know that, if you try to go, Elizabeth keeps warning me, but surely that warning is for other people, not me. I can just do this. What's gonna happen is you're going to be miserable. You're going to hate this. You're gonna, you're gonna not be able to tell if you're fine, tunes any good, , or is it even working?

And again, you'll, you'll waste money. Alright? So we're gonna talk about the anatomy of dataset to 3.5 16 k. why am I going all the way back? Nobody's using 3.5 16 K Elizabeth. I know, but this is where it started. And so this being the intro, I'm sharing this information because if this is enough information for you to just go watch, read the documentation for diner [00:25:00] trainer and go use diner trainer.

We're rooting for you at the FFA. We don't want authors to have to be reliant on us. We teach and we do the things that we do so that authors have the technology and have the education if they need it. So if you can do it on your own, we're like saluting you. If you need our assistance or our help, we're here for you.

That's, that's how this works. So, like I said, data sets are at least 10 examples of a prompt. A response. So system and user and response. This is where it started. There's more options now. 10 is the minimum. If you're doing long form, you really want to do as many examples as you can.

You can use synthetic data, which means it's a bunch of AI writing that you validated and you changed, such as using outputs from a higher model, and then you're feeding it into 16 k to make. Six 3.5 16 K, right? Like GPT-4. Oh yeah, you can do that. We did this it says last month 'cause these notes are from, a while ago.

So we did this actually last [00:26:00] year. Feeding GPT-4 outlines into GPT-3 0.5 16 k. The resulting fine tune wrote a longer and more detailed outline than either baseline. So you can also use only human data. I have a fine tune, that's just my human writing that it's trained on. Data sets are written in JSON l formatting for open ai and minstrel, Google's gotta be the odd one out.

They're using CSV, but there's some limitations on that, so I have a feeling they'll eventually change over. So let's take a look at some examples for a data set. So these are, it has a system prompt, a user prompt, and then it has an assistant prompt. In this case, AI did write some of this, but AI doesn't have to write any of it.

What you're doing is you are, you are putting these together. So that you're basically like when you teach a kid how to tie their shoe and you just show 'em how to tie their shoe over and over and over and over again, you're gonna show the LLM how to tie their shoe over and over and over and over again.

So this first one, fabulous. Greg, a writing assistant. You analyze writing and rewrite it to remove cliches. So I would give some examples, rewrite the [00:27:00] following. She goes between a rock and a hard place because she couldn't decide between Brad or Chad. Her stomach was in nuts over the two men. And little did she know her life would never be the same.

I tried put in as many cliches as I. Now I am rewriting this. She found herself torn between two men who loved her, Brad and Chad. The two men could not be more different in how they made her feel, but she didn't want to hurt either one. Somehow the decision felt final. Two final. Now, you could still say like, this is a little bit cliche, Elizabeth, but it's not nearly as bad as this one.

So what I would do is I would do 10 examples of this and I would put it together, and then when I wanted to go use this fine tune, I would use the same system prompt. I'm like, you're a fabulous writing, existent named fabulous, Greg. And then you would analyze, I would give it some writing, and then it would immediately.

Write back stuff that's not cliched. And it would be consistent. Now, some people will be like, Elizabeth, I don't need your fine tune. I can just prompt this and then I'm gonna get good responses. If that's the case, you're right, you don't need a fine tune. But if you're finding yourself still editing even this part of what the AI [00:28:00] puts out, then you can start using a fine tune to give it a bunch of examples.

So it'll start be, it'll start being closer to what you want. Here's a marketing example. You're a marketing genius named marketing. Marketing Mark. Like Marky Mark, you take a blurb of a book and you write three funny short hooks for social media videos. Now here I use an example of Frankenstein. So this data set, this middle one might, you know, be a whole bunch of different book blurbs.

And I'm showing it examples of like what kind of hooks I want. So here's an example. You thought he was dead. Well, no one be his friend. It's alive and very shocking. Um, okay, and then here you're a writer's best friend who always gives good story ideas mixed with humor named loud Lizzie. So this is for someone who likes to brainstorm, but you want your AI to have a bit of personality.

Now you can always just say in the prompts, you're like loud Lizzie, and you make crude jokes and stuff like that. And like I said, that might be sufficient. That might be like, nah, that's good enough. These are for when people want the AI to always be on the ball. [00:29:00] I need story ideas for my genre of a historical fiction book involving a duck.

Sure. Best friend. I'm so glad we're hanging out today. Wanna go to the mall later? Here's that story idea premise. A duck is stuck at the Eiffel Tower in Paris during World War ii. Historical fiction, outline, quacks and courage. A duck's tail from the Eiffel Tower. Now these are silly examples. They're just to make you guys laugh and to see like your opportunity to fine tune a model is.

Sky's the limit. Anything you can think of that you can give the AI examples for, it will impact and start to respond back in that style. If your first one doesn't do very well, try increasing the number of examples you're giving it, make sure you have clear criteria to judge it.

Okay, so there's now two new versions of data sets that you'll also be learning in this class. The first, the examples I just showed you, where you would do that, you know, 10, 20 times. That's the very first style, which is, um, structured fine tuning. So conversational is where you can now do a system. User assistant, user assistant, [00:30:00] user assistant, user assistant.

The assistant is ai, the user is the human role. When I say the assistance, the ai, it doesn't mean the AI has to write those responses. It means you're modeling what you want the AI to write. And then direct preference optimization, DPO, this is where you give the model a prompt and you say, good, bad, good, bad, good, bad.

Both of these methods are less than six months old. So do make a data set of a variety of high quality story components , for your genre summary, et cetera.

Ask or scene briefs. Make the example. AI responses high quality, consistently formatted scenes. So you can do those kinds of data sets. You can do data sets of high quality writing, and the AI will start matching that high quality writing just like it tries to match your writing style. Just when you're doing one shot prompts inside of chat.

You don't wanna run the character list over and over again and expect the LLM to know your character. We saw this with steps. Steph Fine Tune actually had, a dog that was in a couple of the different, examples. And so [00:31:00] then when she was using the fine tune for other things, it started using the dog's name for characters.

This is because when you have a small data set. Every token in that dataset suddenly gets higher attention. And we're gonna talk about this more in just a second with some visuals. The way LLMs work is everything is pieces of words kind of suspended in this multidimensional space called a vector store, and they're connected by their calculated relationships to other words.

So for example, the words dog and cat probably live closer together than say, dog and lion. Dog and cat they're most often written together in passages of writing, more so than dog and lion. But cat and lion have a different relationship than lion and dog, and all of those words, dog, cat, and lion, are gonna have completely different relationships with the word pet.

So if you can imagine these words kind of like grooving and singing out in this space [00:32:00] of like their location is their. Is their mathematical relationship to every other token that's there that they have a connection to. It's bonkers, right? Just like it's mind-boggling. It breaks my brain too. So when you put them in your dataset, you start to shift them around.

And so when a word or token is in a dataset and it gets. A lot of attention. It basically, it's giving extra attention. That's what the core functionality of, an LLM is the paper that started these transformers, which is what the T stands for in GPT says, it's from a paper called All you need is Attention.

So you're giving more attention, attention to that one word. You're gonna see that word in all of your responses a lot. , which is the same thing that humans do. By the way. Humans have crutch words too. My crutch word is just, so when Steph did this first dataset and it had the dog's name a couple of times in there, all of a sudden the AI that was fine tuned was like, oh, she really likes this word for a name.

So the more [00:33:00] varied your dataset can be, the stronger your fine tune's probably going to be. Over fitting, over training with a small un varied data set that makes the LLM unable to perform other tasks because it's stuck on a very narrow, fine tuned behavior. So in step's case, the dog name every once in a while is not the end of the world.

And we see this too in overfitting, in the base models. If I asked you all, give me a list of five words that the AI constantly writes you, you can gimme that list, cacophony, tapestry, whatever the word of the day is, Willow Creek. Yeah, that's because those phrases, those tokens are overfitted in the model.

They're just overtrained, they have too much attention on them. So you can introduce that with a fine tune, that has that. I did it with my fine tune too. I realized that when I kept getting responses back that Elizabeth was biting her lips. She bit her lips, she bit her lips, she bit her lip. I'm like, oh my gosh, she can't bite her lip this many times.

And then I go to my dataset and I find out that like in four of the 20 chapters I gave it, I have Elizabeth Bitter Lip. [00:34:00] So I basically overtrained that one phrase on accident. Okay. So why can't I learn my characters? Like it knows public domain characters. Let's talk about that. The main training of LLM is billions of pieces of data.

It's like run millions of times and weights are applied to those pieces of data. That's like some of the special sauce. It's not just a statistical analysis of how these words are in relation to each other and the training data set, because that would just be like one calculation.

Adding weights and things like that allows you to start tweaking, the attention given to all of these individual tokens so you can kind of skew the data the way you want to. A fine tune does not add additional tokens to the training data. It just rearranges some of them. That's an important concept.

It's like Play-Doh. So even if you think you have a unique name, oh. Oh, well, I'm creating tokens because my character's name is Rean Anon. The way the LLM will handle Rean Anon is it [00:35:00] will break it down into whatever components it does already have tokens for. So anon might be five or six tokens. Actually, I don't even know.

Let's go find out. So I'll use, , GPT-3 0.5 and GPT-4 riken non non Manon. Okay, so you'll see it's seven tokens. It broke it down into seven tokens that it already has. And what's very interesting is let's see if token number five and token number seven are the same number, and they are, you'll see it's 2 6 3.

So we know token number 2, 6 3 is in there three times. It's the token on O-N-O-N-O-N. Now, let's see, I'm gonna copy this. Let's see what just on is. Token number 2, 6 3. Now let's talk about this. The word on has a meaning, okay? But it only has one identification, one id 2, 6, 3. [00:36:00] So this word has probably a relationship with whatever the token is for table. Table 2 0 4 8. So the AI goes, okay, 2 6 3 2 0 4 8.

Those two tokens have this kind of relationship, how often they are. That's how I understand how they go together when we give it a name like this, a right phenomenon, it's constructing it with the token pieces, and then this whole thing kind of becomes like an entity. And this is actually where the new research is heading in AI development.

They're moving away from individual token relationships and moving to clusters of tokens, like phrases and sentences. That's where the newest, advanced research is right now. . So the point is, we didn't add any tokens when we named our character phenomenon, we didn't add any tokens.

These tokens already existed inside of the LLM. If I use a very common name [00:37:00] like Bob. You'll see that Bob gets his own token. Bob is token 33,488. So I wouldn't think it's a very common token, but you'll see that some common names and common things. Let's see what Liz is. Liz is actually two,

the capital L and then the lowercase z. I wonder how big Elizabeth is. Oh, Elizabeth gets its own token now. Tell me how that makes any sense. I can tell you why Elizabeth is a much older name. Bob is not just a name. It's also a, verb like Bob your head. And also Elizabeth all capitalizes a token.

But if I lowercase my name, it's immediately going to break that into, multiple tokens, which is el, which we know is a part of a lot of different things. And then isbe. Now that's for 3.5 and four. If you look at the newer models, we know that they have made more of an effort to tokenize more common words.

You'll see that Elizabeth is one token. Now it's not [00:38:00] two. I bet you Liz is one token. Yeah. Liz is one token as well. But the token ID is 159,000. So that's why these models got bigger. They have more tokens than ever before, so you're not introducing new tokens, which is what the core training is. And so that's why it knows who Mr.

Darcy is because it literally was training that invented Mr. Darcy or Alice in Wonderland or whatever. And even the token Alice with capital A. The LLM is going to interpret that differently. If you're talking about Alice as a housekeeper or Alice as a leader of a school, as a principal versus Alice on an adventure, it's going to go, oh, there's other words in relationship to that that we wanna bring in, like rabbit or, or things like that.

Or Wonderland. When you're talking about Alice on an adventure versus Alice, who's like the principal of a school. So if that was all like, my brain hurts, let's do, let's play with some Play-Doh. So it's like [00:39:00] Play-Doh, there's only so much in the container. You can use a tool and you can make it into different shapes than just your hands.

So when, this is why you wanna make a fine tune your own, the fine tune makes no permanent change the core training of the model. So when you fine tune, chat, GT four oh, you're not changing Core four Oh, it's your personal fine tune. You can't add or subtract Play-Doh that's in the set. You just have the Play-Doh that's in the bucket each time the LLMs get.

Bigger, we get more colors and bigger containers of Play-Doh to play with. We saw that in the tokens. Liz is 159,136. If I go to GPT-3 0.5, it's gotta be broken up into two tokens. The more tokens you break up a, a word into the more likely confusion and hallucinations are gonna happen. Versus if you have a concept that's only one token.

So each time they get bigger, we get more colors and the bigger containers of Play-Doh to play with. Now what's an overfit? When you try to [00:40:00] do too much and you turn all those colors into that terrible brown color, you know what brown color I'm talking about? When you take all the Play-Doh colors and you mix 'em into one color and you can't Unix 'em and it's just broken, you can't get the pretty individual Play-Doh colors anymore.

Or in the case of an LLM, other functionality. So. Why do we use strange, weird names like Outline Mageddon? Well, because if it's tokens, you know, if there's not something for the name of what we're doing when we're giving it these like unique, weird names like Outline Mageddon or B Marketing Marketing Mark, we are helping the OLM kind of create a new kind of relationship.

We have found success doing these kinds of things. You don't have to, but when we put it in the system prompt, it just kind of helps the AI to understand exactly what we're doing and send that signal in the system prompt. So the specific sequence of tokens as part of the rearrangement of other tokens inside the LLM to rearrange them to be closer, farther away from other tokens in the dataset.

So fine tune is we [00:41:00] rearrange the furniture and that rearrangement had to involve those specific token sequences that were repeatedly in the training dataset. This is how you can get, she bit her lips, she bit her lips, she bit her lip or Willow Creek. If you have that too often in the dataset, you have rearranged that furniture to the front of the room, and of course, that's gonna be the first chair that the LLM tries to have you sit in.

Since these main sequences are gibberish, unlikely to be in a ton of other training, we're helping the LLM recognize right away how we want the furniture rearranged of some tokens and not all tokens in the LLM, but we do not literally write new tokens to the LLM. We only rearrange the tokens inside to change their address.

In relation to other tokens in the vector field. So last little visual aid. If you're imagining this vector field as a bunch of metal bits and screws that were living on a place, on a table, and you drop a magnet, like a fine tune, you'll see that some of the magnets and screws sucked right to the magnet.

But other ones that the dataset, or in this case the magnet, had nothing to do with, they're gonna stay where they are. So if they're not affected by the magnet, [00:42:00] then they won't move. This is literally what a fine tune does. So a fine tune changes how the existing tokens in the training dataset relate to one another in small fine tune amounts.

You got through it. Congratulations. Hopefully there was enough. Education about what it is and what we're doing and why. This is going to allow the AI to talk like you. So some questions that will help you think, to how a fine tune will help you. So I want you to start thinking of these questions.

Take some time to jot them down before you go through the rest of the modules, because this is gonna help you understand this and get proficient at fine tunes a lot faster. What kind of consistent response style do you want from the LLM? One of the biggest things I hear is I wanted to write longer, or I only wanted to write 400 words.

Length is a really big thing for authors. Having a fine tuned data set that consistently shows an output of the length that you want the LLM to do will help it at least hit that length. You may not like all the words, but it'll at least hit that length. What kinds of prompts, [00:43:00] system and user will you be routinely using in hopes of getting this response?

If you're someone who prompts on the fly. You're a pantser or whatever, take a look at your old conversation. See if you're, you're constantly saying, can you expand that? Can you make that longer? You can add that to the first prompt in a conversation chain and then like give that second answer that you did in the prompt chain as the first answer in your fine tune, so that way the LLM skips doing that first bad answer and just kind of skips to that good second answer.

Also, why do, why does the current baseline prompt fall flat? Have you exhausted tricks of prompt engineering, like changing your hyper parameters? That's that temperature and stuff. Styles of prompting, such as chain of thought. In other words, have you moved beyond just basic prompting of like, write me a book.

Are you actually giving like very detailed prompts that you know, sometimes work for you, sometimes don't, and you just want it to be consistent? And then the biggest one of all is like, how will the fine tune be a win? And what I mean by this is for example, I was very happy when I got 3.5 16 K to [00:44:00] consistently write me 2000 words.

That was a win. I didn't even care if they were wonderful words or not, but it actually, my fine tune of 3.5 16 k immediately brought dialogue in. 'cause that's not how 3.5 16 K used to write. Used to always not write dialogue. You'd have to put that in on a second response. And it actually wrote me 2000 words, in a response.

Like it would actually write 1500 to 2000 in a in a go. Those were the wind conditions. So think about your wind conditions for your fine tune. That way if you get stuck or you have questions, you can always email us at Dean at future fiction academy.com or ask a question in our discord. Because if you can understand your wind condition.

Then we can help you troubleshoot and figure out what's the best data set for you or the best model or something like that. Or if a fine tune's not gonna be able to solve your problem. But if you can't clearly communicate what's wrong and what you want in a fine tune, you are never going to be satisfied with your fine tune.

You're just gonna be like, well, it kind of feels like it's better, but maybe not. So make sure you have that wind condition.

Okay. [00:45:00] I know that was a lot of information and we took some field trips. You now know what a fine tune is, okay? You have the same knowledge of fine tuning that I had to go procure for myself in 2023, and so did all the other founders of the FFA. So you're armed and you're dangerous.

If you think you can just. From that Learn Dyno trainer is available. , if you want more guidance, then I recommend you purchase the fine tune class because we're gonna teach it all to you. If you've got a little glimmer of that, like, oh wait, I understand this. Oh, I understand this. Okay, I wanna make the AI right, like me, this is the class for you.

And like I said, we're going to give you nine data sets. I really do think that this is the future. We have automated books now with ai. We've gotten there. We have the Easy Peasy book machine. Everybody has that from last year, where you can just say, make a plan. Right. Edit, make a final pairing that with a fine tune is the secret sauce that makes it so that what the AI writes out is yours.

It's you. It's people not even reading that and going, I [00:46:00] think this is ai. No, I think that's EAW. Um, so fine tunes are definitely a big component of the future. It's also gonna help future proof your career because if you are writing with AI with a fine tune, your writing is not going to look like ai. The common phrases that people are going to spot of like, I can already tell you, oh, one preview.

Constantly says flattery will get you nowhere. Every single book I've run with it in different genres, it has worked that line in Willow Creek because there was a bunch of Willow Creeks inside and now we see a lot of books in, the Amazon store with Willow Creek. It's like exponentially gone up. It's because the AI rehashes a lot of these tokens that have too high attention.

If you're tired of seeing the names Elena and Elara and, oh gosh, Jake, oh gosh, 3.5 16 K. Every male lead was a Jake. These are the things that a fine tune can cure, so, good luck with the rest of the class. If you need any help, just holler.[00:47:00]