How to Fine-tune GPT-3 in less than 3 minutes.

hello everyone today I would like to show you  how you can create your own fine-tuned models   without writing a single line of code so yeah no  need to install python because with entry point   which is a platform that me and Mark have been  developing for the past four months you'll be able   to do that without writing any line of code now  for the purposes of this video we'll be creating   a classifier type of AI model specifically  we'll be classifying these keywords. Now   what do I mean by that? so here I have a list  of 4000 keywords that I got from a Google ads   campaign now I have to analyze these keywords  and to do that I have to categorize them right   because keywords by themselves they don't really  mean nothing to me so I have come up with four   different types of categories completely custom  based on my logic and now I can analyze them   right so I can have the classified I can now see  how much people are converting right and I can   see that on use case type of keywords they're  not really converting that well while on tool   descriptive type of keywords they're converting  really well now the problem is I've only used 120   keywords for this analysis which is of course  I have four thousand keywords it's not really   that accurate right but doing this classifying  this manually would take me an entire day and it   would be mind-numbingly boring right so let's buy  into an AI model to do this for us now to start I   will simply take the examples that I already have  right and paste them in a empty sheet and just   download this as CSV there we go that's all that  I need now I can go to entry point and create a   new project let's go with PPC classifier and  we will make it a classifier type of project   you can also create generators in entry point  like generator type of models that produce some   text like a description of something let's say an  intro for an article a story whatever it doesn't   really matter but creating AI generators is a bit  harder because the output is not black and white   right with classifiers it's either the output is  either correct or it's incorrect so it's very easy   to create a classifier type model so we're going  to go with that now we can create a blank project   or a simple text prompt with category completion  I'll explain what this means in a moment or you   can even use a tool that we have built-in entry  point that allows you to just write the URL and   entry point will scrape the data from the website  and then you can classify the data and use it   however you want right so but for now we're just  going to go with blank project click Place create   now to start you can create Fields And format  them manually which I'll show you how to do in   a second or you can simply upload your CSV just  like that I have uploaded the CSV now we have   to define the columns what are the columns well  so first I have the category column this should   be the completion right and it should be one of  the four predefined options and the Search terms   of the keywords they are just like a text type  of a prompt basically I'm going to go continue   and I can now allocate a certain percentage  of my examples training data essentially   to be validation data and what validation does  is when once you send the data to open AI for   fine tuning the validations are actually not  going to be used for fine tuning they're going   to be used to evaluate the health of the  model so the accuracy of the model right   in my example I don't really have much examples  only 120 so I'm going to go with zero validation   and yeah so the important started now  we're just going to wait two seconds   there we go it's done so that was fast and here  are all my data points now if I had a team and   I had like a ton of examples I could divide them  into entry point and they could help me edit and   label the data right here make sure that we have  really awesome high quality data for fine tuning   that's not the case with this simple classifier  it's very simple plus I have already prepared   high quality data in the CSV so I don't have  to do that here but that's an option and it's   a super useful option I can also change training  to validation right so I can change that train   data points to validation data points later and  so on now well we're actually ready to start the   python we simply have to pick a platform of our  choice so in my case it's going to be open AI   I could also use AI 21 which is another large  language model provider and in the future we're   going to have all of them but for now just open  Ai and I simply have to enter my API key and to   find that I have to go under my open AI account  and just go under view API keys and just create   a new secret key just copy that and then paste it  into endpoint save and just go back to our project   now let's go to our phytons click plus and  we can pick a model in my case I'm going to   pick DaVinci but keep in mind that cutie is also a  wonderful model for creating classifier type of AI   models but in this case I really I don't have many  examples already I'm just going to go with DaVinci   so I can squeeze everything that I have out of the  trading data that I basically provided right and   entry point is going to also calculate the number  of tokens that I'm going to send to fine tuning   and it's going to calculate the cost right  now I have some Advanced options here which   is setting the number of epochs and prompt loss  weights and so on we can leave this as it is   and now we can simply press start and yeah so the  finaling job has started so now entry point is   sending all the examples to open Ai and now we're  simply waiting for them to find the model and   send it back to us essentially right now this fine  tuning job is fine tuning model is named tealpanda   and so entry point will always name your models  with a color and an animal and we'll do we do that   just so that it's easy to navigate once you have  like five ten different models instead of looking   at some complex strings we have color colorful  animals okay so we're going to wait for this to   finish and in the meantime I'm going to show you  templates and Fields because this is really where   entry point shines they might be daunting to get  into and by all means you can just do your models   the way I showed you right here but templates and  Fields really unlike additional possibilities with   entry points so let me show you an example of how  you would use that and when you would use that   okay so I'm going to create another test project  let's go with uh let's see story generator test I   don't know this is going to be a generator just  a blank project and here I'm going to manually   set the fields and the templates okay there we  go so this is going to be the text now if you   were to just use a normal prompt you would have to  actually set up the fields like this this will be   the prompt Target is the prompt there we go and  now we also need a completion Okay so just like   that now the reason why I made it like this now  I can go under templates and the prompt which is   going to be one big variable and the completion  as well so if I go under examples and I create a   new prompt completion data point let's say that  write a story about a female Spider Queen named   right and then we can have the story here  okay so that would be an example of creating   a generator type of air model and we will  go save the next and then we would create   another like prompt write a story about a  orc a male orc named John then John is okay   so that's how we would add the data points now in  this project we can see that I will always have in   my prompts like I can you know predict this I'm  going to have a gender a race and a name okay   so doing this manually every time opens up the  possibility for mistakes and it's really time   consuming so instead what I can do I can go back  to my fields and I'm going to I'm going to delete   actually I'm going to leave the  completion I'm going to delete the prompt   going to create a new text field  which will be named name prompt   I'm going to create a new text field  which will be the race and another one gender there we go okay and I can go under  templates this doesn't exist anymore and I   can go right a story about let's go with this  named name here we go save click on examples   and so the name here is John Grace is I don't know human what is it human  and yeah right now I can just save this and the prompt adopted right so it fits the  template so I can just enter the variables   and all the examples will adapt so this was  icky either Queen female there we go okay now   once I have the prompts Style style like this  now I can go back here and change the template   I know draft story you can save this and all  examples will now change as well right so I   won't have to manually I won't have to manually  change any anything later on the project right   so that's really useful now what about what  if you have a ton of data points and they're   incomplete what if you don't always have a name  for example well in that case we actually build in   handlebar logic inside these templates now this  is really really Advanced use but with Handler   bar logic you can actually create if statements  let me show you what I mean so here we can go with   if name is not no well then I should  simply pick a name since we have the   name but if that's not the case else let's go  with uh let's go with Kevin let's close the if there you go save we go back to our examples  and now we can create a data point like human   male this is a story about there we go Kevin right so it otherwise it would have been  empty the problem would be incomplete but now   we gave it the default value it does take some  energy to learn this but once you do I mean you   can just create and manage such huge data sets  of examples and it's going to be beautiful yeah   let's get back to our classifier I'm actually  going to go with another one that I fine-tuned   a while ago on the same data set just because the  other one is going to take an hour and a half so   just want to show you this right now okay so this  is the final this one is named gray loose not   steel Panda anymore but it's ready and now I can  just click this button and it's going to open it   up in playground it's right over here now you can  test if this works so let me just pick a random   keyword like this one paste it in here and add the  stop sequence a stop sequence here as well submit   which will probably lower the temperature  to zero it's a tool descriptor that's   right that's perfect classification  let's go one more member support yeah the competitor brand so it works now  I also have playground in entry point so I   can quite literally do the same thing so I  simply select my model here run completion   and I can see the well the correct completion here  in the future we're also going to have an option   for you to upload or download the completions or  even change them to them and once you change them   you can actually save them to examples for future  fine tuning that's for classified type projects   for generated type of projects we're actually  going to have an AI generate the examples run   the completion switcher model and then provide you  a few completions like two to three which you can   then pick whichever you like the most which is an  extremely efficient way or to find on your model   even more right so yeah that's coming now let's  use our model in our actual project so I'm going   to go back to my keyword list and I'm going to  go here and write completions or categories yeah   categories there we go and I'm going  to run this function that I made which   calls that model and there we go  completes it just like that easy join PDFs file jump PDF files online for free  that's a use case tool descriptor free online   pdf file so that's right Windows Studio that's  a competitor there we go right and we have our   python model that can classify these keywords  so that's it and if you also want to integrate   your newly fine-tuned AI model into Google  Sheets just like I have over here with a new   custom function well I have a tutorial on  that for our newsletter subscribers or for   our Discord members so make sure you either  join our email list or join our Discord and   I'll send you over that video and if  you would like to use entry point to   fine-tune your AI models well then there's  a sign up link right below it's currently   in open Beta which means it's still free it's  not going to be the case forever so make sure

Channel: Miha Cacic of ENTRY POINT

Views: 1,847

Keywords: fine-tuning, fine tuning, AI, openAI, How to fine-tune GPT3, gpt-3, gpt-3 fine tuning, fine tuned gpt3 model

Id: fOpFhRclaEk

Length: 14min 34sec (874 seconds)

Published: Wed Mar 29 2023