Running the Fine Tune Model

This is our base test and we got one case right at least. Case one and three was wrong. So what we're going to do when our fine tuning is done here now we're going to try to run the same problem with our fine tune model and see if it can solve it I have low expectations but yeah let's give it a go okay so our training is finished so I wanted to kind of just grab all this information here right all the steps let's copy that let's go to chat gbt and ask how does this llm training run look because I don't know anything about training models right I know just the basics so it seems this training run seems to be progressing quite well uh okay so let me just let this run and see what it says you can see the training loss uh consistently decreasing from first step and the last report step is a good indicator of effective training learning this suggests that the model is learning from the training samples and improving prediction over time okay so that's interesting if we go all the way down overall this looks on track just make sure validate model's performance and hold outet okay yeah I think it's fine so now let's actually test this versus our base model okay so if you continue down here we have something called inference here we can actually run our model so I went ahead I uh gave instructions so solve the lead code problem and I fed in our problem right the same as we had in our base test model here solve the problem in Python syntax and we leave the output blank and yeah let's try it okay so let's see what happens here now okay so it's still running here's a step bystep solution applied to problem three uh okay yeah let's let this run and then we take a look at the final solution okay so you can see solution to problem five uh okay let's copy it I just have to wait a bit okay so I just stopped it so let's try this now so I'm going to copy this solution let's head back over to Elite code and let's try it okay so it didn't get it but I thought it was a good try uh we still got the same cases uh right uh I might try it to run it one more time and see if we get something else but you can see it kind of followed our uh instructions so we have the response right so here is the step-bystep solution we have the solution and an explanation so let me just try to run it one more time and see if we get another solution okay so we have another solution example here so let's try this one so so I just H added some temperature here so I set the temperature to 0 z uh let's try this solution oops okay so let's try it okay so that didn't work either so what I went ahead I did uh I found a new problem that is also easy so let's try it on that just for fun okay so let's run this and yeah just wait until we get the answer back okay okay so we have a solution here that looks very complex right let's copy that let's go back here let paste it in and let's run it oh no it didn't get it but yeah uh it's not perfect far from it but what I like is that uh if we take a look at let's put on this you can see the response is kind of in the structure I wanted we get the stepbystep uh solution part it wants to think through that then it wants to create the solution then it wants to explain it so even though it didn't work I'm pretty happy how uh the fine tuning turned out okay so maybe we didn't get the result we wanted but uh what I wanted out from this video was not kind of actually how our fine tune turned out that is can be many things uh and remember we were running the F three a week proba maybe if we did this on llama 3 it could be better uh but what I wanted to focus on in this video is kind of how you can create data sets so that was kind of the idea behind this videos and this is not coming from an expert so don't take this for the final solution uh of anything this is just the way I have been playing around with this and as you saw from probably the the fine tune result it was kind of following our instructions pretty good so I'm going to put up this code if you want to try it out to kind of create your own yourown data sets uh it's very easy to go from text to to Json if that's something you're interested in uh so I'm going to leave this up just follow the link in the description

"WEBVTTKind: captionsLanguage: entoday I want to talk a bit about how I create my personal data sets for fine-tuning so just a warning I'm no expert at this but I just want to share what has been working for me so in this video I'm going to show you how I gather quality data how I structure it and how I use it for fine tuning models like f tree so let's just get started okay so as you probably saw from the title we're going to focus on lead code today so I want to create a data set that kind of solves these lead code problems so let's start with a contest that has been recently ended so we don't get any contamination in the data set right so let's click on this 395 here so what I want to do is find the problems okay so let's go down here so hopefully this is an easy problem I think we want to start with that uh on um on F3 yeah we have an easy problem here okay so what I want to do is uh uh okay we can turn off this now uh okay so you can see a bit better so what I like to do is the first thing I want to start with is actually collecting the problem right so here we kind of have our problem I want to grab every single thing here the examples and the problem and the constraints so let's just start with that okay so let's just copy this right and then I'm going to head over to my notepad uh I create something just called Dash problem one so let's paste in our problem here okay that looks fine uh I also want to grab kind of the Python 3 syntax here and I want to grab this part here so we kind of have how we should set this up let's go back here and I like to paste that in here at the end right okay so the reason I have set this up in this kind of structure is because I have a code here that is actually going to parse this text file into Json so here I kind of want the lines to start with Dash problem and I wanted to start with Dash solution to problem so let me grab that too and go down here and let's just create a new line here and let call it solution to problem one okay and save that so we kind of have our problem here now and now we need the solution to the problem so let me show you how I do that so what I like to do is to use a powerful llm so let's this could be claw three Opus this could be gd4 and and I like to use those to kind of solve the problem and get the structure and the reasoning behind this so if you take a look at our system prompt here in CLA 3 Opus uh you are a professional software Dev with expertise in Python and solving lead code problems always explain in a stepbystep way Chain of Thought in the text before writing the solution to the problem in code okay and then I have a bunch of examples of problems I have solved before so this is basically the same we have a problem and and we have a solution here's a step-by-step approach how we should solve it and we have a solution to problem one right so you can see this and then we have a second problem so I gave this like a many shot examples to solve these problems so if we scroll all the way down here we have our problem 10 that we has haven't filled out yet so here is where we are going to paste in our problem right so we paste in yeah the problem we just copied from lead code we have the class solution and then I go back here to lead code uh I go to Solutions I select Python 3 I found yeah here's a great solution so I copy that and I paste in the solution here I just go solution solution and write this and then I go first create a step by step plan how to solve the problem 10 using pathon Code then write the solution in Python code Python 3 syntax okay so let's run this okay so you can see here we have a step-by-step plan how to solve problem 10 we get five steps right okay that's great we have a solution in Python code here we have the suggested solution so you can see this is not exactly what we got here so we got a bit of a different variant I think that's interesting and then we have this explanation so what we want to do now is transfer all of this into our text file right but of course we have to check if this was correct so let's copy the code let's head back to lead code let's paste it in here and let's uh yeah we can just go here and run it all the test cases hopefully this works uh accepted accepted accepted perfect so yeah we have our solution we have our step by-step plan how to do it and our explanation so I'm just going to click Text here I'm going to grab the stepbystep plan right and down here and let's head back to our text file now okay so you can see we have solution to problem one so I just want to paste this in here step byep plan to solve problem one okay solution in Python code so we kind of need the code right so let's go back here and let's copy the solution let's paste it in here great space and let's grab the explanation too right so let's grab the full explanation of this and I think this is very valuable data right because we kind of combined a problem with kind of the step bystep plan how to solve it we get the solution and we get the explanation I think this is very good data so now this is kind of ready so that means we are actually ready now to kind of turn this or convert this into our first Json example okay if we go back to our code now you can see that three fields we want in our Json file we want instruction input and output so for instruction I kind of hardcoded so this is solve the following lead code problem so that is our instruction and uh input is going to be of course uh the problem right and output is going to be solution to problem so let's see how this works now if we go to our um let's do python all right and let's do 1 million and run it so you can see problems pared written to lead code problems T Json okay so let's go open that and here you can see here is our format now oops so we have instruction solve the following lead code problem and you can see here is our input so you can see it starts with you are given two arrays of equal length if we go to our data sets on top here you can see okay you're given to okay that's good and uh the output starts with stepbystep plan to solve problem one if you go back to our data set and scroll down you can see solution to problem one is actually a step bystep plan to solve problem one did we get all the text so let's go all the way to the end here and you can see we end with extra space I don't know if you saw that but uh if we go down here into our text file you can see the final here is extra space so that means that we got everything and yeah our first example here is kind of done and if we just start over again now and we do like let's say uh Dash problem two right and then we should be able to create an example so let me just fill out a problem two and I'm going to show you how this works okay so I filled it out so we have a problem two here now we have uh I think I fill in like a solution to problem two okay great and now let's run it again okay problems passed and written so let's reload this and yeah you can see now we have a second example so you can see this is not the same because we don't have the same start here this problem yeah solves okay so let me just double check this this probably involves finding okay uh I just want to check that that looks a bit strange uh no it's correct okay perfect so that means we have our second example so what you wanted to do then was just continue like this uh I have a data set already that has 13 problems so this is what we are going to use for our fine tuning so you can see it's pretty big uh it's like this it has all reasoning so I did 13 examples and let's turn that into a data set so I call that uh I think I call it lead code data set right okay so let's just open that and let's create a new Json right like four okay and let's run this uh Le code problems three was it uh supposed to be four but that's fine let me open that and here you can see here is our data set now so here we have 13 examples with reasoning with the solution and with the problem and with the correct instructions so this means uh we are ready to actually run this uh with some fine tuning on 5 3 but before we do that there's one more thing I want to do I want to head over to hugging face so you can see I have my profile here so what I want to do is I want to click here new data set and I want to give it a name right okay and I want to pick public license I don't care uh create that data set and what I want to do next files and versions okay I can add a file so I'm just going to upload our Json file here right and save it and then we kind of end up with uh uh let me show you my data sets and then we end up with something like this so here you can see here we have all the examples so that was the input part right you can see all the examples we have in input here we have the instruction so this is our instruction we gave and we have the output and you can kind of see how many characters this is percentages and stuff I haven't really looked too much into it uh but it's but is good about this is that we can actually now use this in fine tuning over on unslot so that is what I'm going to show you next so if we head over to the unslot GitHub I'm going to leave a link in the description for this we can just scroll down here and we can start using these notebooks so this is totally free right so I'm just going to click on this F3 notebook here okay and here you can see we can start uh with some fine tuning on F three so it's very good explanation here so I'm just going to walk you through this let me just take off this so we can see a bit better right okay so let's just start by uh starting this first part here okay so you can see now we are uh connecting to a GPU so this is going to initialize this this is going to take some time it's not that long but it's just going to start this so what we want to focus on is actually our uh uh our data set here right so let me swim in a bit so here you can see uh we have our model name so this is going to be the F3 mini 4K instruct that we is what we want to train on uh let's scroll down to our data prep here so here you now you can kind of see we are following this pattern here so remember we have instructions input and response and we have some kind of instruction here that is actually going to use all of these examples so you remember our data set right input instruction and output put it's the same as you can see here and remember we need this EOS token here end of sentence token so that's good and now we kind of want to load the data set here so let me show you how we do this so since my um my lead code data set is public I can just copy this URL here so let me show you that so I'm just going to go back here I'm going to copy this and let's go back here and just paste it in here right okay and that should be it now we should find our data set so you can see this is completed now let's fire up this so every time you see this check mark here this means that we can move on you can actually see what's happening here so uh I'm just going to let this run a bit and I'm going to show you when we are actually starting the training okay so you can see now uh we kind of put in our data set here that worked well we downloaded this data you can see 13 examples remember we had 13 different examples so let's just continue down here train the model and you can see it's uploading this we can press I'm just going to press every single play here so we have the Tesla T4 and let's start here now so this is going to start our training and here we can get some more informations you can see we have number of examples 13 number of epo 60 and this is going to run uh usually takes about 20 many minutes I think uh it don't take that so you can just let this run in the background so let's let this run for a while and I'm going to take you back okay so this has been running for a few minutes now you can see uh we have about 16 minutes in total we are running for 5 minutes and you can see the training loss here now is starting to drop we started around one right we are down to 75 so we let going to let this continue of course all the way to 60 steps and I'm going to take you back there and then we're going to try it uh but while we wait for this uh I went ahead I found a new lead code problem from the uh earlier contest so this is going to be count the number of special characters so what I went ahead I copied this I created a problem here so we have the lead code problem I paste it in this solve the lead code problem in Python 3 syntax so now we actually going to do like a base test so uh what I'm going to do is copy this right and let me me do uh AMA run 53 okay and let's paste in this problem and let's see if it can solve this okay that looks a bit strange uh but let's try it so let me copy this let's go back here and paste it in and let's try it yeah there's some errors here so let me fix that and let's try again okay so there was just some syntax error so let's try it again uh this should be okay now so let's run it so remember this is fire three this is kind of our base test and yeah we got one case right at least and case one and three was wrong okay so what we're going to do when our fine tuning is done here now we're going to try to run the same problem with our fine tune model and see if it can solve it I have low expectations but yeah let's give it a go okay so our training is finished so I wanted to kind of just grab all this information here right all the steps let's copy that let's go to chat gbt and ask how does this llm training run look because I don't know anything about training models right I know just the basics so it seems this training run seems to be progressing quite well uh okay so let me just let this run and see what it says you can see the training loss uh consistently decreasing from first step and the last report step is a good indicator of effective training learning this suggests that the model is learning from the training samples and improving prediction over time okay so that's interesting if we go all the way down overall this looks on track just make sure validate model's performance and hold outet okay yeah I think it's fine so now let's actually test this versus our base model okay so if you continue down here we have something called inference here we can actually run our model so I went ahead I uh gave instructions so solve the lead code problem and I fed in our problem right the same as we had in our base test model here solve the problem in Python syntax and we leave the output blank and yeah let's try it okay so let's see what happens here now okay so it's still running here's a step bystep solution applied to problem three uh okay yeah let's let this run and then we take a look at the final solution okay so you can see solution to problem five uh okay let's copy it I just have to wait a bit okay so I just stopped it so let's try this now so I'm going to copy this solution let's head back over to Elite code and let's try it okay so it didn't get it but I thought it was a good try uh we still got the same cases uh right uh I might try it to run it one more time and see if we get something else but you can see it kind of followed our uh instructions so we have the response right so here is the step-by-step solution we have the solution and an explanation so let me just try to run it one more time and see if we get another solution okay so we have another solution example here so let's try this one so so I just H added some temperature here so I set the temperature to 0 z uh let's try this solution oops okay so let's try it okay so that didn't work either so what I went ahead I did uh I found a new problem that is also easy so let's try it on that just for fun okay so let's run this and yeah just wait until we get the answer back okay okay so we have a solution here that looks very complex right let's copy that let's go back here let paste it in and let's run it oh no it didn't get it but yeah uh it's not perfect far from it but what I like is that uh if we take a look at let's put on this you can see the response is kind of in the structure I wanted we get the stepbystep uh solution part it wants to think through that then it wants to create the solution then it wants to explain it so even though it didn't work I'm pretty happy how uh the fine tuning turned out okay so maybe we didn't get the result we wanted but uh what I wanted out from this video was not kind of actually how our fine tune turned out that is can be many things uh and remember we were running the F three a week proba maybe if we did this on llama 3 it could be better uh but what I wanted to focus on in this video is kind of how you can create data sets so that was kind of the idea behind this videos and this is not coming from an expert so don't take this for the final solution uh of anything this is just the way I have been playing around with this and as you saw from probably the the fine tune result it was kind of following our instructions pretty good so I'm going to put up this code if you want to try it out to kind of create your own yourown data sets uh it's very easy to go from text to to Json if that's something you're interested in uh so I'm going to leave this up just follow the link in the description So yeah thank you for tuning in and I'll see you again on Wednesdaytoday I want to talk a bit about how I create my personal data sets for fine-tuning so just a warning I'm no expert at this but I just want to share what has been working for me so in this video I'm going to show you how I gather quality data how I structure it and how I use it for fine tuning models like f tree so let's just get started okay so as you probably saw from the title we're going to focus on lead code today so I want to create a data set that kind of solves these lead code problems so let's start with a contest that has been recently ended so we don't get any contamination in the data set right so let's click on this 395 here so what I want to do is find the problems okay so let's go down here so hopefully this is an easy problem I think we want to start with that uh on um on F3 yeah we have an easy problem here okay so what I want to do is uh uh okay we can turn off this now uh okay so you can see a bit better so what I like to do is the first thing I want to start with is actually collecting the problem right so here we kind of have our problem I want to grab every single thing here the examples and the problem and the constraints so let's just start with that okay so let's just copy this right and then I'm going to head over to my notepad uh I create something just called Dash problem one so let's paste in our problem here okay that looks fine uh I also want to grab kind of the Python 3 syntax here and I want to grab this part here so we kind of have how we should set this up let's go back here and I like to paste that in here at the end right okay so the reason I have set this up in this kind of structure is because I have a code here that is actually going to parse this text file into Json so here I kind of want the lines to start with Dash problem and I wanted to start with Dash solution to problem so let me grab that too and go down here and let's just create a new line here and let call it solution to problem one okay and save that so we kind of have our problem here now and now we need the solution to the problem so let me show you how I do that so what I like to do is to use a powerful llm so let's this could be claw three Opus this could be gd4 and and I like to use those to kind of solve the problem and get the structure and the reasoning behind this so if you take a look at our system prompt here in CLA 3 Opus uh you are a professional software Dev with expertise in Python and solving lead code problems always explain in a stepbystep way Chain of Thought in the text before writing the solution to the problem in code okay and then I have a bunch of examples of problems I have solved before so this is basically the same we have a problem and and we have a solution here's a step-by-step approach how we should solve it and we have a solution to problem one right so you can see this and then we have a second problem so I gave this like a many shot examples to solve these problems so if we scroll all the way down here we have our problem 10 that we has haven't filled out yet so here is where we are going to paste in our problem right so we paste in yeah the problem we just copied from lead code we have the class solution and then I go back here to lead code uh I go to Solutions I select Python 3 I found yeah here's a great solution so I copy that and I paste in the solution here I just go solution solution and write this and then I go first create a step by step plan how to solve the problem 10 using pathon Code then write the solution in Python code Python 3 syntax okay so let's run this okay so you can see here we have a step-by-step plan how to solve problem 10 we get five steps right okay that's great we have a solution in Python code here we have the suggested solution so you can see this is not exactly what we got here so we got a bit of a different variant I think that's interesting and then we have this explanation so what we want to do now is transfer all of this into our text file right but of course we have to check if this was correct so let's copy the code let's head back to lead code let's paste it in here and let's uh yeah we can just go here and run it all the test cases hopefully this works uh accepted accepted accepted perfect so yeah we have our solution we have our step by-step plan how to do it and our explanation so I'm just going to click Text here I'm going to grab the stepbystep plan right and down here and let's head back to our text file now okay so you can see we have solution to problem one so I just want to paste this in here step byep plan to solve problem one okay solution in Python code so we kind of need the code right so let's go back here and let's copy the solution let's paste it in here great space and let's grab the explanation too right so let's grab the full explanation of this and I think this is very valuable data right because we kind of combined a problem with kind of the step bystep plan how to solve it we get the solution and we get the explanation I think this is very good data so now this is kind of ready so that means we are actually ready now to kind of turn this or convert this into our first Json example okay if we go back to our code now you can see that three fields we want in our Json file we want instruction input and output so for instruction I kind of hardcoded so this is solve the following lead code problem so that is our instruction and uh input is going to be of course uh the problem right and output is going to be solution to problem so let's see how this works now if we go to our um let's do python all right and let's do 1 million and run it so you can see problems pared written to lead code problems T Json okay so let's go open that and here you can see here is our format now oops so we have instruction solve the following lead code problem and you can see here is our input so you can see it starts with you are given two arrays of equal length if we go to our data sets on top here you can see okay you're given to okay that's good and uh the output starts with stepbystep plan to solve problem one if you go back to our data set and scroll down you can see solution to problem one is actually a step bystep plan to solve problem one did we get all the text so let's go all the way to the end here and you can see we end with extra space I don't know if you saw that but uh if we go down here into our text file you can see the final here is extra space so that means that we got everything and yeah our first example here is kind of done and if we just start over again now and we do like let's say uh Dash problem two right and then we should be able to create an example so let me just fill out a problem two and I'm going to show you how this works okay so I filled it out so we have a problem two here now we have uh I think I fill in like a solution to problem two okay great and now let's run it again okay problems passed and written so let's reload this and yeah you can see now we have a second example so you can see this is not the same because we don't have the same start here this problem yeah solves okay so let me just double check this this probably involves finding okay uh I just want to check that that looks a bit strange uh no it's correct okay perfect so that means we have our second example so what you wanted to do then was just continue like this uh I have a data set already that has 13 problems so this is what we are going to use for our fine tuning so you can see it's pretty big uh it's like this it has all reasoning so I did 13 examples and let's turn that into a data set so I call that uh I think I call it lead code data set right okay so let's just open that and let's create a new Json right like four okay and let's run this uh Le code problems three was it uh supposed to be four but that's fine let me open that and here you can see here is our data set now so here we have 13 examples with reasoning with the solution and with the problem and with the correct instructions so this means uh we are ready to actually run this uh with some fine tuning on 5 3 but before we do that there's one more thing I want to do I want to head over to hugging face so you can see I have my profile here so what I want to do is I want to click here new data set and I want to give it a name right okay and I want to pick public license I don't care uh create that data set and what I want to do next files and versions okay I can add a file so I'm just going to upload our Json file here right and save it and then we kind of end up with uh uh let me show you my data sets and then we end up with something like this so here you can see here we have all the examples so that was the input part right you can see all the examples we have in input here we have the instruction so this is our instruction we gave and we have the output and you can kind of see how many characters this is percentages and stuff I haven't really looked too much into it uh but it's but is good about this is that we can actually now use this in fine tuning over on unslot so that is what I'm going to show you next so if we head over to the unslot GitHub I'm going to leave a link in the description for this we can just scroll down here and we can start using these notebooks so this is totally free right so I'm just going to click on this F3 notebook here okay and here you can see we can start uh with some fine tuning on F three so it's very good explanation here so I'm just going to walk you through this let me just take off this so we can see a bit better right okay so let's just start by uh starting this first part here okay so you can see now we are uh connecting to a GPU so this is going to initialize this this is going to take some time it's not that long but it's just going to start this so what we want to focus on is actually our uh uh our data set here right so let me swim in a bit so here you can see uh we have our model name so this is going to be the F3 mini 4K instruct that we is what we want to train on uh let's scroll down to our data prep here so here you now you can kind of see we are following this pattern here so remember we have instructions input and response and we have some kind of instruction here that is actually going to use all of these examples so you remember our data set right input instruction and output put it's the same as you can see here and remember we need this EOS token here end of sentence token so that's good and now we kind of want to load the data set here so let me show you how we do this so since my um my lead code data set is public I can just copy this URL here so let me show you that so I'm just going to go back here I'm going to copy this and let's go back here and just paste it in here right okay and that should be it now we should find our data set so you can see this is completed now let's fire up this so every time you see this check mark here this means that we can move on you can actually see what's happening here so uh I'm just going to let this run a bit and I'm going to show you when we are actually starting the training okay so you can see now uh we kind of put in our data set here that worked well we downloaded this data you can see 13 examples remember we had 13 different examples so let's just continue down here train the model and you can see it's uploading this we can press I'm just going to press every single play here so we have the Tesla T4 and let's start here now so this is going to start our training and here we can get some more informations you can see we have number of examples 13 number of epo 60 and this is going to run uh usually takes about 20 many minutes I think uh it don't take that so you can just let this run in the background so let's let this run for a while and I'm going to take you back okay so this has been running for a few minutes now you can see uh we have about 16 minutes in total we are running for 5 minutes and you can see the training loss here now is starting to drop we started around one right we are down to 75 so we let going to let this continue of course all the way to 60 steps and I'm going to take you back there and then we're going to try it uh but while we wait for this uh I went ahead I found a new lead code problem from the uh earlier contest so this is going to be count the number of special characters so what I went ahead I copied this I created a problem here so we have the lead code problem I paste it in this solve the lead code problem in Python 3 syntax so now we actually going to do like a base test so uh what I'm going to do is copy this right and let me me do uh AMA run 53 okay and let's paste in this problem and let's see if it can solve this okay that looks a bit strange uh but let's try it so let me copy this let's go back here and paste it in and let's try it yeah there's some errors here so let me fix that and let's try again okay so there was just some syntax error so let's try it again uh this should be okay now so let's run it so remember this is fire three this is kind of our base test and yeah we got one case right at least and case one and three was wrong okay so what we're going to do when our fine tuning is done here now we're going to try to run the same problem with our fine tune model and see if it can solve it I have low expectations but yeah let's give it a go okay so our training is finished so I wanted to kind of just grab all this information here right all the steps let's copy that let's go to chat gbt and ask how does this llm training run look because I don't know anything about training models right I know just the basics so it seems this training run seems to be progressing quite well uh okay so let me just let this run and see what it says you can see the training loss uh consistently decreasing from first step and the last report step is a good indicator of effective training learning this suggests that the model is learning from the training samples and improving prediction over time okay so that's interesting if we go all the way down overall this looks on track just make sure validate model's performance and hold outet okay yeah I think it's fine so now let's actually test this versus our base model okay so if you continue down here we have something called inference here we can actually run our model so I went ahead I uh gave instructions so solve the lead code problem and I fed in our problem right the same as we had in our base test model here solve the problem in Python syntax and we leave the output blank and yeah let's try it okay so let's see what happens here now okay so it's still running here's a step bystep solution applied to problem three uh okay yeah let's let this run and then we take a look at the final solution okay so you can see solution to problem five uh okay let's copy it I just have to wait a bit okay so I just stopped it so let's try this now so I'm going to copy this solution let's head back over to Elite code and let's try it okay so it didn't get it but I thought it was a good try uh we still got the same cases uh right uh I might try it to run it one more time and see if we get something else but you can see it kind of followed our uh instructions so we have the response right so here is the step-by-step solution we have the solution and an explanation so let me just try to run it one more time and see if we get another solution okay so we have another solution example here so let's try this one so so I just H added some temperature here so I set the temperature to 0 z uh let's try this solution oops okay so let's try it okay so that didn't work either so what I went ahead I did uh I found a new problem that is also easy so let's try it on that just for fun okay so let's run this and yeah just wait until we get the answer back okay okay so we have a solution here that looks very complex right let's copy that let's go back here let paste it in and let's run it oh no it didn't get it but yeah uh it's not perfect far from it but what I like is that uh if we take a look at let's put on this you can see the response is kind of in the structure I wanted we get the stepbystep uh solution part it wants to think through that then it wants to create the solution then it wants to explain it so even though it didn't work I'm pretty happy how uh the fine tuning turned out okay so maybe we didn't get the result we wanted but uh what I wanted out from this video was not kind of actually how our fine tune turned out that is can be many things uh and remember we were running the F three a week proba maybe if we did this on llama 3 it could be better uh but what I wanted to focus on in this video is kind of how you can create data sets so that was kind of the idea behind this videos and this is not coming from an expert so don't take this for the final solution uh of anything this is just the way I have been playing around with this and as you saw from probably the the fine tune result it was kind of following our instructions pretty good so I'm going to put up this code if you want to try it out to kind of create your own yourown data sets uh it's very easy to go from text to to Json if that's something you're interested in uh so I'm going to leave this up just follow the link in the description So yeah thank you for tuning in and I'll see you again on Wednesday\n"