The REAL Reason Your GPT-3 Fine Tunes Suck (and how to fix them)

Key Takeaways

what I'm about to explain to you will completely change your understanding of fine tuning and give you a 10 times clearer understanding of what fine tuning is why we do it and how you can use it for your own purposes fine tuning is such a powerful and crucial role in the entire Ai and large language models ecosystem that I want to make sure all of you are 100 clear on what it's for and why we do it hi guys welcome back to the channel on my recent fine tuning video which will be available here I've got a lot of questions from the community and I started to realize that a lot of people don't actually understand what fine-tuning is and why we do it so I thought I'd hop on here today and give you guys a quick explainer and sort of demystifying video talking about what fine-tuning is why are we doing it give some examples of how we can apply it to businesses so that you're 100 crystal clear and you understand this key aspect of life language models and artificial intelligence the first reason that you'll fine tunes May suck is because you're expecting chat qft-like output from your gpt3 fine tune now I must admit that I2 was fooled by this initially and it took me a little while to fully understand that what we are interacting with and the API through the gpt3 API is not the same as while we're interacting with on chat gp2 you've got to understand that chat GPT is an application that has been built by open AI on top of tpd3 gpt3 essentially functions as a huge autocomplete engine and it's able to do a bunch of different tasks because of that what openai have done with chat gbt is they added a ton of reinforcement learning and fine-tuning on top of the gpt3 underline large language model in order to get it to this really conversational and excellent experience that we all are familiar with unfortunately what we have access to through the apis and that we're able to fine-tune is not the same as chat GPT therefore when we make calls to the gpt3 API we're not going to get back nice chatbot style dialogue from it it is just raw autocomplete response from gpt3 why what you may have noticed is that when you're querying your fine-tuned model it may either get cut short by its token limit or Ramble On endlessly and this is because there's so many different variables that you can change and tweak in order to get the large language model which is gpt3 in order to reach the final product or say if you're trying to make a chatbot like chat gbt there's so much work that goes in between what you have access to In The Raw format versus the final output that you're probably looking for so in summary the first reason that you'll fine tune tomato is because you are expecting chat TPT like output out of the raw language model that is gpt3 that we have access to program managers and now the second and what I think is the most likely reason that you'll find to instruct is because you don't properly understand the fine tuning process and why we do it the biggest misconception that I see around fine tuning today is not mean that I myself feel free to as well is that fine tuning is meant to onboard more data into a model essentially to teach a model new things this is incorrect the fine-tuning process is actually intended to help a model recognize new patterns and respond to those patterns appropriately that's right fine tuning is not meant to teach a model new stuff it is absolutely crucial that you understand fine-tuning as pattern recognition training here's a super simple example of fine tuning on a simple task say you have online business and you're hiring at the moment and you're receiving tons and tons of resumes from people all over the world with many different formats but you're getting all of these resumes into your inbox all the time these resumes from different people obviously don't have any strict format so the information of the person can be scattered all over the document in many different places and formats you could take all of these documents and scrape the PDF for all the data that you could find if you scrape all of these PDFs there wouldn't be any uniform structure to the data that you received let's say you wanted to take all of that complicated and messy unstructured data from the PDFs and then reformat it to be really easy and understand so that you can input that data directly into your backend systems what you could do is create a spreadsheet with two columns the first one containing the unstructured data from the PDFs and the second one containing that data but structured how you'd like it to be with each row representing one person you could get a VA to go through all of these and essentially restructure the unstructured data into the format that you would like and repeat this for the entire 100 or 200 rows that you have what you would then have is hundreds of rows of promptly completion pairs you can funnel into the gpt3 model through the fine tuning process that I cover in my fine tuning video which will be linked up here because you now have this fine-tuned model you can call it programmatically and whenever you get new resumes that come in you can take the unstructured data that you scrape from their PDF on their resume put it through your fine-tune model and get it out in a nice structured way that you wanted it in the first place and then put that directly into your systems this can easily be done by integrating it with zapier or something similar another example could be sentiment analysis say for example if you write social media marketing agency and you were looking to track the overall sentiment on one of your Instagram or Facebook ads or on all of the ads that you're running at the moment you could fine-tune gpt3 to become an expert at classifying the sentiment of social media comments to do this you can just grab a few thousand social media comments from Facebook or Instagram and then put them in a spreadsheet the same way that we did with the previous example and have all of the data in one column and then a sentiment label in the other if you get a virtual assistant to go through and label them all as positive or negative or neutral once you have that data labeled you can then funnel it into the gpt3 model to the fine tuning process and you've essentially made gpt3 sort of an expert at tinsman classification for social media comments this is useful because social media comments are a little bit different to the typical kind of text that gpd3 is used to predict a sentiment on so the nuances of what makes a positive or negative or neutral comment on social media may be different to the kind of language that it's used to seeing on blog posts or any other sort of internet archive material that has access to as a social media marketing agency you could then use that fine-tune model programmatically to begin sorting through all of the comments on your social media ads and then determining an overall aggregate a sentiment score for each of your ads to determine how they're performing and what the description is from the people you're displaying these ads from so notice how in both of these models we are not trying to teach the model any new data we are simply getting it to recognize a pattern and respond in a certain way we are making it better at a very specific task so now you may be thinking okay how can I put more data into these models to expand their knowledge for other purposes now this is a very Hot Topic in artificial intelligence and large language models at the moment because there's a token limit on how much information you can put into a API request how can we include other data sources while working within this 4000 token limit that we have on models like gpt3 now one option that you have to expand the knowledge of your model with new data is called semantic search now this is quite a complex topic and I don't think I can cover it fully in the scope of this video but I'll give you a quick rundown so you know what I'm talking about somatic search essentially allows you to take in a corpus of text Data say you have tons and tons of PDFs of box by student author or legal documents you can take that Corpus of text Data put it through a specific function that you can get access to the libraries programmatically but essentially these functions are going to vectorize all of the information chop it up chunk it up and then put it into a multi-dimensional vectoring system so that it sort of tags what the content of these different chunks of data is and then remembers it for later and stores it in a database so all of your text data is taken and chopped up and stored by similarity within your database you're then able to prompt your model when you put your input in it's going to top that up in the same way and vectorize it and see what's similar to it and then bring that data back and then provide you with an appropriate completion using any data that was in the database if it's appropriate so this is a very useful and popular way of adding more data to your models and giving an x a few more things a super simple example of this would be to get around the limitations of gpt3 only being trained up to 2021 data say you want to train it on some new Wikipedia data you could scrape all the Wikipedia data and put it through a semantic search vectorizing function hold it in a data and then whenever you query about their specific Wikipedia article it would go okay I'm looking for something similar within a database and we'll look for it pull it back and give you an answer that a normal gptc model wouldn't because of its data and time limitations on what it was trained on now if you guys like a little bit more information on semantic search and how myself and others are using it in large language model applications and I can do another video on it if you'd like me to let me know in the comments below if you want some more information it is a little bit technical but I'm sure you guys can handle it another option you have to get these models access to more data is by allowing your large language model of choice to interact with a database one of the many functions of large language models like gpt3 is a natural language to SQL query because of this ability to translate natural language into a SQL query we essentially are able to take a natural language and then get access through any information in the database of our choice this is typically done with libraries like land chain which allow you to essentially chain together different operations with different models for different purposes in order to get a specific output so land chain allows you to take in some input take that input into a SQL query perform that SQL query get the data back whenever like that and then translate that again into natural language and then give that to the user so a lot of things hidden behind the scenes but something like land chain allows you to take in the natural language and convert that to essentially database query and get it back to the user I'm actually working on a project right now with a client of mine who came to the channel myself and my development team are working to get a database lookup system that takes a natural language create a database and gets appropriate responses back in natural language so it's definitely possible I'm currently doing it and I'd really really recommend this for numeric data because semantic search can't really handle numeric data does database method can if you have any burning ideas you'd like to have a chat to me about in terms of the session of feasibility or planning the development now or just get my thoughts on the idea in general and you can book a call with me down below it's in the description and in the pin comment my Consulting link so if you'd like to have a chat then you can reach me down there now both of these methods are really really awesome ways for you to extend the knowledge of your large language models and once you combine it with libraries like land chain you can start to make some really really complex applications and it's really important to understand how fine tuning fits into all of this creating fine-tune models that are great for a specific task can be so important in building these larger applications in line chains teacher so if you have a fine tune model that say like we used the example earlier that takes in the resume data that's unsorted and can sort it that is a sort of key step in structuring data that you could then pass on to different parts of the application so keep in mind that fine tuning can be a very very crucial tool when you're looking to build your applications because these models are so powerful if you can make them more powerful at even one specific task they can really bridge between different parts of the application and sort of take data from one form and convert it to another so that you have can sort of connect it all together and make a really really powerful application at the end of the day now regarding land chain I will be making a video on it next week so make sure you're subscribed and you hit the bell for that because it is absolutely crucial that you as an entrepreneur understand land change so that you can begin to sort of frame up what you can do with these language models when you connect them together in a chain once you understand land change as an entrepreneur you're going to be able to see opportunities that you just didn't know existed before I better wrap it up here but that's all for today thank you so much for watching I hope I've been able to clear up some lists and misconceptions that I definitely had at one point and I really hope I've been able to clarify things for you because understanding how fine tunes fit into the whole ecosystem is very very important and not spending too much of your energy expecting to get something out of it that it simply can't do so as always if there's anything you'd like me to make a video on in future please let me know down below if you have any questions about what I've gone over in this video comment down below I'll be replying to all of them and if you're looking forward to seeing the next one please subscribe to the video hit the Bell so you don't miss it and leave a like on this video If you enjoyed it really really helps my channel that's all I asked of you that's all for today thank you so much for watching and I'll see you on the next one foreign