Udaya Sai Chikka

So I have been I'm with us. So I've been working on this field, like, for, like, almost 5 and a half years. So, initially, I started working as a Python developer, and, uh, I guided my proofs of, you know, like, uh, you know, helping his PhD completed, uh, you know, for doing, uh, made Python based problems and automation, all that stuff. And I got introduced to the machine learning way back in 2018. And from that point of onwards, you know, I had been working on cutting edge AI ML techniques. Uh, initially, I worked as a Python developer, and, Uh, later, you know, got into PCS and where, you know, I got an opportunity to work on, you know, natural language processing problem, like, chatbot billing. Uh, I took that opportunity and build the, uh, you know, chatbot for a US based banking client, uh, within 3 months, and, uh, it was a great success. And I got promoted to, like, a leading engineer where I got introduced, like, More challenges is client facing role, uh, where not only, you know, I help up I help them, like, uh, to build the process, But, also, you know, uh, I convert those to the, uh, you know, client requirements to technical requirements and converting those technical requirements to no or neutral tasks, where I'll be handing those tasks to my teammates and, uh, you know, completing that work and, you know, reviewing their work and, you know, reviewing their work and, you know, taking that. And after that, after TCS, you know, uh, I'm the 1st employee of a a research based company, uh, Landstar. Uh, it's a Canon based phone. So my role was, like, a over there. I lead a team over, or, like, 5 people over there where we work on, you know, natural language processing, chatbot development, uh, Charge GPT generated via applications. Uh, not only that scalability, DevOps, Uh, code reviews, all these things. You know? Uh, I have actually worked on that. And, uh, also, uh, we we we were the, like, First one in the market to introduce, uh, uh, generative AI, uh, into the risk the space. Uh, and, uh, you know, there I work for, like, 1 and a half year. And from, like, last, uh, 1 1 1 and a half year, I've been working on these large language models, uh, like, BERT, Chargept, GPT 3, all these things. And, uh, my my my primary skill set was to, you know, Roy, these large language models, scale them, and also, you know, uh, make sure, uh, make make sure they're they're actually accessible to the general public at reasonable amount of cost. So we're actually building something very similar to ChargeGPT by using open source model, Uh, 75,000,000,000 parameters, uh, for the real estate space. And, uh, recently, uh, you know, uh, we we also, you know, Got a chance to work on with Vertex AI, uh, from Google, uh, where we actually worked on, like, you know, building, uh, auto utilizing our ML and the model garden. Uh, we were actually, you know, able to, you know, build, uh, some good, uh, modern endpoints, you know, with, uh, make make them available for I downloaded through API, basically. So yeah.

So the loss function, uh, basically, it's a a very important parameter. Uh, when choosing a loss function, you need to make sure that the type of the data that you have and also make sure how it is actually working, how it is actually performing over, Uh, different epochs, basically. Uh, not only that, uh, the the the criterion for loss function, uh, depends upon, uh, the type of the problem that you are also, you know, uh, selecting. Like, for example, either you are selecting regression, uh, you wanted to select another problem. If you sit back, uh, if it's a classification, it's a different problem. Uh, but, also, at the same time, uh, if you're actually dealing with, like, you know, general, uh, text or, for example, uh, we are, like, a text based, uh, you know, transformer based applications you need to have, like, different loss function where the importance of, you know, uh, human feedback, uh, the importance of things, how we where it went wrong was actually really important in these loss functions. So while developing a model, uh, definitely, like, it's really, uh, important to make sure to check the lost scopes. And, Uh, my search was to, you know, to try to see, like, which loss function is actually working for the nature set of epochs, Uh, try to run with the different set of loss functions and see how the loss scores are actually whether they're converging or they're, like, they're diverging, uh, with the test and prime datasets and also, like, you you make sure that, you know, the coves, uh, whether, like, whether they're With that, you know, you understand whether they're, like, fitting or under fitting. And, also, this loss function needs to provide valuable feedback back to the, you know, model so that it actually makes the right decisions. So the loss cost can actually tell you whether it is actually giving, uh, right amount of feedback, uh, back to you or not. So whether the loss function was, like, you know, it's like whether it's, uh, giving very high feedback on, uh, when, like, It's giving very slow feedback. It actually depends upon the last function, the type of the problem that we are solving. So it actually very, very depends upon the, uh, dataset, uh, that you have, whether there is any the dead set drift is actually there or not. And, also, uh, whether there is any, uh, data imbalance is actually there or not, uh, whether it's a classification problem, a regression problem, or it's like a, you know, like, uh, generated AI. Like like, for example, uh, you know, like, text generation, uh, thing or, like, I would say, like, a major major generation thing. Uh, it It actually depends on how well, uh, for its image classification, you need to see, like, how well the images were there, like, whether they were, like, high quality. The pixels were really good. The data is actually good or not. So all these things, parameters, you know, like, um, my my I'll choose all these things, uh, while actually solving our we're actually, you know, designing a deep learning model. The loss function, uh, is really important here. So, uh, I I'll consider all these aspects.

So large scale dataset, uh, is actually, uh, you know, like, when the data is large. Uh, so, definitely, you need to make sure that, uh, you're not actually you don't have any written data. So my 1st goal would be, like, uh, you know, to convert all the, uh, you should first analyze the data. You know? 1st, uh, you should not, you know, run into a training first. So you should first analyze the data. You should first analyze and then transform the data. So you in the analysis part, you know, you plot some graphs on top of it, uh, observe the skew skewing the data, you know, perform some techniques, uh, like, you know, like, uh, whether there is any, uh, redundant data is there. Uh, you apply techniques, like, you know, the data retention techniques. And, uh, so you transform the data. Uh, like, a play log on a certain column, uh, see if it is, like, making any in the data, whether the data is actually you know, like, uh, whether we need all the data or not, uh, is to maintain, like, a perfect split in between the, you know, and test, uh, split basically and how well you're actually choosing that split. And, um, you know, also, with the large scale dataset, you know, you need to make sure that are actually giving right amount of feedback back to the users. So whether, uh, you want to be using deepening techniques, uh, you wanted to use conversion neural network or, you know, normal regular neural network or, like, it's a, uh, you know, like, probably a transformer based approach or fine tuning a model, um, all these things. You know? So, definitely, a large scale dataset, uh, you should not you know, when the data is there or when the sufficient amount of data is there, uh, it is definitely possible to build a good machine learning model. So when the data is less, switch a challenge. Similarly, if the data set is larger, so large scale data set is also there, uh, when the scaling you know, you need to perform a scaling operation aspect because there's a very high chance that the outliers are in the dataset. You know? You need to bring back them to the normal scale and make sure that, you know, your data falls under the bulk curve. And, you know, uh, if if it falls under the bulk curve and it's under the normal scale range, uh, you know, it works really well for the convolutional machine learning models. But, of course, you know, uh, with the advancements in technology, we can address all those things. Uh, but at the same amount of 10. Uh, we also need to make sure that, you know, uh, we are not overfitting the data. Okay? Uh, so you apply you first transform the data, anal analyze 1st, uh, plots and graphs, transform the data. And then based on the transformation and based on the data, uh, I mean, like, its type of application and, you know, probably based upon the characteristics of the data. Uh, you you observe. You see whether it's, like, a perfect, uh, you know, like, based on the target variables, you see whether it can be linearly separable or not. So if it's linearly separable, you know, you have a conventional method, like, so for example, methods. Uh, it can do that. And boosting debusting algorithms, like, all those things, like, well, they'll work really well for support data, uh, like, an operation based and all those things. And, also, uh, we we also do have, like, you know, transformer based on deep learning methods, like, correct neural networks, uh, deep learning, all these things. So you can use that and to, uh, well, uh, you know, large scale datasets, you know, to actually try in the data model. Uh, so yeah.

So it's it's really important that, uh, when you actually define a neural network architecture, uh, you need to make sure that you at least follow, like, 4 parameters. Okay? You you need to see whether the nodes, uh, and then the edges and then the curves, all these things. Like, for example, if you if you if you, uh, know if if if you can observe a normal, uh, any any any problem, you know, uh, for the first thing that you need to do, for example, is check and image classification problem, probably. Okay. So how how how do you see the image classification problem? So first, you identify, whether the pixel is there or not, okay, in the in the 1st network in the 1st layer. And in the 2nd layer, those dots actually form like, uh, you know, like, lines, uh, whether they're inclined or, like, straight lines and all those things. And, uh, in the in the next like, the combination of, uh, all these layers add up and, you know, these dots forms 2 lines, and these combination of lines forms 2, like, you know, shapes, like square, rhombus, or, like, in a circle, all those things. And, also, these these 2 layer will all these combinations will add up you know, formations of objects, like, for example, head, whether there's in head or not, or let's say, if the circuit is there or not, all those things. So my my, uh, approach would be, like, to define, uh, at least, like, uh, you know, 4 or more 4, uh, a layer model, uh, with each layer having, like, you know, maintaining, uh, to cover at least, like, all these, uh, you know, shapes, I would say. So the input, uh, in I mean so based on the input, you know, the input layer will have to, you know, convert that to, you know, like, these things. And, also, make sure, I I really don't want to go to the dense layer, but I I definitely want some convolution because of the convolution. You know? Your number of operations and the, uh, calculation will be really easy for you. And, also, uh, I kind of, you know, based on, uh, how well my, uh, you know, the number of layers of the dimensionality is increasing. Probably, I'll I'll include some, you know, dropout mechanism, uh, to make sure, like, you know, Hey. Uh, I want my model to be learn not learn on any any data. Sometimes I want to miss the data intentionally, so I improve also. So convolution, dropout, all these things with the perfect loss function, uh, you know, we are able to, you know, uh, come to a, you know, like, balanced complexity and performance, uh, here, basically. Uh, and the cool thing here is, like, you know, you you can actually, you know, sense that in your mind itself, uh, how well your model is actually performing. And, uh, not only that, uh, but you also need to see, like, uh, when the number of layers were increasing, okay, in the model, uh, the complexity is increasing, it's really hard to give up. Okay? So you at at the same time, uh, not all, uh, datasets, you know, not all, uh, classification, but not not all neural network art pictures we'll be able to like, it's not a one solution for all problems. Right? So, definitely, you know, you need to make sure that, you give a little bit of edge, uh, depending upon the problem. And, also, make sure you keep an eye on the improved you know, increasing the complexity. So, you know, you need to kind of drop fit and also give more importance to the internal stages and also initial stages. Because if if you are not capturing the immediate information, immediate in the in the stat, no, uh, like, you will not be able to have, like, at the end. So import data is important, and you need to make sure you choose the

Okay. Uh, so finding loss function. System loss. And what kind of loss function should be? Yeah. So For generate to text, uh, I mean, uh, you are actually, uh, try you you're actually, uh, I mean, the loss function that was defined here, uh, might not be a good, uh, because it's actually good for erection, I mean, uh, regressive problems, basically. So here, uh, what it is actually essentially saying was, like, if the target variable was, like, continuous variable, like, Any price of the house or, like, age or something like that. Uh, it will be really good loss function because, you know, it is actually, Uh, you have a target, and you you model picture something. And, uh, this loss function is able to, you know, uh, do the thing. And, It it was able it was able to give you, uh, like, how far it went. Okay? Uh, so that will be really good estimate for, uh, like, regression problems, but not for generative applications. Uh, because for generative applications, it's, like, really different. Uh, so it actually works based on the, like, human feedback, basically, like, you know, you want what is the next proper word, okay, uh, you want? And, uh, like, there may there might be, like, a number of, uh, proper words, uh, because, you know, in we have lot of synonyms, and we have, like, lot of, Uh, promo webs and all these things. Uh, so there might be, like, lot of lot of, uh, you know, possibilities of the next text. Right? So, Definitely, it's not really a good loss function. Generally speaking, like, The lost functions for generative, uh, text is, like needs to be really depending upon, like, um, like, the basis of, you know, I want a human to be involved, you know, while, uh, doing this kind of, uh, technical. That that's how that's where, like, you know, fine tuning and, you know, like, But there's, like, uh, RHLF and, you know, PFT, like, prime reduction, the time binding methods. So all these things comes like, uh, you know, Like, we have, like, a different, uh, methodologies, like, supervised, semi supervised, and, you know, uh, unsupervised, different looking methods. So, Definitely, I consider the for the generative text, uh, we should be, like, you know, we should not use our we should stay away, from, like, conventional loss functions and, uh, rather choose, uh, what is actually applicable for them, uh, specifically for such type of applications, Uh, is my opinion. And, uh, not only that, not only for TensorFlow, but also PyTorch and other frameworks as well. It's respect to the framework that you are using, or, like, whether you whether you are actually opting, like, a cloud based, uh, like, on this, like, Vertex AI or, like, Gemini or, like, any Any other, like, SageMaker or or any other, uh, you know, like, tool cloud based tool you are using? You need to make sure that, you know, you you do the data analysis part, uh, data evaluation part, and, uh, the timing part. Correct. And for that, you choosing the last function is also, like, really important, and it's it should be should be really, uh, wise as well, mission

Yep. So far, recently, we are actually, you know, uh, building a generative AI model. So, uh, we're not only implementing a generative AI model, but I'll I'll tell you. So, like, we are actually working on open source, uh, last language model, uh, it's called Mistral. We're actually using a soundmillian parameter from Mistral, which was actually open. So we are trying to build something, uh, you know, for real estate space. So we wanted to, you know, uh, create a, you know, charge GPT interface kind of a tool for leadership people, which can actually help them in cold emailing, uh, SEO and, you know, uh, publishing their, you know, uh, leadership material and all that stuff. And, um, the problem that we faced was, like, you know, uh, the the available the availability of the data and the availability of, uh, you know, deployment method and strategies, uh, for open source metadata, open source LLMs were, like, very less. So, uh, we have, uh, tried a lot of options. Like, we tried on going with, like, VLLM, uh, one of the methodology to, you know, uh, deploy, uh, large language models and also we can, uh, deploy this in house solution in house solution, and, also, we can actually, uh, go with, like, their, uh, cloud offering, uh, to access that API, uh, but these are, like, all are all costly. So what we did was, uh, you know, we we kind of, you know, implemented our own, uh, in house deployment methodology, wherein we actually download the entire 7,000,000,000 model and, you know, uh, kind of, you know, make that inference possible, uh, by combining that with FastAPI, uh, which is a, like, a Python, uh, API based framework, which can actually do the inference in real time, uh, and we tied that to, uh, you know, our, um, what to say, like, uh, our SaaS based, uh, you know, platform, wherein we are actually making calls directly from our, you know, interface, basically. So the challenge, like, you know, was mostly in around, like, you know, deploying and, you know, scaling that, uh, and making sure that it's actually available. And, uh, it is not actually a hallucinating and all those things. And, uh, we need to, you know, do the fine tuning part as well for that model. Uh, so we fine tuned on, like, you know, custom generated dataset, like, for the real shift applications. We kind of fine tuned it. And, also, in fine tune, we kind of, you know, uh, decided, like, which methodology that we need to go through. And, uh, we try different, uh, approaches, uh, like, uh, p, f, your, and, uh, human loop and the feedback and all those things. Uh, and we end up, like, you know, having our model running, and it's actually working really well. And we're really happy and what the progress of the model. So, uh, it's not only about the problem or the challenge. It's it's also about the implementation and how we are actually dealing this situation and, uh, uh, how we are actually dividing that entire task into simple pieces. Okay? Uh, you have this big problem and, you know, you need to first understand, like, how we're actually solving each part each piece of it to solve the problem entirely. So that's my strategy, and that's how we kind of, you know, did that.

Yep. So, basically, uh, in in the in the given code, So it is saying that transfer model object has no attribute from from pre print print. Uh, so, basically, Then the main reason was for this, uh, thing was the package, the transformer mode, and whatever was there, uh, has no method Call dot from pre prime. Okay? So, like, uh, it's actually like, uh, the thing is that you need to, Uh, import the transformers module, and we need to make sure that it is actually available in the Hugging Face library. Hugging Face, uh, you know, like, uh, transformer Space space. Type in face space. So you need to get the model name and user tokenizer and all those things. Uh, but to make sure, you know, you use a transform I think this transformers model library. Your transformer model, uh, is there. You you imported that, Uh, but doc from Deepgram is, uh, no has no attribute. That means we're actually not using the correct, uh, package or correct import here. So you need to use the correct import and make sure that the model is actually available in the transformers, uh, space based study. So that might be the reason, Uh, why it is not actually giving you the, uh, why it is actually giving you the, you know, like, the error? So that that might be that might be one of the reasons. And, also, uh, you need to make sure that, you know, you have, Maybe it might work for, like, you know, the previous versions, like version 1, version 2. There might be different versions. Right? So, you know, also need to make sure, uh, which version you have installed. And, uh, for that version, whether the code has been any change or you need whether any duplications or removals, Uh, the particular code might work for, uh, some other version might not work for the latest version. So you also need to make sure that, uh, you know, you check the version correctly. And, also, uh, you need to make sure that you have, like, uh, like, all the dependencies installed, uh, for that. For example, top of the transformers library, whether you have, like, Call the, uh, dependencies for install correctly or not in your environment or in virtual environment whether it's like, If if you still regard there, like, you go to Stack Over for check with the even from the community, I think this community whether anyone are actually facing the issue and, uh, see how we're going to Google that. So and, uh, that that's that's how I these are all the things that I I try to, user, you know, to resolve the problem. And, uh, if nothing is actually working well, uh, probably, uh, I'll see, Uh, what is actually causing the issue by going through the source code of, you know, like, the transformers library. And, uh, you know, see, like, uh, probably, I'll submit a, you know, issue, uh, in the in the in their, uh, GitHub issues in their GitHub page. And, probably, if I'll be if I'm able to solve, probably I'll The problem I'll explain the problem. I'll to see, like, how do we clear this problem and, uh, you know, submit API possible, uh, with that. So That's my, uh, approach to solve this problem, basically. So yeah.

Okay. So Python is also, like, really good, uh, you know, framework to, uh, build the machine learning morals. Uh, so, yeah, as I can see here, So with the transformer block Uh, here, you know, uh, you first the layered emissions are actually going well here. So transform block you have in it, and, uh, all the definitions are actually going well. The forward pass implementation x was, like, you know, you giving that. Uh, but, uh, your your the problem is that the potential is is that, you know, you're not you're giving you're passing the data directly, uh, like, 3 times. Like, uh, the attention layer, like, you you, like, uh, you need to pass it, like, through phase. Like, you know, 1st, you need to pass to the previous layer, and then you pass to the next layer. And, uh, it has to go, like, step by step. Right? The attention parameter, like, you know, you also need to, uh, see, like, whether you wanted to do anything, preprocessing with the data, like, uh, all those things. Uh, but the Forward layer and the forward layer. But you directly going to the self recognition and, you know, passing, like, the, uh, parameters, like, these. Uh, it's It's not it's not really a good implementation, I would say, uh, for this kind of problems. And, uh, especially, Python has really good documentation, uh, on the Internet, so definitely, uh, make sure you check the right, uh, like, the best practices, The best practices for, you know, uh, developing such kind of, um, you know, uh, uh, transformer blocks and defining that. And, uh, yeah, I hope, uh, the issue can be solved easily, uh, if you can pay attention in the forward layer. And, also, while initiating the the layers as well. Um, the transform block, uh, we need to follow the approaches for the neural for them in the general approaches, uh, like, you know, like, um, what to say, like, you the deep learning and, uh, you know, the conversion in your network and all those things, you need to try, and then go to the transformers. Right? So, uh, the importance of initial, the transformers is really important. So you make, You should give importance to the tension layer of it. So that that only, you know, you'll be able to, you know, succeed in all these kind of, uh, solutions, and, you know, you'll be able to define your network architecture really well with PyTorch. Uh, and then, you know, uh, it's really easier. It's like just like cup of coffee for, like, you know, uh, everything, uh, how we are doing and, you know, all these things. The standard structure is that clearly has to be good in the forward, uh, method and, also for the initiation layer as well. So, yeah, I hope that answers

Yeah. So state of the art generated models, like, for example, PaLM or, you know, Chargegptgptgptg, all these things, uh, we can these were, like, you know, pretty much latest and, you know, all those things are in they don't have really SDK in, like, each and every language. So most probably they have SDKs for, like, Python, I would say, like, are probably, like, all these things. Right? So, uh, in such cases in such cases, what when you wanted to degrade that with the legacy systems where your applications were, like, building, like like like, 10, 15 years back, uh, like SAP or Java or like PHP applications, uh, know to interface or to integrate those, uh, into those applications, it's really challenge for you, uh, to, you know, delete all these things. So one of those that you can do or what I can do, what I'll do, uh, actually get this in, uh, you know, degrading this, uh, with the ServiceNow, basically. ServiceNow responses platform, uh, wherein it actually handles the user request and, you know, do the stuff. So we sit with the client and the requirement. We design the solution and all those things. So the product that we did was, uh, we kind of, you know, take part of this generative AI aspect entirely into, uh, the outside of legacy systems, like building an API, deploying that generative model, are, you know, accessing that through API, basically. Whatever the stuff that you wanted to do, do it outside the system. Okay? And interface that system with, uh, API call basically. Uh, pretty much any legacy system, uh, depending upon the time of the legacy system, there might be possibility that you interface that through API or probably an extension or probably, you know, like a custom script. So depending upon that, you know, uh, whatever the out system, uh, in system. Okay? I'm talking about the system, saving system, and out system. Outstream is, like, you know, the general generated way model, which was actually fine tuned for your use case, deployed, ready to go, ready to take the calls and all. The out system used to be interface within systems and legacy system. Uh, so you you write the RAPL, model where Haida would say. So if it's like a a job based on, like, you know, something like a web based application, you know, your middleware is like an API. Uh, if it's like an annotation or like that, you know, it's like an action, uh, things like that. So it's really important to divide the systems. Uh, the legacy system has to be different than the out system, which was, like, the machine learning or the generative AI system. And the middleware is, like, the connecting between these 2. So then, you know, you'll be able to, you know, uh, not only it actually helps you to, you know, uh, disturb the systems, not only helps you to, you know, not issue the system, but also in the long run, it is easy to manage and efficiently, you know, find any problems, or debugging is really easy when you do run the systems and, uh, you know, integrate the the existing system, which was really important in my opinion. So yeah.

So, uh, there there might be there might be chances, uh, that when when we're actually using generative AI models, of course, uh, the data might be skewed or, you know, probably, there might be a drift in that. Uh, so impact on the generated data, like, you know, like, uh, it's still important that, you know, uh, in my opinion, uh, when we're actually giving input to generate via model. Okay? Then there are there were, like, different layers. Right? So we have we have a input. Uh, we have, uh, when whenever we we ask the question to generate, we are. Uh, first of all, you know, it needs to or go through a sequence of steps. I needs to go through sequence of steps before it teaches to generate AI. And, also, uh, when The output is actually coming back to the, uh, user or back into the, you know, system or back to us. It needs to follow or it needs to go through, like, system of layers, basically. So, you know, when when we give the input, it needs to go through, like, you know, vector stores, uh, to see, like, whether there was any, like, you know, you you pass the context. Right? You along with the input, right, you pass the context and find any similar queries were actually answered previously or not. All those things. Right? Vector stores, fine cone, all these things. You will be able to, uh, build a vector store and, uh, you know, do all this process of any similar content and adding that to the context of your input message, all these things. Right? So, similarly, uh, when you're actually, uh, collecting the output at the, you know, uh, at the last layer or of the thing. Basically, you know, you need to see you need to perform similar kind of approach. Uh, that that way only, you know, you'll be able to see whether your data is cool or not, uh, and, uh, you know, whether, uh, the model is actually performing, uh, not it's kind of not performing, basically, in my opinion. Okay? And, uh, see how you kind of, um, able to do it, basically. So that's that's kind of really that's kind of really important, okay, for you. Uh, and, uh, if if it's actually kind of doing that, uh, you know, the good solution would be, like, you know, to fine tune that. Uh, you know, you fine tune the large language model, uh, you know, like, with pretty much, like you have Mishra, Lima, Baklava. All these are open source models. Okay? All these are open source models, and we have, like, lot of, uh, open source, uh, committees, like, clearly getting we have a lot of, uh, good people on the Internet, like, you know, doing some crazy stuff. Right? So you took the model and, you know, uh, try to see, like, how we can actually fine tune the model to mitigate the, you know, generated content and, uh, you know, how you're able to resolve this issue, basically. So, you know, uh, probably I'll pick, uh, machine learning model and, you know, see how it goes with the already known answers and, uh, see, Uh, if if the model form is not good, I fine tune it. Uh, if the if the parameters were too much increasing, I probably I'll I'll, uh, know how to, like, a prime attrifying methodology to reduce the number of parameters and kind try and try to reduce my cost, and at the same time, Uh, also make sure that, you know, that my, uh, generated content was not skewed, and it's kind of giving right answers and also correct answers, and stuff, you know, not hallucinating, like repeating lot of stuff and, uh, things like that. So that that's my, uh, approach, you know, uh, to solve, like, uh, this kind of generative AI score models. Uh, but, yeah, at the end of the day, like, uh, no. You need to monitor the performance of these models, uh, regularly, and then you opt for, like, MLOps, uh, like, you know, uh, deployment scaling, CICD. Everything, you know, you you need to set up a pipeline, uh, wherein these MLOps, uh, things, uh, machine learning operations things, uh, work really well if you do that, and, you know, you'll be able to, uh, you know, identify the issue in the system and, you know, We're early in the system. Uh, you are in and the best

Uh, for a pretrained model for chatbot project Okay. So, probably, uh, I'll choose, uh, Mistral. It's one of my favorite, uh, you know, model, uh, basically for chatbot projects because it works really well, and it's a small model. Fine tuning is easy. And, also, uh, for the given tasks, they actually, uh, no for the given task, it actually works really well, Because why? Because it's since it's a, you know, a large number of size 7,000,000,000 parameters, uh, I can easily fine tune it. And, also, uh, the nature is that it's open source model. Since it's an open source model, you know, uh, there were, like, lot of, uh, things were actually a lot of people did lot of work on that and, you know, uh, did that. Uh, if a number of parameters were not, like, uh, you know, important to me, uh, I choose LaMDA. Uh, LaMDA, basically, or also if the multimodal prompting is kind of important for me. Uh, but it would only text based to chatbot. I'll go with the and, uh, also depending um, will things have thing here is that, you know, Uh, my do some research, uh, study the architecture, and, uh, depending upon the architecture, whether What was the what was the manifest Uh, well, the architecture, you know, really tells you whether it can be suitable for our chatbot projects or not. So, yeah, all these things actually matter while selecting a, you know, a predefined model, basically. And, again, in the chatbot project space, uh, there were, like, a number of, uh, things happening. So I probably tied that with, uh, some open source, uh, you know, like language chatbot framework, uh, you know, like Rasa. Asahi is one of the, you know, leading, uh, chatbot frameworks. So, probably, I'll tie Rasa. I'll use Rasa's, uh, what to say, the chatbot framework capabilities, and I can actually easily integrate, uh, this with any, uh, you know, easily with any chatbot framework. So that's the essence of it. You you you convert that into an API. You run it on inference, or you wanted to, you know, build an Android app. Select everything is possible, uh, with this. So, yeah, that that's that that's how I choose. I'll choose a open house framework. At the same time, also make sure whether I can actually deploy it easy, uh, and, also, whether, if if some if someday, like, whether there if there is an issue, uh, whether I can do the, you know, backup, uh, clearly or not, uh, how well the data is actually, you know, with the parameters, uh, it's doing, uh, like, whether it's challenging or not, all those things. So yeah.

So, basically, for fine tuning, a GPT 2 model, uh, you know, like, domain specific language is something you need to you need to make sure, unit. It's it's it's kind of really it's kind of really different what he's actually trying on. Right? So, uh, I rather wanted tune So to use an approach called, uh, no, r h n. There might be, like, different, uh, uh, fine tuning methods. Okay? Tuning, semi supervisor fine tuning is there, supervisor fine tuning is there, or, you know, our total unsupervised fine tuning is there. And depending upon the type of application, whether you are actually doing AI model fine tuning or non generative way of fine tuning model, it's different, basically. Okay? And, uh, they were, like, obviously, uh, so is there and, uh, you know, like, not supervised fine tuning, if I would say. But, um, you know, you, I would say RHL is, like, one of the better, uh, because it's a domain specific language. And, uh, the human feedback really actually helps you, uh, to make sure, you know, the model the GPT 2 model understands the data pretrained transformer model, understands, uh, the domain correctly. And, uh, also, you know, uh, the human and the feedback actually helps them to, you know, correct, if it is going in a wrong direction. Right? So it's how how good is that? So but here, the thing is that you need to make show that, you know, parameter the number of parameters that you are doing, like, will increase, obviously, when you find them. Right? So you need to, like, opt for, uh, strategies like, you you know, parameter action, fine tuning, uh, and how to you know, we we dig out those number and to reduce the number of parameters and also at the same time, if your model was, like, being too complex and, you know, it's picking enormously large, but it's it's becoming difficult to maintain that. Right? Office, of course, will influence cost will increase and, you know, uh, deployment cost will rise up and all those things. And, uh, doing all that, if if it's work, it's good. Otherwise, you know, it's like like time of waste. So my approach will be, like, uh, you know, uh, humanly feedback because it's, like, too much specific language. And, uh, yeah, uh, I'd rather choose that.