Kundan Singh

Help me understand more about background brief instruction. Okay. Sure. Uh, myself, I have total experience of 8.5 years, and relevant experience in data science is more than 5 years. In June 2015, when I completed my BTEC, I started working on Python. And, uh, later on, I did different certifications with respect to data science from Coursera and migrated myself from Python development to data science. And, uh, I have done my master's as well from an IT War angle in artificial intelligence and machine learning. And, currently, I am working with our systems international private limited. And here, my formal designation is as a senior data scientist, but sometimes I work as a team lead as well. As of now, I am, uh, currently working and leading a project which is, uh, related to generative AI. We are creating an AI model that, uh, and we are creating an AI model, and our end goal is, like, we are going to create an AI chatbot who's gonna interact with end user in a video format. So, like, um, like, any 2 persons do video calls. So in the same format, our AI chatbot is gonna interact with the end user. So how exactly would it work? Like, uh, it basically contains 2 parts. 1 is related to, uh, my large language model part, like NLP part, and one is related to converting text into video. So I basically lead this complete project. So I basically handle and lead up both kind of scenarios, NLP related part as well and computer vision related part as well. In in a related part, uh, we are basically using, uh, lava 2 here, and we are fine tuning it with respect 2 different documentation with respect to different scenarios. And, uh, a part of that, with respect to computer vision side, so how exactly would it work? Firstly, an end user would basically speak his questions. So using so firstly, his audio would be converted into text, and that text would be given into, uh, LLM model as an input. And after, uh, providing output to that LLM model, the output will basically, uh, provided to our, uh, AI model, which basically converts that text into video. So converting text into video, which is basically going to, uh, be as an output to the end user. So here, uh, we can use any person's face. Uh, just provide some photos with respect to that question, and we are going to convert that into avatar of that, uh, a j network. So here, uh, we used transformers bark transformer we used for converting text into audio. We used wave tulip. We used GFP gain, uh, a part of that OpenCV and, uh, GFP gain, we also used and different other, uh, like, FFFP easy to use there. And apart of that, I have good experience with respect to traditional machine learning algorithms, uh, like classification, regression, a part of that, uh, like, uh, object detection related things, like detecting different anomalies from in terms of images as well or in terms of different data as well. So these kind of different things I have performed till now. So, yeah, overall because from the starting of, uh, my journey as a data scientist, I am working in service based organization, so different kind of experience I have.

Okay. So training a large language model, uh, with respect to, uh, like, Like, training a large language model with respect to our current scenario or our current data. So different techniques we we can use. So, uh, techniques like, There are different scenarios, basically, different techniques or end scenarios we can perform. 1 is related to, like, what kind of Training, we are doing. So we can use parameters based training like EFT. And even though, uh, different Training in in the sense, like, complete training of the, uh, all the weights related training with respect to large language models. And a part of that, Uh, as I said, like, what I've I normally prefer, like, instead of fine tuning with respect to existing weights of that Last language model, uh, instead of of fine tuning all the date database, we should train with respect to, Um, with respect to the new neurons and with respect to the, uh, do you basically call this technique like PEFT? And, uh, we can use LoRa for that, and we can use Q LoRa. It's, uh, although a new technique with respect to this one. And, uh, auto tokenizer, we basically use when we provide our, uh, own data to Train with respect to large language model. Uh, as I said, like, for example, we are, uh, for example, we are basically using Um, LAMA 2 model. And with respect to LAMA 2 model, we are fine tuning it with respect to our questions and answers. So in the same specific scenarios and specific format, We have to, uh, like, create our model, and then we have to provide, uh, that specific scenario data to be trained on. And, uh, yeah, these things we can perform. And, uh, a part of that, uh, other technique, And a part of that yeah. I think that that's the thing that we can implement.

In detail a method for optimizing performance of a deep learning model without overfitting. So, uh, overfitting, uh, overfitting is a condition which, uh, we identify and get to know. Like, uh, let's suppose we have trained our model, and with respect to training data, uh, it's performing well. Let's suppose it's providing an accuracy of 93%, but with respect to testing data, it's not performing even though, let's suppose, 85% accuracy. So this scenario, uh, and this condition is called overfitting, like, with respect to trained data, it's performing very well, but on testing data, it's not performing well. So, uh, to, like, tackle these conditions, so what things, uh, like, we can do? Firstly, different scenarios and different methods we can implement. Uh, 1, uh, 1 method, like, with respect to with respect to complications with, uh, of our neural network. So our neural network should not be much complicated. Uh, it should be I would not say, like, less complicated, but it's it's complications should not be on the higher side. We should use drop out layers after a few steps so that, uh, we would not take on overfitting problem in the future. And apart of that, uh, like, uh, leaky ReLU activation function we use. So using activation function, we can also do that a part of that optimizer like Adam and these optimizers we can use, and, uh, a part of that data augmentation, we should also perform before sending our data to be trained on. And, uh, yeah. And apart of that, uh, we should make sure that, uh, with which data we are currently testing, it should not be completely different from the data with respect to which we have basically trained our model. Like, there should not be data mismatched related thing. Like, our training data and testing data soon be related to each each other. So, uh, these kind of things, uh, these kind of methods we can implement.

During the development of a deep learning model, you have to implement a specific So So here, what I can see in in a, uh, like, forward scenario, uh, the potential issue is given in implementation of the transformer block, uh, that lies in, uh, forward method attention mechanism. And and a typical transformer block, the attention mechanism usually involves multi head self attention followed by, uh, followed by a feed forward layers. However, the provided, uh, code that you you have provided you have written here likes details on the actual implementation of the attention mechanism, specifically with respect to the, uh, multi head attention with the query. The key and value, uh, like, matrices, which is essential for the proper function in the architecture. Yeah.

Uh, debugging a model. So, um, how do you approach debugging a model that performs well on training data but poorly on unseen data? So here I can see the problem statement. Problem statement um, it's like, uh, our model is performing good on training data, but poorly or nothing means it's basically fit, uh, like, um, overfitting with respect to the data. Um, so with respect to this, that we can implement different methods. Let's suppose um, it's related to deep learn, uh, deep learning model. So with respect to that, we can implement dropout. We can implement data augmentation. Um, we can, um, we can change the we can basically less out the number of neurons with respect to layers, and we can less out the number of um, layers of number of layers from that deep learning model. And apart of that, um, we can use leaky ReLU activation type of functions. We can change the activation functions, and, uh, optimizers, we can use um, a part of that, uh, early stopings, we can implement so that our model will not be trained further after us, but some specific scenario. Um, and, um, uh, some kind of, uh, like, um, we can also implement. And even though um, we can do implementation with respect to, uh, like, key normalization things can also implement. And in terms of let's suppose we are working with respect to some regression problem, so, uh, with respect to that, let's um, suppose we are getting overfitting here, so we can, uh, implement some regularization method over there. Um, so we can implement rich or lasso method over there. And apart of that, we can do different tweakings with respect to, um, uh, tweaking with respect to our, uh, model parameters. Let's suppose we are creating a model that is um, using decision trees. So decision number of decision trees and number of, uh, like, heretical layers, we can increase, and um, then, uh, we we can basically, uh, remove the problem related to unseen things. And, uh, these kind of debugging things um, and, uh, with respect to debugings, like, um, that that till now I explained, it's related to, like, how we can, um, implement the things to resolve this problem and to debug on just only identification. Like, what we can do, we can um, check our model accuracy with respect to our training data and with respect to our current data that we are currently getting um, to be, uh, provide predictions with respect to this one. For the testing data, we can check its score and we can check, uh, and compare the score with respect to our training data. If um, the num, uh, the number is very huge. The difference between, uh, the scores is very huge, then we can say it's, um, it's kind of this problem. And, um, sometimes, like, we have deployed this model, so data drifting also there. Um, data drifting can be like, data drifting or model drifting, so these can be the things of scenarios. Um, and here, uh, at using envelopes, we can also implement and and identify with respect to this one, and it can provide us notification over there. Let's suppose, um, uh, we have, um, identified, uh, like, with respect to test data whenever our precision would, um, uh, like, less than this, uh, this much of threshold, then we would get notifications. So that kind of thing, we can also implement. Go to the next question.

How do we approach selecting of loss, uh, when designing a new deep learning model? Okay? So, uh, loss function. So it would basically depend, like, what kind of, uh, problems we are implementing. So different kind of loss functions, we can implement and check. And, uh, loss functions with respect to, like, if we are going through, The regression problem, then, uh, the normal loss functions would be RMSC, MSC, And, uh, apart of that, if we are going through, uh, like, uh, classification problem, then kind of scores would be there, And the part of that, uh, hold on a second. Let me check If anything else I can provide you information regarding this? Okay. I think, uh, that's the thing that, Uh, I can provide. So As I said, like, would depend on regression classification. After that, I would check, like, uh, what kind of, um, like, if If it is related to, uh, classifier, then cross entropy binary cross entropy, we can implement. And, uh, if it's related to regression, then mean scared error, root mean scared error, and, Uh, some kind of even we can create our own loss function as well, like, uh, loss with respect to kind of some segmentation, and we can create our Some old and modified loss function as well. Some scenario we basically use. Uh, and even though, uh, let's suppose in in In current in diff, uh, current scenarios, what we are getting let's suppose we are creating in, uh, diffusion models here. Sometimes we create our Own customized loss function over there.

In the process of training a generative model with TensorFlow, you are defining a loss function. Here is a snapshot of code. And, uh, and, yeah, code is like custom loss function and considering the loss, Uh, considering the losses that generate text to why might this loss function be, uh, inappropriate, and what kind of The loss function should be generated. Huge for this task. So as per, uh, all my previous Questions answered, like I provide, like, we can create our own custom loss functions. So as, like, custom loss functions, we have provided in a pseudo code. Uh, let me read the question again. So here, with respect to this loss function provided, we are basically calculating the mean So you difference between the true end of the predicted values. Okay? And, uh, while the function may not work for some task, if, uh, we And even though some, uh, in the specific scenario that you are currently talking related to generate generating the text, uh, text, basically. So text in recent task often require, uh, a loss function that considers the probabilistic nature, uh, nature of a language. A more suitable loss function with respect to text generation is gonna be like, um, it it could be like, uh, categorical loss, Uh, entropy or specifically when dealing with the large language model, the, uh, loss major difference between the predictive probability distribution of the words And the actual distribution of the word, like, 1 what encoded on all these things. So, um, so with respect to this one, in terms of Text related thing, I would say it's not a very wise of good, uh, loss function that we can implement. So instead of that, using, Um, like, uh, using categorical cross entropy helps the model to learn and generate, uh, the text by penalizing the, uh, derivatives from the actual word distribution. So that thing we can implement.

Uh, imagine you are using a Hugging Face, uh, transformer library, and, uh, you encounter the following pseudo code. Uh, so the code is like, uh, model transformer from pretrained. And here is the model name. And uh, so upon running this code, received an error. External object has no code. What might be the reason for this? So the main reason that I can see and identify just by looking at this one like it may be the cause that, uh, like, transformer model object, uh, hasn't any attribute kind of, uh, from pretrained. Uh, okay, so to resolve this issue, you should ensure that you are using the correct class from Hugging Face. So again, phase transfer library is to do of, uh, transfer more model. You should use a specific model class available in the library, such as, like, we can use, but model g p t two model. Uh, distill BART model. Yeah. This and apart of that any other thing. Uh, like, here, uh, for example, like, we are using BART model. Okay? So instead of transformer model, we can use BART model then a dot uh, from pretrained, then we can use like that. So that's the main thing. Uh, that's the main reason that we are getting error with respect to this one. Yeah. Uh, you just define here, like, transformer model, but uh, instead of that, you should define, like, what kind of transformer model, uh, you are currently implementing and using, uh, like, bug model, t five, or anything. So that's the

Uh, how do you handle the challenge of integrating state of the art generative models into, uh, legacy system. So, uh, so different, uh, scenarios and different, like, challenges we can implement, uh, with with respect to generative AI models. So firstly, let's pick up the generative AI models in terms of, um, let's suppose we are using we we recently created an diffusion model. Okay? So with respect to the diffusion model, So we have basically created a diffusion model that is generating images from the text. Okay? So different challenges that we can, uh, implement and take a like, uh, maybe because handling a diffusion model or handling any other generative model. Um, firstly, our, uh, we have some specific or some limited amount of GPU memory available here. And when we are going to influence with respect to our model, so it's, uh, going to firstly reserve or keep, uh, the complete model into its memory. Okay? So here can be few things, like, let's suppose we have deployed our model into a Flask framework, And then later on, what we are getting, like, we are getting 4 to 5 request in a time. So, uh, when it's going to influence 4 to 5 request in a time, it's surely, it's going to provide us, uh, error with respect to, like, out of memory related things because, uh, of 1, um, GPU. Let's suppose we are using 16 GPU of GPU, and our diffusion model is already taking more than 12 GB of GPU before containing, uh, the train model, uh, into its memory. So on the time of in print, sometimes, uh, let's suppose it's inferencing with respect to, uh, one request, so it would ten, uh, from 12 GB to maybe 15 GB. So with respect to one request, it's going to its peak value. But if we are going to send multiple value, then it's going to out of memory. So here, we should, uh, like, queue our, uh, request with respect to this one. And a part of that, uh, like, implementing MLOps implementation with respect to this one is also a hard problem. Like, we, uh, we should have a sufficient, uh, kind of machines over there. Uh, like, we can easily upgrade, uh, our GPUs over there. We can easily create different replicas with respect to GPUs. So these are the some specific kind of, uh, like, uh, challenges we basically face. And many times, we face challenges with respect to library is related thing as as well. Let's suppose, currently, we are using PyTorch related thing, uh, PyTorch. Let's suppose two point zero having CU 11.8. So maybe with respect to this, our model is providing good thing, but, uh, but maybe there is some changes or update with respect to Python. Later on, it's not providing better better scenario, better, uh, like, results and it start providing errors because we just implemented some, uh, system updates, or we just implement we just install a few libraries, and later on, we identify our, uh, like, library has been updated, and now it's not, uh, not incompatible with respect to CUDA ourselves. So these kind of issues we, uh, also phase with respect to this one. These different kind of challenges we also face. So let's go to the next question.

Design a system to identify, to identify and handle data anomalies in real time as new data is uh, just trade for training your model. Okay. Uh uh, uh, so, uh, with respect to this one, we can, uh, our pipeline basically should be capable of handling various data formats and sources and volumes. So, like, real time monitoring should be there. Real time monitoring the sense, implement the monitoring mechanisms to access incoming data for anomalies at it, uh, it enters the system. So firstly, before entering any data into system, we firstly take a, like, whether we are getting, uh, anomalies over there, and we should implement different, anomaly detection system with respect to the data. Uh, and a part of that, uh, we should implement different rules having some thresholds. We should implement some data handling and correction part over there just before sending the data into our system and into our, uh, model to be trained on. And after that, we should implement some feedback loop related to the model, uh, improvement. So these feedback can be related to either manual feedbacks use, uh, by providing some some kind of QA or, uh, different persons or some, uh, as I said, like, human feedback or even though, um, some kind of, uh, feedback looping with respect to, uh, some different scenarios are in rules as well. So there are features for that, but, yeah, we can implement that thing as well. And even though, uh, we should implement uh, documentation and logging with respect to our data and our systems. Like, there should be, uh, we should maintain and comprehensive logs and documentation regarding detected anomalies and their handling procedures and then how, basically, they are uh, impacting with respect to our model training. And a part of that's, uh, like, uh, we should implement compliance and governance with respect to that. Like, data governance policy should be there. And, uh, Yeah. These things we can implement. And a part of that design a system to identify. Yeah. I think these things we can implement. And even though let's suppose, uh, let's suppose after some specific, uh, time, uh, let's suppose after 1 month, we uh, we are seeing some, uh, model drifting or data drifting then what we decided, like, we are, uh, like, going to retain our model with uh, to previous data and the new data with respect to which our previous model was not trained. So at that scenarios, we we also implement, uh, like, data anomalies with with respect to this part. And then we we should send this data to between, and we should implement data clipping as well, and sometimes we should implement, uh, like, uh, we should implement, like removing those data points. And even, uh, in real life scenarios, what I see, uh, let's suppose, uh, we are creating an a classification model, and it is basically having 2 classes, and here, uh, imbalance data we currently have. For 1 class, we have already 10% of data and other 90% of data we have for another class. So let's suppose what, um, minor class, uh, we have different data points where, uh, few data points are having, uh, like, anomaly values. So here because we already have less data, so we should not delete those those data points. But with respect to, let's suppose, uh, majority class, uh, data points if we are having few data points where we are having anomalies with respect to data. So here we can delete those things. Uh, or, um, but normally, we should implement clipping with respect to what it

Propose a approach to fine tune GPT model, specifically with respect to client domain specific language. Uh, so, uh, here, with respect to this one, like, what things we can implement? So firstly, we should prepare our data Prepare our data in the form of, uh, in a specific form, like, in a special one which is being accepted by GPT 2 model. And then later on, uh, with respect to this one, we should, Uh, we should choose, like, what kind of model training we are basically doing, uh, after doing some kind of data preprocessing. In terms of data preprocessing, we should, uh, like, do some cleaning with respect to our data. Uh, we should remove some noises, some relevant information, uh, or any inconsistency. And after that, we we should implement with respect to this one. Although we can implement some as well. And apart of that, uh, we, uh, we should firstly choose, like, what kind of variant and pretrained of What exactly the variant or version of, um, g p two model we are going to fine tune, and then we should Uh, implement the process of fine tuning. So with respect to the, uh, this one, we should utilize the trans although we are going to utilize the transfer learning for fine tuning, We are going to use the EFT, uh, implementation like we are we are not going to train the complete model weights. We are just going to, Uh, um, train with respect to some specific scenarios with respect to our data points, and we should implement, uh, some Training strategy. Like, training strategy in the sense, like, adjusting learning rates, adjusting, uh, the batch size, Training epochs, monitoring the models, and, uh, monitoring the model performance through evaluation metrics Uh, specify, uh, the, uh, the metrics which are specifically with respect to our client's domain and a part of that domain, Uh, specific prompting and sampling, we should implement, like conducting, uh, alternative training and, Uh, alternating loops with domain specific prompts, um, an example of inquiry, uh, example of different prompts and different scenarios, uh, which would basically provide some, uh, like, which would increase our model to provide outputs uh, which are, uh, actually being provided as an examples of whenever we are sending our data to be trained on. And a part of that validation and hyperparameter tuning, we can implement. And, Yeah. Um, and a part of that, uh, evaluation and refinement, we can implement and, Evaluation in the sense, like, after let's suppose we have trained our model for We have trained our model with respect to, uh, with respect to our client's domain and client's domain data. So we have trained that, and later on, we can Uh, do evaluation, like, what kind of accuracy, what kind of results we are getting. If they are acceptable, then good. If they are not acceptable, then We should create different scenarios, and then we should refine further with respect to our model, and then we can, uh, like, uh, deploy our model and then do can different testing. And even though that kind of, like, uh, you we can implement MLOps, uh, which can auto train our model with respect to some specific scenarios. Uh, like, after a few months, we it would go to train or something. So these kind of things, we

How would you validate the output of a part based sentiment analysis? Would you ensure its accuracy? So, uh, like, we have, like, uh, as you're saying, Specifically, how we validate the output sentiment analysis. So we have created a bulk model, and we have fine tuned it with respect to our model. And we are fine tuned with respect to Sentiment analysis. And now how we are, uh, going to ensure its accuracy? To validate the output of, uh, BERT based sentiment analysis model and ensure we, Uh, model analysis, we should ensure these kind of accuracy, like, uh, these kind of like, things or, um, like, scenarios we should implement, Like, uh, just a second. We should basically calculate the accuracy of our dataset, uh, we should calculate the accuracy uh, of our model with respect to our test dataset? And, uh, with respect to that, like, like, on the time of training, we have already did Train test split and training data we are providing to the BERT model to be trained on, then we are, uh, calculating accuracy with respect to our test data, And then we are basically performing some matrices over there. Like, uh, it would basically depend on business case to business case. Uh, in some business cases, Uh, we do prefer with respect to precision and some recall, and sometimes we just do focus with respect to f one score And, uh, to, like, evaluate the performance or to check the performance metrics. And sometimes, Like, we just see the confusion metrics and do some identification and evaluation with respect to that. And, uh, a part of that cross validation techniques, we can also implement, like, Uh, we have created different, uh, like, folds with respect to our data, and then later on, we can check validations with respect Two different folds and the best accuracy model we can take, and we can, uh, deploy that thing as well. So that kind of Thing we can also implement, and a part of that, part of that? Yeah. Continuous, uh, model deployment and training that we normally use in MLOps. So that kind of scenario, we can also implement. Like, let's suppose, uh, we just trained our model, and after 2 weeks, we are getting Some drifting with respect to data and model, and then later on, we can auto fine tune, uh, like, auto model and, uh, can compare its accuracy with respect to previous model and auto deploy that. So So, yeah, these kind of things that, uh, we can implement. So let's go to the next question.