Generative AI Full Course – Gemini Pro, OpenAI, Llama, Langchain, Pinecone, Vector Databases & More - Ep247

our laptop has RTX 490 thatbasically means we are women and scienceour laptop isvery powerful there will be electricitycost involved there will be multiplethingsinvolvedright so everything is over here you canprobably see with respect to this rightit is somewhere around 27k sorry not 27lakh it is 27k as as I just saw 0. I didnot leave that part okay so but you canjust understand the cost is keep onincreasing okay so here you can see thatRSC uses Nvidia Quantum in Infinity bandwhere product cluster is equipped withRoc you can probably seethis C see see see CO2 emission duringpre-training you have to also give thisinformation if you want to publish theresearch paper total GPU time requiredfor required for training each modelpower consumption PE power Peak capacityper CPU device for gpus see how muchcarbon is emitted right 7B is thismuch power consumption 400 watt 400 watt350 watt 400 wat total total GPU hourstake 33 lakh 31,000 no no 33 laks 11,000hours GPhours who has this much time guys if astartup in India will spend this muchtime in trainingdone I don't know this is how many yearslet's say 24 into 12 uh 24 into 365 justdo how many hours will be there how manyyears it has basically trained rightcarbon P for print pre-training and allthese information are basically thereright and then here also you canprobably see the comparison size Codecommon sense reasoning World Knowledgereading comprehension math mlu mlu isbasically human level understanding uhBBH and AGI right now I've have told allthis information now let's understandhow this models are basically trainedhow llm models are trained okay so tillhere everybody happyyes everybody happy with the teachingthat I'm actually doing so now we aregoing to move towards how llm models aretrained and we will discuss it step bystep so guys clear ornot clear or not just tell me give me aquickinformation so here I'm going tobasically write the stagesof stages ofstages of training so first informationhere I specificallyhave I will justdraw the stageone so this is my stage one based onthat research paper I'm basically goingto draw okay so this is nothingbut generativepretring okay generative pre-trainingsecondstage so second stage is nothing butsupervisedfine tuning which we also say it as SFthe same information what is writtenover there that research paper samething I'm writing thirdstage third stage iswhatreinforcementthrough human feedback this is my thirdstage Okay so initially in this stage ingenerativepre-training we give huge data so thiscan be so any any llm model basicallytakes internet Text data or any documentText data in PDFs in all all thoseformats and here we specifically createor use this generative pre-technique now generative pre-trainingbasically means here specifically we usetransform architecturemodel Transformer of bir architecturemodel the outcome of this is what theoutcome of thisis the outcome of this is we basicallysay itasbase let's say if I probablyconsider if I probablyconsider so I will write this is mybase Transformermodel what is this the base Transformermodel okay now this base Transformermodel is then base Transformer modelbasically means whatever Transformer Ibasically trained on I will basicallysay this as base Transformer model okaynow the base transform model is in turnconnected with supervised fine tuningbecause same model will be taken and andsupervised fine tuning will be done ontop of itokay top of it right now this understandthis base transform model will be ableto do various task like textclassification text summarizationmultiple things it will be able to dookay now here only we will not keep itin case of uh llm model we will take itto the next step the next step issupervis finetuning now here what we arespecifically going to do we are going tousehuman trainersalso we are going to involve humantrainers to put some kind ofconversation some kind of conversationand here we will create some more customdata this is important to understandhere we will try to create some morecustomdata right so some more custom data willbe created in this case in thisparticular step and those custom datawhich is created it is created basicallyby whom by human trainers I will talkabout what exactly is human trainer whenI probably Deep dive more into it okaythen it based on this custom data wewill train the specific model and theoutcome of thismodel outcome of the model willbe okay just asecond oops it got closed let's seewhether it is saved ornot I hope so it should be saved oh myGod okay apologies the system gotcrashed I don't know what happenedbecause of that the entire material gotdeletedsad can't helpokay so how much content I had actuallywritten I don't know whether it's thesystem got crashed or the scribblenotebook automatically gotdeleted sad to hear about it but it'sokay I don't think so anywhere itis okay I don'tknow generative AI the materials gotdeleted I'm extremely sorry I don't knowwhat happened over here but I'm not ableto seethat materials got deleted yeah okay noworries anyhow you'll be able to see inthe recordings so don't worry about thatuh let me continueokay let me continue okay okay now let'sgo step by step I was just talking aboutsome important things over there sofirst step I will go with respect tostage one okay so stage onegenerativepre training okay this is basically mystageone now what all things we basicallydiscussed inthis okay in generative pre-trainingwhat we specifically do is that we useTransformer architecture okay so herewhat we are doingwe basicallyuseTransformers Super beneficial for NLPtask and then along with this we takeInternet Text data and document Textdata so this is nothing butinternet Textdataanddocument Text dataokay and this is what is my stageone okay stage one now once we trainwith this specific Transformer we whatwe get we getbaseTransformermodel we get base Transformer model nowwhat this base Transformer model isbasically is Cap capable of right whatthis base Transformer model is capableof you need to understand this specificthing okay this base Transformermodel is capable of doingtask here I will write down all thetask numberone textsummary I will save this saving this isalways better so that if it gets deletedI can open it so the task which is ableto do is like task textsummary sentimentanalysis third task can be somethinglike textuh wordcompletion

(Editor: code)