Large Language Models Bootcamp- Information Session

Key Takeaways

just make sure we are okay we are live okay thanks everyone for joining we are going to go ahead and get started my name is rajal I'm the chief data scientist and the lead uh one of the lead instructors at data science dojo and we are going to talk about our large language models boot camp today so we have been um around for quite some time of the uh earliest companies in data science and machine learning upscaling um a lot of graduates lots of companies on our portfolio um Global presence and uh we also happen to be the first boot camp on large language models and generative AI in Industry so we are going to only focus on the large language models program today so basically uh the idea of the boot camp is uh um is allowing people to uh give them the opportunity to learn the skills that are needed to build Enterprise applications using large language models and generative AI when we go and use chat GPT Cloud uh Bard um they work great you know uh great tools but there are some challenges when we start building Enterprise applications um I think the uh from our experience the foremost uh challenge that we have is the Regulatory and um governance challenge um many companies we have a customer uh we have a few customers actually who are using uh an llm uh product that we have built think of that as your own chat GPT your custom GPT deployed within your own uh tenant um and a lot more uh you know a lot more more than what chat GPD offers um we have built that and uh the customers are using it the foremost reason customers come to us is primarily for regulatory uh and uh and data governance and AI governance type challenges when you upload data to chat GPT if you allow your employees to use chat GPT there is a chance that uh uh that your your intellectual property maybe some pii your HR Etc they might get um exposed and in worst case scenario it may be used by some of these companies in building uh some of this models and you know a whole set of issues can actually result as a um as a result of that but there are other uh other issues even if you are able to build these applications locally uh which anyone can of course but there are other issues that you have to worry about right so now you have built this application internally within your organization how do you handle all the issues around you know the context window and how the token usage how do you build an application that is uh cost efficient uh how do you worry about the the latencies of responses how do you worry about hallucination how do you worry about um you know the correctness of response and correctness is uh in a classic machine learning uh setting correct correctness could mean well uh if it is a fraud pred if it is an actual fraud the prediction should be a fraud um correct prediction in that case if the actual uh transaction was uh not a fraud your model predicts it to be a fraud prediction fraud uh transaction then that's a wrong prediction but what does correctness mean in the context of uh a large language model um and there are different evaluation approaches uh you know you would worry about you know whether the answer is complete or not whether it has all the factual information that was needed whether um you know is it giving too much information even if it is factually correct is it giving all the factually correct information but not complete so uh there are so many other nuances uh that you have to deal with when you're building an application that has some business implications uh whether you should use open source model or a close- source model and how how do you deploy them how do you use them um how do you actually deal with the the knowledge being outdated how do you deal with the brittleness of prompts and so on so there are a lot of business challenges that you run into and uh in our boot camp we handle all of this we actually have designed a very comprehensive curriculum uh and the curriculum has been uh designed by people who actually built um who have built these uh products uh Enterprise products that are being used by some of the you know 50 plus 100 plus billion dollar uh companies uh globally so um so this is not your uh Hey clone this get repository and build a rag Pipeline on a single PDF file you're talking about you know an infrastructure that can handle um 10,000 files being uploaded and at the same time or 10,000 and or pointing it to an API and building a rack pipeline or um you know worrying about all the latencies and the how the architecture should be um how to set up hybrid search how to set up U um uh keyword search how to set up only sematic search uh how to put put guard rails and so on so I will talk about all of these but it is one of the most comprehensive curriculums there are Great Courses out there uh but the courses are in Pockets there's a one course here on Vector databases one course here on embeddings another course here on fine tuning we cover the entire stack um in the the best detail possible and I will expl I will show you the curriculum I will show you some of the content that we have to give you a sense um why we think that uh after attending this boot camp we are very confident that you will will be actually able to build an llm product uh within your Enterprise whether you are cons you could be a consulting company um providing uh building llm applications for other people llm solutions for other people or you could be a Founder for a startup building a product you could be um engineer a Dev manager who's building this uh internal capability or uh you could be uh a technical product manager who's actually guiding these products and building um and helping in with the devs uh you want an in-depth understanding as a product manager as a technical product manager to be able to build these uh applications so we'll talk about all of that in a moment uh so the the boot camp goes like this definitely a very comprehensive curriculum all the mainstream uh llm tools libraries and packages uh will touch upon all of them with practical EX exercises um and then there is a comprehensive project uh with mentoring support at the end of the day uh at the end of the week it's a 5-day uh 40h hour boot camp we start uh I will explain the entire curriculum how do we how we actually go through the curriculum one by one uh you just have to bring your laptop uh what we have done is uh when you set out to learn uh when you set out to learn uh um how to how to do machine learning or how to do uh build predictive models in this case how to build llm applications there are a lot of moving Parts uh you may need some GPU clusters you may need um endpoints for an llm uh you may need um some kind of uh uh uh infrastructure to actually run all the library all the all the code samples and Etc so uh what we have done here is we have set up an infrastructure that is training ready right so you bring your laptop as long as you have web browser will take care of the uh of uh everything else and uh for any cost in terms of GPU clusters uh that you need for uh for our fine-tuning exercises anything that you may need for um uh for any uh token costs for uh for inference uh anything related to other exercises and other python code samples uh that you need you do not need to install anything we have pre set up environment um that you can come in and start using and hit the ground running uh so these are our partners uh thanks to some of the partners I mean they're providing us uh resources uh they are providing us compute resources and in some cases uh we have instructors from these companies who are coming in and teach ing uh at the boot camp so so the next uh boot camp is happening in um it is actually starting on Monday on December 2nd so these are some of the partners in our ecosystem some for the next boot camp in in the December boot camp these partners are going to be involved and in the past we have had these Partners I mean depending upon their availability you will will invite speakers um talking about real Enterprise uh use case is um for the respective uh problems that they are solving in Industry um so I think what I'm going to do here is um I'll go over this uh this uh typical architecture of an llm application and then after that I'm going to go through the curriculum um so if you look at this here um uh the at the at the Cod of it the very um um the the the core component of uh uh an llm application at the heart of this llm application is the the large language model right so it could be um you know the anthropics of the world and it could be the open AIS of the world you know gp4 gpt3 or gp4 Omni it could be llama 2 llama 3 llama 3 one um so you need uh someplace to host them you could be self-hosting them you can get a hosted model uh in Azure in AWS in data breakes but that's once that is in place you need a lot more than just the llm uh it turns out uh that uh I often actually tell people that you still it is still 95% uh soft engineering ai ai is there but you need to understand the concepts but you still if you a good software engineer um or if you're a good software uh you know uh technical product manager in general you should be able to actually build these applications uh you know the and U really the uh AI part is in most for the most part it is used as a blackbox and I will explain what that means in a moment um if you're using these llm applications the uh another uh another component or maybe another significantly important component is a vector database uh you build a search index a semantic or a keyword search index on some cases hybrid search index you build that search index uh and then combine your llm with that that uh that search uh framework commonly known as retrieval augmented generation we go and we will go and talk about rag um uh or retrieval augmented generation in quite a lot of depth uh when you push your documents into a search index very often you uh very often you need actually to um if you're building a sematic search engine you need an embedding model uh so an embedding model is also an a critical component by the way everything that I'm talking about um the reason I'm talking about this is that uh the and in the boot camp you will hear the theory of whatever I'm talking about in quite a bit of depth and then also the practice of everything so in this case the embedding model if you uh so you you convert your documents into embeddings embeddings go in a vector database and Vector databases have both you know your standard uh search functionality on top of uh um the sematic search functionality using embeddings and then you have have other things like Lang chain how do you how do you build a multi-agent system how do you build uh the concept of memory and understanding of appreciation of history in your llm application uh how do you um make an API call from within your agent so you want to connect to the Google Maps API you want to connect to Bing or Google uh search API how do you do that um how do you create templates for your prompts um and so on so that that part is actually handled by Lang chain um and Lang chain actually um has you can see that I mean it is um uh lanin is not the only one I mean we have llama index and you can you can write your own python code as well but you know Lang chain is by far Lang chain and llama index are by far the most popular uh Frameworks uh in this case um then you have to set up observability uh how do you uh how do you log things do you see the traces if there was an error is was there a user feedback how can you monitor that um how do you look at the metrics how do you look at the latency how many tokens are being used how many uh what was the cost of uh inference um um then there are guard rails I mean how do you put guard rails such that a toxic a prompt that is toxic is not actually uh or whether the prompt or the response if they are not uh you they're against your policy how do you ensure that there is nothing against the policy that is happening um then uh there is there is deployment as well uh I can keep going I can spend a lot of time on this but in general we cover this entire stack uh to to be able to uh you know uh we demonstrate um or we actually practice and we learn and practice appli um pretty much the entire stack I mean so whatever you may actually run into while building this these applications so what I'm going to do here is I will give you a curriculum overview I will give you an idea uh within the next 15 to 20 minutes uh what we have uh in our um what kind of content we are covering in our uh boot camp and then uh and what kind of labs we have what maybe show you some lab samples show you some slide decks and so on okay um let me actually make sure that you can see my screen okay so let's actually start with um um so a common question is what are the prerequisites really if you are interested in building these applications uh uh if you work in um you know you have some appreciation understanding of Technology you should be able to attend I mean and it was sound to good to be true but when you see how we have structured this how the labs are um uh really you will uh you will uh see how we have done it how we have organized it and then um and then happy to take and answer any questions so before you come in uh the boot camp uh the language of choice uh is python so if you are not a python programmer we have uh some a two to three hour tutorial that you uh like very highspeed tutorial where we up you know bring you up to speed with the basic python Concepts and uh once uh we um uh once you have done that uh when you come in on the first day of the boot camp we start on Monday this um in uh this case um this uh the next boot camp is starting on Monday December 2nd and um the boot camp is going to be uh some people attend in person in Seattle other people they are uh located remotely um so some people they attend it remotely from you know wherever they are from the comfort of their home uh so you can enroll in the same boot camp the same interaction you know you can see people in class people in class can see you um you know sometimes it ends up happening that online attendees they tend to be actually far uh you know sometimes they are actually more engaged than the inperson attendees I mean there was there has been one case but in general it's a very interactive experience uh so we start on Monday uh on on Monday we start with the general high level overview of uh the entire ecosystem uh what is an llm uh um what is a prompt how does a prompt work uh on an llm um very simple idea when why do you need the bings I mean what does an embedding mean why do you need a vector database and um very you know high level um explanation of why do you need a vector database why do you need embeddings uh then we go and talk about you know the risks and challenges when you're building these application how do you customize your model uh using fine-tuning rag or some kind of hybrid method um and in the within the first two to three hours you have a a fairly good understanding of the end to end ecosystem of you know what does it look like what does it mean to actually uh use an llm application once we have done that um we have this session on uh uh the attention mechanism text embeddings and Transformer architecture um and the way um you know we do this is uh you can see that uh you know we start with so even if you don't don't have any background at all we start with the the very basics of why I mean you know what is discriminative learning what is generative learning um uh then um uh what is a Transformer what does it mean to use a Transformer and then we slowly gradually actually explain the whole idea of attention mechanism we talk about attention mechanism and we build uh build it very slowly but surely we build this uh this whole momentum where you um from intuition to actually uh you know you are able to understand what actually this whole Transformer and attention business is about so you can see dot products Etc of course I cannot see tell you I cannot explain everything to you but you can see that you know we talk about you know multi-head attention here you know what does similarity mean um you know what are similarity matri and matrices and really get that geometric uh geometric and uh you know algebraic interpretation of you know what does all of this really mean um then we have practical exercises uh then we move on to Vector databases once you have the embeddings what do you do with those embeddings well we store those embeddings in uh we store those embeddings in a vector database so what is a vector database so if you look at this uh you know we go and uh start with drawing parallels between standard Vector database to you know uh your sorry um uh your relational database uh how are vector databases different from relational databases so if you look at this what does a query look like uh and what does a vector database really do and then we talk about different types of searches you know Vector search IC search Hybrid search filtering um and why it is needed uh we talk about different indexing approaches most notably hnsw I mean how do you actually uh when you have hundreds of millions of embeddings in a vector database how do you retrieve these embeddings uh as quickly as possible so you can see that this this session itself more than 100 slides here uh in this session we um we spent about three hours explaining um how does a vector database actually work very detailed session you know ideas around chunking what happens to chunk uh chunk uh individual chunks how do you uh you know how do you add meta metadata to your chunks I mean why is it important to um add metadata uh to different chunks and so on um then we uh move on to exercise if you look at this here so um we do six to seven exercises here so we start with a simple Vector search what does it mean to do Vector search right so well um now you will click here and um everyone uh each of uh the attendees gets their own instance of the this lab environment uh any changes you want to make any new code samples you want to write uh create um you have persistent storage uh with us uh in the our lab environment uh now you look at this uh you know I'm creating a data set importing the data uh you know querying the vector database and so on now then we go to similarity search then we go to hybrid search and hybrid search is that when you're combining both the keyword uh your standard traditional search with the uh with your uh schematic search right so if you look at this uh loading some of the questions if you look at this here I am you know I will go and open this file uh let's look at this here here I am going to here so we load this file a small data file we import that uh I will import it in a vector database and now we'll set up a hybrid search engine you can see that there is you know we are showing you different query parameters so I think I can run it I hope this and then you get the an running instance of a vector database right so you don't have to worry about setting up your own Vector database yeah I have to let me see did I run this okay so I think this uh uh vvr server is not running this is a vector database I mean if this was running I would have been able to show you but you can see that uh What uh uh what we do is we give you these examples of uh how do you per perform a hybrid search how do you um how do you search on a specific property how do you create filters and facets in your uh Vector search index um then uh generative search and then you have Vector compression what if you have too many vectors and uh compute uh and uh computational resources they become uh challenge so how do you compress your vectors while maintaining accuracy then how do you set up semantic caching just like you have regular caching uh in software uh semantic caching is uh um in scenarios when you uh when you expect that very similar queries are going to hit your search index how do you uh how do you make sure that you're not actually generating those tokens again you cash the results U what is the capital of France uh and um tell me about U um what is the capital of France that's a question um what is the name of the capital of France right so these two queries they are the words are different but the intent is the same so semantic caching actually you know helps you uh helps you save on compute and infrastructure cost inference cost uh by caching the uh similar results um then uh we have uh we set up a very very basic now uh people are um uh when people are um by the by this time and this is day two by the way we are already in day two I should have mentioned so by the by lunch on day2 uh you are um you're very familiar uh you're very familiar with the idea of rag you know the general overall ecosystem and now uh verba is uh you know think of this as a very basic uh a rag application so you go and build this rag application uh using Vera um and we give you uh you know the infrastructure where you can actually deploy this ver application um now um there is uh uh so vb8 is our partner here um in in this you know helping us with the the curriculum and then also delivering the curriculum and also providing the infrastructure then uh neo4j they are also our partner in uh this this whole thing then we talk about how can knowledge graphs actually help you in um how can knowledge graphs help you in improving um rag so we have a session on graph rag uh Adam Adam CI from uh neo4j um he actually talks about uh uh he talks about you know the the whole idea of graph rag so and um U we finish that um this uh uh this discussion on how can we improve um how can we reduce hallucination using um graph rag um and then uh we move on to let me actually go here then we move on to Lang chain uh we spend almost I have to say we spend about a day on lch six to eight hours and uh we uh cover Lang chain in um as much depth as possible uh you know I'm not sure how many of you are familiar some of you may be familiar the fact that you are here so I'm assuming that some of you are already familiar so we start with model iio uh we talk about uh uh we talk about why do you need templates why do you need example selectors right so in this case for instance you know if I launch it you will see the lab comes up so every I mean I cannot click on each of them but uh you know basically that's the idea the labs are already running all the dependencies ETC they are already installed and uh what we do here is um uh we get we discuss why prompt templates are needed once you understand why prompt templates are needed well uh here is how you create a promp temp right so and then if you look at this I'm going to run run run uh in even even if you're not the best of software developers if you know you know if you know how to read this we ex we actually discuss what does chat open API API mean what are the other libraries that are there what is promp template right so you know so what are some other libraries that are available right so if you look at this you can see that we go through them really explain in a manner that it makes sense to even some users who are somewhat non-technical right so uh and then you know um well this is creating an a connection or an instance of a GPD 40 and then I am you know running this prompt template and then we talk about uh uh we we talk about what does this mean what does this total tokens mean output tokens uh you know what does this um you know this whole response mean and how are you going to build an llm application using this and now um imagine now retrieval how do you connect your application with different types of data sources Google Bing Dropbox SharePoint Salesforce HubSpot um really any possible data source how do you do that how do you create chains such that you know your uh llm application it is a series of function calls as opposed to a single uh function call uh let me go here maybe I can show you one of one more example from chains here it's coming up uh um and I'm going to run it and see that I'm actually now using chains here the same thing this API key that I'm pasting you everyone gets an API key for their usage during the boot camp uh uh and then we run it uh and then if if you look at this this is uh an example uh that we have created a chain first you ask if we ask the user okay you know which suggested area uh and uh we have set up this prompt template where uh we talking about your your job is to come up with the classic dish from the area that the user suggest and now the next part of the chain is given a meal give a short and simple recipe on how the how to make the dish uh at home and the third part of the chain is given the meal um suggested dessert that goes with that meal and then when I go ahead and run this basically we we go we go through this um we go through this entire uh we go through this entire uh sequence and explain um um you know um one more thing is a lot of these things um uh uh a lot of these things can um um you know we can go through these uh this process but then we go back and show you how how we Lear leverage this in a practical application what does it mean to actually have uh have uh These Chains um and uh how would you use it in a real application then I mean this is the router chain and so on uh then we talk about memory how do you build the concept of memory in uh your agent or your llm application how do you create multiple uh multi-agent framework how do you build this agentic framework and uh an agentic framework uh we include we cover Lang graph as well uh so we will be talking about Lang graph uh uh how do you build more complex uh agentic workflows using langra um and uh and a whole lot I mean I'm sure I'm missing something here uh but this is uh you know maybe I can just working with Lang graph uh I click on this and Lang graph is going to be you can see that we're talking about the agent State and nodes I mean how do you create these nodes uh and for those of you who are not coming from a programming background uh but I mean are coming from product management and the background I can assure you that most people uh are they are just fine because the code is already there you just have to understand how things are flowing but for those who are more coding Savvy they should be able to actually they should be able to uh uh modify this and build application so one thing is that I mean if you're a product manager you will leave as an llm and gen product manager uh but if you're a product manager uh professionally you're not going to leave as an llm application developer right so you will remain uh the role that you have will remain the same but now you have added a very good thorough understanding of uh llm and generative AI landscape um so this is a common question but the the amount of code that you see it should not deter you from uh from attending the boot camp because really I mean you have to run run run if you look at this I'm I'm essentially just running this uh this code uh we we discuss the code hey what is this doing you know what is this you know tick token encoder what does this do and so on so we we actually explain the code uh and as long as you understand the the that part I think you should be okay um then I will go back Lang chain is done we talk about you know building a basic semantic search engine uh we talk about really the challenges in advanced rag I mean how do you deal with the um when you're setting up a rag application how do you really um deal with different challenges in rag I mean so how do you deal with chunking how do you set up um you know chunk augmentation how do you uh understanding intent uh um how do you mitigate huc hallucinations uh you know how do you decide how many chunks should be how many chunks should be there um then uh you know what embeddings and how big should the embeddings be whether you should compress vectors or not so there are a lot of nuances there uh that we talk about in a lot of detail uh we also uh talk about fine tuning um in fine tun tuning we um we go through uh the idea of uh uh you know fine tuning transfer learning indepth quantization low rank adaptation qora which is a combination of both uh and then we uh we give a you know everyone gets a GPU cluster where they can actually find T A Lama to 7 billion uh model forbit quanti model uh and we compare uh the one with the um with unfin tuned or with a a model that is not fine-tuned um um we also go through the details of evaluation evaluation is something um that may not come AC come in as uh you know the first thing to learn but it is incredibly important to for you to know how to evaluate uh your uh large language model application so we talk about we start with the standard machine translation uh type metrics um all the way to you know things like perple perplexity score and uh you know ragas you know metrics around um metrics around U you know evaluating your context and evaluating the generation responses uh I think I'm uh there was something that I want to say that I'm forgetting now yeah and we also cover the data sets and uh which data sets are best in What scenario so for instance if I go down to evaluation if you look at this once again same thing practical exercises right so because a lot of this material I mean uh one YouTube talk here another deep learning AI course there you know one udm course there I mean but the a lot of these things are fragmented but what we have done is everything is coming together in one week you have a very solid understanding of what uh what um what does building an llm application entail um so I will show you just one example here if you look at this um I'm talking about you know evaluation using ragus and then when I do go here again um the lab is there uh loading and splitting text documents creating embeddings and setting a vector store we go through this step by step and the the labs by the time we get here you already by now you absolutely clearly understand uh what a vector store is and you know what is happening here you know what a prompt template is you know what a chain is um and so really uh initially we take it slow but later on during the week uh you can see that a lot of things they are coming back and you're putting all of the context together and if you look at this here uh we have uh buties uh we have uh you know uh we have created this uh uh lab to solidify your understanding that you the things that you learned in theory now you're applying it uh you know what is context position what is context recall what is faithfulness what is answer relevance see and so on um we uh talk about observability and guard rails uh and monitoring of your models uh how do you how do you put some tracing and logging and uh monitoring on your uh llm application we talk about that in quite a bit of detail then uh we uh on the last and we also talk about deploying your fine tune model um that is another session that we have and then finally on the last day of the uh of the boot camp we uh show you how to build your own llm application using um you know using different tools and then once again um once again um you know that part is that we give you boilerplate code and from there you can actually um you can get started and then you have exercises now add this chain add memory to your application and do this uh make it multi agent make it agentic and so on so let me go and see if there are any questions okay uh okay Harold your question is how much to attend the boot camp you're referring to uh so um the boot camp is $5,000 but I think there is uh depending upon availability depending upon the seats we have remaining I think right now don't quote me on that the price at at the moment the price is around 3700 or $4,000 I'm not sure please check the website I think uh go to the website and uh the team can actually post a link directly or set up a call with us and we will be able to let you know then there is more questions uh what is the cost of training if we are joining in from India uh there's no difference in cost because I mean our cost of doing business is the same but happy to talk to you Manny uh please you know feel free to set up call and then happy to actually talk to you uh okay uh uh yes the uh recording is going to be shared uh for sure okay um let me see okay there are some comments there can you share the uh link to the Jupiter notebooks no the these are behind a login we cannot share the links to the Jupiter notebook but if you're looking at something specific we are happy to you know if you can set up call we we will be happy to actually show you um uh the notebooks uh let me see do we have uh and there is in the Q&A tab I think Carol I have already answered your question um it's uh okay so I think let me see what I have remaining here uh we have gone through the curriculum already as I said I mean knowledge of uh Python Programming you can just go and uh you know uh um if you if you have some background in uh python you should be able to actually um follow everything in the cont uh content uh very easily um uh this is the roughly the technology stack that we cover uh during the boot camp for different exercises so the final uh project for instance uh it is deployed in hugging face and uh I believe it is streamlet but we do use some hugging face there as well um and you can see that this is the roughly the technology stack uh and this is just you know the high level idea what what is the content uh what is the content that we are uh covering so um and then we also have sessions around you you know our own learnings while building a product I mean what should be the architecture how should be how should you set up your search index how should the microservices be set up and so on so it's really uh you know U really a very um carefully distilled information um that we share b as a practitioner U uh let me see attention mechanism Vector databases know we talk about all of these things there is sematic search prompt engineering fine-tuning we talk about that then we have uh Lang chain we cover Lang chain in as much depth as possible possible we probably one of the most comprehensive course discussions or treatments of Lang chain out there multi-agent applications we talk about this monitoring the guard rails uh is there uh Advanced rag right so really I mean building a rag application so I saw a comment yes I mean yes you can find this information online but really uh if it was that easy um then everyone would be building llm application right so with all due respect right so um yes I'm uh um yes you can find a lot of this information online but the way it is how it is organized how it is how we connect the dots how the exercises are set up uh you know it uh that is what makes all the difference okay uh llm evaluation I've talked about it uh and building an llm application finally that is there um these are some of the instructors that we uh expect I think all of them are confirmed except one but this is uh what we expect next time I mean here are I I think I'm sure you have seen uh these testimonials I think the team was actually putting these links uh there uh um in the chat window and I'm assuming that uh on the social media if you're watching the live streaming uh but I mean this these are our uh you know Partners uh and customer customers who actually have attended in the past we have had people travel all the way from well as far as Australia and uh the Middle East and Asia and Europe they have been traveling to us many people attend remotely some people they prer in in person uh or in class learning so if you look at this these are all the company these are some of the companies that have already attended our training so and these are you can see Microsoft Google is there Arco is there you know and uh uh Starbucks is there Fidelity is there right so these people these companies uh have actually attended our training um so we have been doing this for uh a little over an year already um the next boot camp uh Seattle and online in parallel uh you know some people attended in attending in person others attending uh remotely um but same learning experience same discussions uh and the same curriculum uh we uh we're making it a bit easier for people who cannot attend in um who cannot travel for one reason reason or the other um okay so if there are no other questions uh let me just give it a moment and if there are any other questions I'm happy to answer otherwise we are uh coming to an end to the uh of the training or rather this uh info session okay thank you so much everyone I hope I will be seeing at least some of you at the weekend thanks