Evaluating Large Language Models | Community Webinar

Key Takeaways

my name is sonali and I'm the co-founder of Lighthouse AI uh like u s said in the beginning before uh prior to this I was a lead AI scientist at Progressive Insurance and broadly I have been in the space of AI for six years now I'm very excited to present to you today the current and Bre best practices to do quality assessment and evaluation of large language model based applications so let's Dive Right at it generative AI is disrupting Industries across economies as Per a recent survey business and Enterprise leaders remain focused on generative AI ahead of their ahead of other emerging Technologies and 80% think generative AI will disrupt their entire Industries while 20% think it will cause significant disruptions undoubtedly in few years every company will be a generative AI company in some format Enterprises are integrating Ai and generative AI into many different products a recent survey by Menlo Ventures revealed that generative AI is typically being used in COD co-pilots to provide code intelligence and generation data extraction summarization of blong documents content generation building chat Boot and many more some popular use cases are to build co-pilots and chatbooks of various types to assist in or automate many parts of a business a very popular way to do this is via building a document Q&A system the rapid adoption of generative AI in Enterprises isn't big part due to the ease of integrating AI via apis and libraries such as Lang chain Lama index and many more it literally takes only days or perhaps like a day to create a proof of concept version of the application doesn't matter whether you are building a chatboard these days or a summarization app or a Q&A app it takes just few lines of code while it's very amazing to see prototypes being built rapidly it still isn't a guarantee that the app is safe and reliable so going from POC to production with llm apps is not a very straightforward part typically generative a applications suffer from issues of hallucination inconsistency and toxic responses security issues like vulnerabilities to prompt injection ction privacy issues like data leaks misalignment with business rules lowquality long form conversations and many more all of these issues they limit the Prototype from being productionize faster so anme who could be a data scientist AI engineer or software engineer in your company who is now building the products spends close to 3 to 10 weeks in manually evaluating the outputs typically this is done by eyeballing examples after extensive manual validation by smmes and refinement by developers the entire validation process has to be again repeated for each application that is built so while it takes few days to create a PC it takes weeks to get it into production these risks are present for apps built using Clos Source Foundation models like GPT and even open Source models like llama mstell and so on unfortunately in the current state of the current state of guard they do not work well in in real AAL use cases so today I'm excited to dig into these topics with you all first I will share some examples of the common issues present in llm apps then highlight the importance of evaluations then I'll share the several ways to conduct evalu uations and give you some resources related to evaluations and finally we will take a look at the demos and the code so let's Dive Right at it first let's start with the biggest concern with llms which is hallucination you might have already been familiar with hallucination but for those who are not familiar hallucination in case of large language models means llms make up information because they trained on inter large number of Internet large uh data set from internet so they typically uh give answers from their memory in business context retrieval augmented generation or rag is a go-to technique to ground llms in private information um private information of the company and is believed to eliminate hallucinations as it derives the information from the data sources itself is it really the case we tested that hypothesis in a finance context we find that despite using rag or retrieval augmented generation the latest llms today which is gp4 and Google's latest Gemini 1.5 give only correct responses 65% times they make errors such as giving incomplete responses adding external adding extra information that is not present in the context H or h completely or not providing a response at all all of these experiments were performed on a fin QA data set available on hugging P you can go to the you can go and explore this data set as well as it's given in the slide here so rack does not solve hallucinations here is another example of the type of errors that llms make or llm based apps make both GPT 4 and Gemini 1.5 they get confused in Technical and specific domain specific terms here is an example where when asked about physical security both llms responded about hardware and software security very easy to get confused right many other error types are present as well such as Gemini hallucinating information from from its memory because like I said before these are trainer large number of internet data so it is very easy to get information from their memory and not as um as opposed to saying what is actually present in the document another common issue that is seen in generative a apps is of security via prompt injection and jail GDs here are two examples where on the left a customer service chatbot is made to swear by smartly prompt injecting and on the right it is made to give an unrealistic offer so on the left you see an actual example of tpd Delivery Service chat and on the right is the chevrol chat that hallucinated quite easily so what is a deadly combination here you have put models into production without doing the actual Q QA testing so hallucination uh combined with prompt injection is Can it can be a really dangerous recipe for your Enterprises if you do not do the right QA or the evaluations another important problem is that of privacy problems including data leads various news reports have highlighted this after the release of chat gbt and this issue with and this issue with the leading llms leak sensitive information like training data information and many more now that I have highlighted two very important issues there is another very under discussed issue as well who uh and which is very important case of llms and that is called inconsistency to variations what is that so we have identified that when you provide that the same information in different ways to LGE language models such as writing dates in in different ways or rewriting the entire input with different phrasing the responses by LMS can drastically different can be drastically different such issues also make the system unreliable and yes there are two queries uh that and there are two queries that were given to a single drag system but the outputs were very different as you can see in in one of the output it gave the correct information uh that the remaining availability of company's authorized share is 22.6 billion when it was given the right uh when it was given the dates in this manner just tweaking it a bit and giving the dates in a different way first day of April in 2023 and changing the May 4 to fourth day of May 2023 now it says that that uh the information was not mentioned in the provided context information this can very well happen to the app uh to the apps and the Enterprises so with this example you can now imagine how sensitive these apps are to minor variation in fronts okay now that I have shown you uh various different ways to U various various different issues that the llm apps space so let's Dive Right at the evaluation the llm evaluation is an active area of research because of the reasons that I just mentioned and let's see how people are currently evaluating llms and generative apps in the Enterprises meno Ventures did a survey where they found a common way uh in which Enterprises are handling uh handling evaluation is actually uh 66% of the companies they actually evaluate generative a apps manually around 45% have developed some sort of internal tools or mechanism to do uh to to do the evaluations or labeling the third of the most popular way is to collect data directly from the users so once the system is put into practice but you can imagine how um how risky that is fourth the llms are used to evaluate around evaluate llms and that is done 25% of the time but that is increasingly becoming a popular Trend these days llms as a judge lastly the companies use external software or academic benchmarks as well uh as well for the evaluations while manual evaluation is the most popular technique it is it clearly has many challenges it usually is done superficially by ibing a few examples on the other hand if done thoroughly it can be extremely time consuming it is also not clear in a typical organization so who is supposed to do this manual work is it the data scientist is it the developer is it the ml engineer who is supposed to do it or is it a separate QA team which is sitting and which most organization do not have by the way because of all these issues validation is often skipped allog together and either the apps that end up serving for the internal team or the employees or if they make into uh production or if they're going to customers it it it has a lot of risk it can have legal risk brand risk or um uh many other type of risk that I explained before so skipping evaluations and quality assurance can actually land companies in hot Waters the challenge with evaluation and quality assurance of a fullblown production application is that it requires a multifaceted testing the evaluations need to be conducted for hallucinations definitely for hallucinations rack components inconsistency to input variations what I showed before and presence of toxic inputs of topic responses security issues privacy issues fairness bias and in the future world of AI agents um evaluation of function and API call clearly the popular practice of rag evaluation is only a small subset of multifaceted evaluation that is necessary and that is what the Enterprises are doing right now so when working in a Enterprise scenario it is actually not sufficient to Simply rely on academic benchmarks for the performance of llms you might have already heard like when a when a model is released into the public like say Mr or Lama or or what not they are actually benchmarked on academic benchmarks and these really um and if you go to the leaderboards you can see that they are uh outperforming one after the other um uh the benchmarks that is there but can we actually rely on academic benchmarks when we are developing uh apps uh for the Enterprises it's actually not the performance on academic benchmarks and real world use cases they're typically not aligned even consider the finance example that I shared of few slides ago the latest llms get only 65% of accuracy despite their performance being close to perfect in many academic benchmarks so you can't really improve a system what you can't measure so let's take a look at the uh landscape of eal techniques right now we can map the eal techniques on a scale of how much manual effort is needed so let's draw a line and put manual on the left and automated on the right the academic benchmarking techniques lies somewhere in the middle since the Benchmark tests are typically created with some manual work they're very high quality data sets by the way but the applicability to Industry uh use cases is very less the second one is actually the manual uh human evaluation methods that can be mapped all the way to the left and the third is a fully automated method which is llms evaluating llm that are mapped to the right finally there is one more it's called human in the loop or human plus AI Loop technique that leverage best of the both wordss let's look at both at the pros and cons of each of these techniques and how to use them let's start with number one which is uh academic benchmarks the advantages are that they are large scale and are already publicly available since these data sets are precured in Lab setup they have the golden responses many of these benchmarks have many tasks and scenario setups however some of the biggest challenges with academic benchmarks right now is that they do not match the Enterprise data or task Academy benchmarks are created with public tasks and data which is not the same setup as within the Enterprises thus the performance on academic benchmarks does not correlate or does not translate one-onone to Performance in in Enterprise data and in production um so there are few benchmarks that are very popular and um the um the large language models or the foundation models uh they are typically evaluated on these benchmarks uh some of them are Helm glue and superglue so Helm is a recent Benchmark developed by Stanford researchers uh and Helm stands for holistic framework for evaluating Foundation models it's pretty large actually and it is a very like it says it's a holistic framework so um it covers around 119 models 116 scenarios and 110 metrics these scenarios they include question and answering information retrieval summarization language modeling text classification and other tasks this is because the large language models that are developed they should be capable of Performing old all of this St the second technique which is most popular uh is the uh which is the most popular eval technique is actually a human evaluation there are several advantages to use a human evaluations you can actually hire subject matter experts or smmes to do the evaluations or you can make your team your data science team or your ml team do the evaluations since you're interacting with the evaluators you can train them and create a trustworthy labeling process there is only a onetime effort to set up The annotation pipeline however using human aice also has certain challenges as you can really imagine it can be very slow and cannot scale to thousands of models or systems let's say from uh today let's say in five years you are deploying um 10 you deploying hundreds and hundreds of models generative air models and you are doing them you are doing eval for each and every model manually how Compass some that would be and it would be very expensive by the way both in terms of so expensive in terms of time and cost evaluators so the evaluators may not perform consistently and there can be inconsistency between the labelers finally training and managing smmes uh to do the task is actually a very time consuming process so if you decide to leverage human EV human evaluations there are few uh choices of tools to use some popular ones are Prodigy label studio and ARA the functionalities among all of them are pretty much the same you can upload the data that you need to be evaluated and then the evaluators will do do the task by the platform so let's see some examples here is an example of Prodigy showcase in a text classification in a correct or wrong setting you can give let's say like this is U text classification thing is cream cheese really good in mashed potatoes and then you can classify whether this is a recipe or this is a feedback or a question similarly arila is actually open-source Library available on hogging face and here is a screenshot of what arila does to rate Q&A chatbot and it does uh it does rate the quality of the record from 1 to five and it collects the reasoning behind this papers so once uh they get the feedback they can improve the system so like I said before if you cannot uh evaluate you cannot improve the system so this is one technique to improve the system by collecting human feedbacks label Studio on the other hand it provides a pipeline for you to collect human evaluations and uh annotations and use them to build an rlf pipeline to continuously improve your generative AI system the third evaluation technique are automated evaluation which are very quite popular these days by the way uh here an uh llm based system is used for evaluation this technique is practical since they allow automated evaluation with no or very little manual eff effort and is scalable to many tasks but one question that always arise is who is actually evaluating these llms and is it accurate popular Frameworks include Ras lsmith arise Arthur and trence now let's take a look at the evaluation criteria for rag evaluations first uh it is divided into three types first is the similarity metrix it can be word overlap how the word how the words they overlap between the ground response and the rag response next is the blue score and the root score and remember that all of these uh the scores that you are seeing here there need some ground responses so we need to need to have some ground response before using this metrix the second category is the semantic matx first it does the embedding similarity uh embedding similarity the similar it gives a similarity score between the ground Pro response and the um and the response from the rap I'll actually show this in the in the demo that I I have saved it in the end and then coherence and consistency how uh coherence and consistency with respect to the context that is there the Third type is the llm eval metrics and uh specifically uh this is for Ras which says if the answers are correct if there is respectfulness if there is context relevance and context Precision Precision and recall the last two um these can be done if you have ground TRS if you have some ground TR loers and here is an example of how to use dragas tool to do the automated evaluations they have uh four uh key metrics uh faithfulness um answer correct correctness context uh precision and context uh recall uh context precision and Rec and context relevance like I showed before um yeah and it can be just few lines of code and then it can be used to evaluate finally there is one more uh it's called human in the loop or human plus a eval technique that leverages both best best of both the world evaluations are highly trust evaluations here are highly trustworthy more than automated evals due to the involvement of human or who can provide oversight to this evaluations AI optimizes the work and reduces the manual efforts making the evaluation process faster uh than fully manual effort so this goes in a loop let's say first the llm they eval the uh llm they do the evaluations then the smas say they use an AI assisted framework to do the evaluations and based on the feedback the system is again improved so this goes in a loop and this Pro this makes the process scalable one short coming is that it requires high quality human in the loop framework and the current available technique is given by Lighthouse Lighthouse can be used to conduct holistic evaluations spanning rag hallucinations security privacy of topic and toxic evaluations since it is a human plus air framework the rag hallucination metrics are more nuanced spanning correct and complete correct but incomplete which means uh that there are some information that are missing correct plus extra which means that some additional information is present fully incorrect huc hallucinations and finally no response and it also gives a holistic uh holistic eval on topics such as privacy security of topic and toxicity so using the human plus AI in the loop framework several test cases can be generated with a few clicks and the human expert can actually go and provide feedback in an AI assisted manner not by one by one but in an a assisted Manner and make edits via the UI itself so the evaluations can be conducted with few clicks to generate a detailed scorecard and error analysis now let's uh take a look at the demo I'm trying to get into my Google Cod I'll be sharing my Google collab in second okay okay so let me share um have to get the sharing so okay so I can share this entire Google collab in the end or I can just share uh the GitHub link for everyone who is trying to follow this uh so this will be available to everyone in the end um uh first we have to install a few packages of so okay so what we are trying to develop here is we are trying to develop a rag based system uh which will um which will get answers from a website link and then we are also creating a question an answer or the benchmarking data set on which the rag system will be evaluated on and I I can show you like how you can build your own evaluations so first you have to install the required packages I have already installed that for the time sake and then give you open a API key so okay uh here I have given the data science Tojo I have taken a the a link from the blog of data science Tojo which says um oh okay let's go to that first data science Tojo block opens uh open source llms for Enterprises the benefits use cases and challenges so let's say I want to do a Q&A on this uh on this link so here I'm using the uh uh a data loader which is called the web base loader that basically scripts the information from this link and then it loads the data uh the chunk size and the chunk overlap are nothing but like when you have the data uh how well you can divide it into into different chunks so that the uh it can be ingested into the LMS why this is necessary because the llms they have a token size limit so they cannot um depending on what llm that you are using there is a limit to the tokens that can be given as an input and chunk overlap is what is the overlab between different chunks Uh current and in this case uh I will use the token text splitter what this does is actually splits the given chunk that we generated before as per different tokens you can use other Splitters as well such as recursive character splitter character splitter and many other splitters but here I have taken toen Tex splitter so the document uh is split as per the chunk size and the Chun overlap and the edics are created using the open a embeding so for someone who's not familiar with the embeding the embeding so the um the large language models or any other neural nwork models they cannot um understand directly so they have to be um they they have to be given information via embeddings which basically converts this natural language uh to numerical values so that is what the embedding is let's run this and so this creates the split document uh which is actually which has the chunk uh the chunk values the second step is to generate the questions and the ground truth answer the I have given a prompt below which I actually instructs the llm to generate a question for each given chunk and also generate an answer to the question the answer is return returned in a Json format so this QA template is actually uh just used as an use as an example this is um uh not uh done via any uh this is not done uh VI any promp tuning or anything so it says please generate a meaningful question where the answer is present in the given uh paragraph also generate the answer to the question using the information in the paragraph do not generate any extra information not present in the given paragraph and we are giving it some examples as well and when we you give examples in a prompt is it is actually called a few short prompt technique so uh here I'm giving it some examples where I give like a context or a paragraph and then the output is in a Json format which is the uh the question uh that it will generate and it also generates the answer as well so this is just um to give to um to to tell the llm that whenever you are generating the question and the answer both the question and the answer should only be related to whatever information you have in the paragraph So Okay uh now now we now based on the uh the template that I have written we will generate uh so I'm using the L we will generate a chain so first we will uh generate the prompt template which uses the input variable the input variable here is just the paragraph or the input context that we will give here the input context that we will give here and uh remember that um uh these are just examples on the input context that we will give here will depend on the chunks that we have created so the llm chain um and the llm chain is created so we are taking the qf prompt which we have created here and then the llm is a chat open AI you can uh use any other llms that you want and we are keeping uh the temperature as zero uh what is temperature so temperature is the how much variability you want in your resturants so if you have a temperature zero that means it gives very precise information every time if you increase the value that means it it becomes more stochastic more creative LMS can become very creative and it and it gives creative results every single time so you can play with this but for Simplicity and uh for this task because it's a q Q&A generation task and we want accurate question answer we have kept the temperature zero I have taken the model uh as 34 Taro preview you can explore any other models that is out there so you can go models openi you can go to models here and then there is a whole list of models here that you can check out and I'm using like gp4 sorry gp4 turo preview and you can any you can use any other um models that is there just be mindful full of the context tokens the context window if you use a large context window then it can cost you more so be mindful of that before using before using any of these models now let's take uh the top two splits only and it uh because uh for Simplicity let's take the top two uh splits and first so let's create a dictionary where it will store the chunks the question the chunk like if it's chunk chunk one or chunk two or chunk three and then it will store the questions and then the CR trips so uh the Q&A is generated um see run this because I see my colel got restarted let's see if I can run this or it's giving me some error yeah my colel got started so let me uh start from here from the beginning or let me I can actually show this and in my local local actually I have a visual studio let me grab that so as you can imagine I'm very prepared for this let's see okay okay so this should generate the question answer it takes few second and I create a data frame from the question answer generated so let's see what it has generated um the here this is the chunk the chunk it refers to remember like there there are just two chunks here and the questions uh the DF questions so what are the benefits first question is what are the benefits and challenges of using open uh Source large language models or llms in enter prises and then the second one is how many individuals follow LinkedIn newsletter for insights and data science generative a and large language models so the ground TRS it also generates the ground TRS let's see what the ground TRS are uh it gives me the contr controls the benefits of using include cost savings andan data control and the ability to tailor ai2 so and let's see what the second answer is so over 95,000 individuals trust the Linkin newsletter so um now that we have the q1l generator Let's uh build a rag application uh the rag to building a rag application is again using L chain and I've used Vector store as choma uh choma Vector store um you can use any other Vector store that is out there like Pine phone cant and so on so what rag does is like three steps first it does the chunks it uh it stores those chunks in a vector store so storing and then it using a retriever it tries to retrieve the uh the ex the the context or the chunk based on what quy you have given so that is exactly what it is doing here and I have again used another chain which is the retrieval QA chain and the llm that I've used is a chat open AI CH and the model is gp4 T preview now someone may argue that the model is same as generating the QA so uh you can you can choose whatever model that you want it doesn't matter it's just for examplar purp purposes and then and then the chain type is stuff that it U this gives the entire context as input so let's run this so what it it does it just based on the questions that we have generated here is it tries to find the relevant answer from the entire uh document that we have given on the uh the document here is the website that the block from data science D so it takes a while depending on um what model you have chosen it can it can be very it can take a lot of it can take a long time as well so let's see if have a few second to run okay PR run now or let's run the data frame that was created so it gives you the answers um that are generated using R and then the context like from what context it write the answer so uh this is uh I I showed you that um I told you before like these apps they are very fickle and they are not consistent to the variations in uh the prompts so let's say I modify I give do an experiment and let's say I modify the question one question and um use the generative AI term instead of llms the actual questions can be very complex like no one is going to ask like very simple questions right whenever you are Q Q&A with a document depending on what depending on the users they can ask like very complex questions and because they do not know what is uh there in the document which is generally the case the document size cannot be just this one page uh URL link it can be like 8 to 100 pages and these are never never straightforward so let's say uh the modified question is why should I use generative AI in my company and I run theun this this and then I run the uh run the query let's see what it gener okay uh so it says integrating generative AI into a company can offer a wide range of benefits enhancing creativity decision making here are several reasons if you can see the context it has actually gotten information from its memory it says shareholder of open a global LLC which despite for a profit company blah blah blah it actually gotten the information from the memory rather than getting the information from the document that was given um so you have to do like careful uh QA quality testing on what llms you are building what prompt you are giving and U you're asking if you're U if you're are telling if you're doing some uh checks before actually depl now like I showed before one of the most popular technique is actually llms evaluating llms so let's give an uh evaluation prompt where an llm I'll say the llm to evaluate um you are given so you I'm giving this evaluation prompt where I'm saying that you are an evaluator you're given a question a correct answer and the given answer so your goal is to just compare the answer and the given answer and label if it's or so okay um I have done this before so I can give you the um okay this run because I've already loaded the open AQ information in the environment file so it didn't require it to run uh let's see now I'm running that on the ground TRS and the answers so let's see what the score I get okay I get an eval score that both of them are true that there is there is U it it the answers that I get these are in um these are correct or these are correct with respect to the round TRS that both are same and so it gives the true labels here now let me see if guys okay now I'm going to show you another example which is using ragas which is like an open-source technique where um the uh it it uh the metrics are answer relevancy faithfulness context to recall and context prision so let's run this on and let's see the result that we get okay uh we got the faithfulness score as one and then it gave an end here and then the answer relevant see were high the context recall was uh one here and then the and it Z here so so okay so uh this is uh this is what I wanted to show like different uh different evaluation practices that is out there and um there is another which is eval with Lighthouse human in the loop framework but I'm not going to go into details over there but I can uh share the link to anyone who is interested uh so that's it um that's uh yeah that brings uh see yeah so that brings to brings an end to this presentation and I can take the questions from here so will this slide deck be shared after the meeting yes uh and the and the entire recording is there as well so it will be shared so how can it perform automated manual evaluation techniques without risk of exposing the underlying data set um so so yeah you have to have certain um card rail so that you do not uh expose the underlying uh under underlying data set you need to so whenever the llms whenever you are giving U like I said like there are several prompt injection techniques which will try to uh get information from the underlying data set so you have to have a right Cil you need to have uh um a right sort of uh monitoring techniques as well after deploy so that uh if anything uh if it has revealed anything then it can uh stop it right away or even before deploying anything you can do a lot of uh you can uh do a lot of um tests uh as on different prompt injection detector or different uh uh or different Gil techniques so that you know that the underlying data set is not exposed so yeah [Music] and will the speaker touch on the question of how to select llm for people who want to install DIY localized llm for use with personal data so I will not cover uh that but uh there like you can use several localized llms and you can just uh use that with your local use that with your personal data um or you can use different apis that are there uh for example with any scale and other so you can use that for llama 2 and Mixr and so on so but uh for the time sake and because this is only focused on evaluation techniques I won't be uh discussing on uh how to install DIY localized llms to use with personal data would like to understand evaluation tool of framewor LM uh response to natural language question with code generation like SQL query um and SQL query is then executed against the database to generate yeah so U so if I understand it correctly Jes you asking that U uh in case of code generation how you can how you can know that the SQL query when it is executed it is able to generate the relevant answers so yeah for that you need to have um uh you need to you need you need to have like uh some sort of uh some that there in case like there uh you need to have some sort of uh uh ground truth on where you can evaluate your SQL queries U let's say you have the data sets uh on which uh the databases on which you are running the uh the SQL query on you can create some benchmarking data sets and then you can uh try to find the answers that the answers that you get whether they are the relevant answers to your question you can do some uh kind of um uh matching to see if there is some kind of uh similar between the question the SQL query that you have asked and the answer that has generated so far there is uh there are like very less techniques who Focus or very few techniques that have known who focus on SQL query but so that is uh definitely a very interesting task and that is uh um that is a very interesting research are as well so and yes I will give you give the URL to the collab notebook we'll do that open a I key do you need 20 per month um so so you uh so you can use the API key and it is like per API call depending on what uh you are using I think you asking about chat GPT plus which is 20 chat GPT 4 which is 20 per month but this is different that this is per API call so depending on how many uh API calls you make you'll be charged for that how does your product differ and what additional value does it offer compared to the methods uh you share so ours is actually a human in the loop framework it's a it's an end to end eval framework or end to end QA framework as well that does an holistic um uh holistic U that takes an holistic approach it generates benchmarking data based on prompt injection input VAR iation to input prompt like variation depending inconsistency related to prompt variations that I showed and then the toxic inputs and uh so on so it generates the benchmarking data set B based on that and we support several apps like uh several different kind of apps like summarization uh chat box um currently focused on chat box but we are expanding that to summarization task classification task rag apps and so on so yeah it's a human in the loop framework so it does in an automated manner which in an AI assisted manner so that once you uh once the llm does the eal it can take the eals and then uh the it can take the answers and the answers that are performed that have the ground truth or the areas that need more attention or what we call as a risky areas it uh uh it gets that information and then the SMS they try to go and see why it has performed bad or they can edit the information or they can give the rating and based on that feedback we again do like an llm alignment or we take the feedback and then improve the uh improve the system so yeah that that is how it differs because three models are statistical in nature they always give different how always give different outputs how to come across this in validation yeah I mean uh the models they are very they are very statistical in nature and they will give different outputs so you have to have like rigorous test cases rigorous test benchmarking data sets on which it will perform and this cannot be any state static benchmarks these have to be uh like Dynamic benchmarks so it has to be uh in an evolving process so that you know that how your model is performing yes I can give you link to the demo how do you feel um eal will evolve in the next few years today it feels like sort of putting a gold star sticker on your model but it feels like it needs to evolve past that uh eval above X person by of government organization for it to be useful for business yeah eval will become the important technique in next few years eval is actually a short short part the entire is like the end to Q&A uh sorry end to Quality assessment that need that needs to be done and um uh more and more that I think about it it will become uh U like uh it it will it will become a very essential part because there'll be more and more apps in future and more and more these eval techniques will evolve definitely llm EV valuating llms will become uh will play a very key role in that but there should be some human component as well which is doing the checks and which is putting the checks that whether it is doing the right thing or not so definitely this will evolve a lot and there is a lot of research that is going on in this uh in this field um yeah so that's that what would be the difference if instead of open a we use something locally train llama your example is quite specific and look need full gen model so you can definitely use a locally trained llama model any other model and um uh you can try uh the examples on that so this is uh just to show you give you an example as to how you can build your own Evas how you can see that with slight change in uh variations or input context how the uh answers they differ or or gives the answer from from the memory so that's why the eval techniques are needed and of course you can customize your rag apps and whatnot but it is just to give you an idea that if you do not have this eval technique how would you even know that your rag model is performing better or not so so yeah so you can use any kind of eal techniques uh any kind of uh model not only open AI but um you need eval technique for that yeah yes I'll be sharing the slides on the python code just after this meeting where are you getting the I don't know what's the where are you getting the questions from do you do we not have access to the questions you're answering and I'm sure you have access to the questions are answering right now maybe the host can guide you for that but if you go to the Q&A session you can see that the questions are right there AWS has llms like CLA or better could you please compare this with GPT models um again like sure sure that can be done but this is out of the uh this is not what will be covered in this uh webinar now but sure that can be done with Cloud at Bedrock okay I guess that's that's about it are there any other evaluation techniques yeah uh that's it for the questions I believe uh yeah are you done Tali with the crew Q&A I think I'm done like there's this this last like are there any other evaluation techniques on like there are U there are there are a lot of evaluation techniques that I just showed um and many are evolving as well uh like you can um uh like you can use open source techniques like Ras or you can use closed Source techniques like that Lighthouse provides so yeah there are sever e techniques and these e techniques would vary depending on which uh apps you are trying to evaluate is it the summarization app or is it the um SQL query kind of stuff that come asked or is it the chat chat so yes so these these EV techniques they value a lot I I don't think that Bedrock is a model I think you are trying to bedrock is like a platform where many different models are hosted s the so I think you want to ask about CL GPT I haven't done that assessment but if I'll do you I'll post it on the link so yeah thanks for the question