I Analyzed My Finance With Local LLMs

[@YoutubeCom_11]

Update: Ollama now works on Windows normally

[@Thuvu5]

Yayyy! Great news!

[@TailorJohnson-l5y]

Are you a real human? I have NEVER seen an author on youtube cover so much incredible knowledge in such a short video. This is absolutely AMAZING!!! Thank you

[@martingrillo6956]

Her being an AGI would make perfectly sense

[@Scarsuna]

@@martingrillo6956 Or she used her AI skills to generate the teleprompter output she's reading? ;)

[@whatifi-scenarios]

This is great. We're in the process of integrating LLMs into our "what if" scenario modelling platform and this gave me a few ideas on next steps. Sharing this video with my dev team!

[@noahchristie5267]

Incredible intro video for the semi technical about how chat gpt and similar models will be used in daily life to improve the mundane tasks, with a side of cautions about incorrect answers and computational limitations! Great balance, I’m already sharing it around our team 😊

[@Thuvu5]

Thanks a lot for your comment and for sharing it around! Really appreciate it 🤩🙌

[@johndoughto]

Awesome structure to convey a "simple" idea, without getting down into the weeds with how truly complicated it is. Thanks!

[@AshishRanjan-jn7re]

Great video... My 2 cents: we can force LLMs to respond only in json format by stating it in system prompt, so you get consistent parsable response always (I've tried with gpt4), also you can provide list of possible expense categories to avoid grouping them together later (like 'Food & Beverage' and 'Food/Beverage')

[@martinmoder5900]

Yeah, it is very powerful! However, is llama2 also providing this?

[@NicolasCerveaux]

@@martinmoder5900 llama2 and even gemma:2b does that too, but when I tried it still generated "new" categories, and the json answers would be "odd" like sometime it would modify the name of the expense.

[@siliconhawk]

​@@martinmoder5900 llama 3.1 (the new one) is pretty powerful so it should be able to do it for you. given enough compute power

[@roberthuff3122]

🎯 Key Takeaways for quick navigation: 00:00 💲 *Reviewing Income and Expense Breakdown* - Explained the process of analyzing financial transactions. - Talked about classification of expenses into categories. - Spoke about using low-tech ways and an AI assistant for classification. 02:16 💻 *Running a Large Language Model Locally* - Discussed different ways to run an open-source language model locally. - Listed various popular frameworks to run models on personal devices. - Explained why these frameworks are needed, emphasizing the size of the model and memory efficiency. 04:18 📚 *Installing and Understanding Language Models * - Demonstrated how to install a language model through the terminal. - Showed the interaction with the language model through queries in the terminal. - Assessed the model's math capabilities, showing a failed example. 06:48 🎯 *Evaluating Expense Classification of Language Models* - Checked if the language models can categorize expenses properly through the terminal. - Demonstrated how to switch models, correctly installing another model. - Showed the differences between the models and preferred one due to answer formatting. 08:24 🛠️ *Creating Custom Language Models* - Explained how to specify base models and set parameters for language models. - Demonstrated how to create a custom model through the terminal. - Discussed viewing the list of models available and building a custom blueprint to meet specific requirements. 11:46 🔄 *Creating For Loop to Classify Expenses * - Discussed forming a for loop to classify multiple expenses. - Detailed how to chunk long lists of transactions to avoid token limit in the language model. - Mentioned the unpredictability of language models and potential need for multiple queries. 14:32 🔍 *Analyzing and Categorizing Expenses* - Demonstrated how to analyze and categorize transactions. - Showed how to group transactions together, clean up the dataframe, and merge it with the main transaction dataframe. 15:14 📊 *Creating Personal Finance Dashboard * - Detailed the creation of a personal finance dashboard, that includes income and expenses breakdown for two years. - Introduced useful visualization tools such as Plotly Express and Panel, giving a short tutorial on how to use them. - Demonstrated the assembling of a data dashboard from charts and supplementing it with custom text. 17:02 📈 *Visualizing Financial Behavior Over Time* - Demonstrated the use of the finance dashboard, drawing observations. - Concluded with a note on importance of incorporating assets into financial management. - Highlighted the value of running large language models on personal devices for tasks like these. Made with HARPA AI

[@xugefu]

Thanks!

[@voonoo2059]

Xin chao Thu, thanks for your great video. That's so mind blowing to see beyond the usual usage of ollama local AI.

[@jteichma]

Thanks for the great overview of using aa local LLM Thuy! Very useful and informative.

[@etutorshop]

OMG this is inspiring I always wanted a 3rd party view about my expenses without loosing control of my data and this video hits the nail on the head.

[@Thuvu5]

So glad to hear! Good luck with your project 🤗

[@bimoariosuryandaru325]

This is great! I was recently experimenting on a personal finance tracker dashboard and connect it to a chatting apps, so the user could easily input their financial activity by only typing it. On the process, i try to use chat gpt to simplify and generalise the format so we can input the data faster, never have i thought that it could be done by a local LLM. Looking forward for your next video.

[@BullMoves365]

I love the content. Also, I have not seen anyone can program so fast!!!

[@kevinmanalang9182]

Hi Thu! Last year I had referenced your panel dashboard video to build my personal finance dashboard. I like seeing how you built yours. Your content is very useful. Thank you!

[@SebastianSastre]

Thank you for sharing this dear! You covered the basics and shown the path to a great first goal with your own custom on premise and well licensed LLM. Huge!

[@Thuvu5]

You are so welcome! Glad it was helpful 🙌

[@Codad]

This is such a great video. Thank you for making it. I had no idea this sort of thing was possible and I'm finding all sorts of ways to take advantage of it now.

[@PauloLeiteBR]

Excellent video, I used the concepts to enhance a project that I had already started in R and it worked fine, but so slow in my computer (like 5 min to analyse 10 registers). Now I know the concepts and I`ll keep experimenting with other LLM models. Thank you!

[@tolandmike]

You just earned a new subscriber, Thu. I mean, wow. Very inspirational to see what you built on a friggin laptop, no less. Goes to show you don't need thousands of compute cores, either. Ver very cool. 🎉

[@Thuvu5]

Wow, thanks you so much! Indeed, we definitely don't need to go broke buying super computer for this 🙌

[@SamFigueroa]

I've noticed that most LLM understand that you would like a CSV formatted output and you use that to get more consistent output.

[@n0n4m3y3t]

thank you for including the repo!! it makes the content 10x better!

[@gridaranbirthuvi]

Great video .. The one project which I wanted to take up during my holidays .. Learn in the same time have a view on my personal finance ..

[@brunogillet7132]

Thanks so much ! Being investigating AI for just one month, having so much to learn again (and that's cool), your videos really help. Being not a natural english speaker, it was a bit fast to follow, but no issue : It was clear, precise, and... I will find time to listen to it up to be sure having got any lesson from it. Same apply to your other videos, but change nothing : ( It could even help me improve my English level ;-)... )

[@Thuvu5]

Great to hear!

[@luismoriguerra669]

this is one of the best videos I watched about llms

[@smiley3239]

Thank you! it's quite hard to follow up with this ollama thing, and you explain it so easily. thank you!!! please mae more of this!!!!

[@borismeinardus]

Love the video! The beginning sets up the project perjectly and the tutorial is very easy to follow!

[@apvitor]

You are a very good presenter, easy to follow. Nice content

[@youthresearches]

As always, high-quality content from a highly competent woman!

[@Thuvu5]

That's so kind of you, I'm trying to be ;)

[@Arsenik2]

As a data scientist, I am blown away by your video's theme. You successfully managed to keep it simple to attract the interest of the majority and mention about technical details that is beneficial for more technical people watching this video. Best wishes!

[@NRICHMEMotivation]

I am blown away by this video! If only I can get my CPA to do the same. I guess I’ll need to learn to code.

[@jman9545]

Super cool! Great channel. Excited to watch more

[@bereniceflores81]

Always good to see more people bringing data skills to understand personal finance.

[@icemelt7ful]

As a Javascript coder, this was a mindblowing video, I had no idea Python was this powerful.

[@anissaa1017]

Thank you so much for sharing this with us!! I’ve been looking to do this for years but just thinking about the task ahead, I would give up. I will definitely analyze my own financial statements. Thanks mucho gusto!!

[@soky2466]

Incredible video, I love how you simplified all the process. Your content inspired me I will try it on my personal projects as well

[@Thuvu5]

Awesome, go for it!

[@winhater]

I never ever ever comment on anything, but goddamn - what a great video/tutorial. Just finished playing with the notebook and I learned a ton!

[@Thuvu5]

That’s so awesome to hear! Thank you so much for commenting ❤️🤗

[@kylonguyen-we5mx]

Thanks sis, you're awesome!

[@LukeBarousse]

"Although, as you can see I can't retire anytime soon" 😂😳 Thu, this was a pretty ingenious way to label data; one of the biggest part of our time is data cleanup and this helps speed it up

[@LukeBarousse]

out of curiousity, why did you choose ollama? (vice something like LM studio)

[@Thuvu5]

Haha, yeah I thought I'd saved much more.. 😂 Definitely, I hope to explore more analysis use cases for local LLMs. I heard about LM studio but somehow I just like the setup with Ollama better. I guess they are very much the same in the backend.

[@FaruqAtilola]

Trust me, clicking the video and scrolling through the comments, I was anticipating your comment to be at the very top😅

[@gr8tbigtreehugger]

This was an excellent video - many thanks for sharing!

[@mrbarkan]

This is incredible, a bit far fetched from my skills and time in hands. But surely inspiring!

[@akinwalehabib]

Amazing work you put in here. This is inspiring

[@TheInternalNet]

I learned so so much watching this. Thank you so much.

[@sanatdeveloper]

Awesome research as always!

[@positivitywins8957]

Amazing job explaining this!

[@olivermorris4209]

Thanks Thu, great demo of Ollama, sorry your arent going to be retiring anytime soon😢 I really like the multimodal model support in Ollama, llava is a great model to try and runs on not much RAM.

[@Thuvu5]

Thank you Oliver! I would absolutely not mind making videos until I retire though 🤣. The multimodal support is interesting, I haven't tried it out yet but will look into those models a bit more 🙌🏽.

[@thinkingmachine7760]

Thank you so much. 🥰It is so well explained and a very cool project. I think LLMs are a powerful tool and running them locally will make it safe to share critical information with them.

[@Thuvu5]

Thank you, really appreciate it! ❤

[@korntron]

Outstanding video, especially for this beginner. Didn’t know you could run the models locally. Those ollama layers look like docker, fascinating how the context is setup. Time for me to spend some cycles on all your vids, not just the couple I’ve casually looked at. Thanks!

[@Thuvu5]

Glad to hear you found the videos helpful! Thanks for stopping by 🙌🏽

[@pw4827]

Me too. I thought you need to have some monstrous supercomputer and spend weeks on configuring everything to run one of these models locally

[@juliusprojeto]

Thank you very much

[@Turbo_Tastic]

this is great.. thank you for the breakdown of all these options

[@ricb4195]

I loved this and hope to try this out for myself (though my programming skills are very rusty)

[@leonardvermeer7908]

What an amazing video! This is definitely a personal project that I've wanted to tackle and while I'm familiar with other languages, I'll definitely use your video as a guideline.

[@andrewshatnyy]

Wow this is fantastic video. Thank you, Thu!

[@franklimmaciel]

Thanks for this great video.

[@dasurao7736]

Your videos are well thought out .. Keep them coming - Dont want you "retiring soon" 🙂

[@Thuvu5]

Haha thank you for this! Don’t worry, with KZbin I don’t want to retire anytime soon 😉🤗

[@agyeirichmondowusu9670]

You earned a new subscriber today. Thanks for how intuitive this video is. I also love how you pronounce "O-lla_ma"😹..kidding

[@Thuvu5]

Haha, thank you for the subs! 🎉

[@Jaybearno]

Cool project! I'd like to try it myself. One interesting idea is to have the LLM generate a memo field for each transaction (which can be controlled via prompting). Then by embedding these and doing hybrid retrieval, you can search in natural language as well as by metadata for transactions.

[@Thuvu5]

That’s an interesting idea! Would love to see how well the retrieval works 🤗

[@PhilSmy]

Great video. Very inspiring. Also...I used to live in Amstelveen (20+ years ago!). Funny to see that name in there.

[@Thuvu5]

Oh haha, the world is small! 😀

[@chrisumali9841]

Thanks for the demo and info. So detailed and analytics are great. Have a great day

[@SteelWolf13]

Nice. Might give this a try over the weekend. Just need to figure out how to get my banks data.

[@michaelmraz2707]

How to make LLM learn and be able to correctly identify new categories? For example, creating an income statement from the list of all journal entries, but LLM need to identify each entries and correctly categorized it. Say, there's an entry for a plane ticket and wages paid to XYZ. The LLM reads the entries and correctly map it to expense item "travel expense" and "salaries/wages" expense. This is similar concept to your video, but more broad with the ability to learn.

[@haralc]

There's no language models that can do math. It can answer 2 + 2 = 4, because it has seen people talking about it, but it doesn't really do computation.

[@Dom-zy1qy]

Can RAG not used to do simple calculations?

[@joe_hoeller_chicago]

Actually no. It depends on which LLM, some like Orca2 are trained in math.

[@gammalgris2497]

The LLM cannot but an artificial neural network can maybe help as its just a pile of linear algebra. But then you have to think about what you actually want to do. Do you want to find spending patterns? There are easier ways to do math. Finding categories is an arbitrary task. Check that the LLM doesnt' mix up your numbers/ spending numbers.

[@GeekProdigyGuy]

​@@joe_hoeller_chicagoLLMs trained to do math are like dogs trained to do math. It might do mostly OK for a bit, but errors are a matter of when, not if.

[@Tyrone-Ward]

ChatGPT said, "LLMs are quite capable of performing mathematical tasks, including arithmetic, algebra, calculus, and even some advanced mathematical concepts. They can solve equations, perform calculations, and provide explanations for mathematical principles". So you're wrong.

[@lionelshaghlil1754]

Thanks, That was inspiring indeed :)

[@gaelanmelanson3532]

Such a cool project!

[@muhannadobeidat]

Thanks for the video. Nicely done and presented, educational with an interesting use case

[@lucasjenkinson]

This is a life-changing video

[@yezarniko9621]

That what I'm looking for !!! Thanks

[@_stition9777]

Thank you so much for making this video. Subscribed, this is exactly the content I look for

[@IdeationGeek]

I see how this is useful for being one's own accountant :) Super!

[@bhavyajain3420]

That's awesome. I would also use Llama to write the code for generating plotly charts/dashboards haha!

[@hrgagan9192]

Wow absolutely wow, thank you for such a great project, so many ideas ringing in my head. Cheers

[@mustafadut8430]

If you want to give data as many as the number of tokens of the model. You don't need to calculate and know by hand. Instead, you can do this with "chunks" in Langchain. nice explanation thank you

[@chocolatecookie8571]

I have a great admiration for the younger generations who know how to do all this tech stuff. It looks very complicated to me.

[@Thuvu5]

Haha, that’s so kind of you. I’m sure it’s less complicated than it looks

[@AlexandreRousselet]

J'ai adoré, vidéo super clair allant droit au but et qui nous la joie d'aller découvrir le code

[@bengriffin6157]

Very well explained. Looking forward to you posting the github repo.

[@Thuvu5]

Thank you for watching! I've added the repo link in the description 🙌🏽

[@haqk4583]

Thanks for the great intro into how to get started with local LLMs. I'll give it a go after Tết 😄

[@Thuvu5]

Happy Tet holiday! 😀🎉

[@EricSchroeder-cc4hf]

very good! thank you for sharing!

[@DorianIten]

Amazing. Thank you for sharing this, I learned so much!

[@bhusanchettri8594]

Great insights and well explained!

[@kcm624]

Awesome video, learned a lot of new tools and want to try this out. For the dashboard, wonder if using Excel would be easier? Not sure.

[@davidtindell950]

I just read about the latest Meta LLAMA model that is supposed to be better than GPT4 for s/w dev! I hope that we can run it as a LOCAL LLM ! Thank You for this timely vid. ...

[@Thuvu5]

Ooh that’s pretty cool! 🤩 So great to hear many models are approaching GPT4 capabilities 🤯

[@therealpattypooh]

Great video to start using LLM! Thank you for sharing!

[@TheBenJiles]

Thanks so much! It giving me inspiration for using this in a security analysis context.

[@nimeshkumar8508]

Thankyou so much for this video. I relly like the explanation. Thanks

[@nguyentrananhnguyen7900]

ayo, i'm just doing my first step that's logging every expenses i got since the start of this year i'm just thinking about doing some sort of software that help me manage my expenses and savings and this is exactly what i think of thank you for the high quality video

[@anuraagpandey8316]

incredible, loved the content.

[@TeaForecast]

Very concise and informative video. I appreciate it.

[@relaniumz]

What vscode extensions do you use? I'm especially interested in code auto complition extension.

[@Thuvu5]

It was GitHub Copilot 😊

[@aitech4future]

I was wondering where I listened to this music. Amazon learning has this background music. Thanks for sharing :)

[@atenciop123y]

Thanks again for another wonderful video. Ollama is now available on Windows as a preview. I used that preview version on the solution you shared here and it worked great! 🙂 Can you recommend a tutorial on the panel library? Thanks in advance.

[@TheSabatuer]

FYI The multiplication example you tried wasn't accurate, the 2nd input number was different than the example you tried. 49,792 x 857,294.2 = 42,632,383,271.8! 45 mil is still way off

[@DarkSoulGaming7]

Thank you SOOOOOOO much for this !! this is an awesome tutorial

[@Thuvu5]

You are so welcome! Glad you like it!

[@qbitsday3438]

Love it , i am subscribing instantly , i have a lot of questions.

[@GeorgeZoto]

Excellent video and practical application, you didn't get to cover pydantic much which solves a current challenge with LLMs. As for the dashboard, maybe another framework or approach with less or no code could be be more efficient :)

[@marijnstollenga1601]

You can get rid of the randomness by setting the temperature to 0, or controlling the seed.

[@Thuvu5]

Thank you, this would be better indeed!

[@marijnstollenga1601]

Btw another good trick, at least with llama.cpp you can define a grammar for the output. So instead of coaxing it and validating, you can _force_ it to output e.g. json, or even a more specific grammar! @@Thuvu5

[@marijnstollenga1601]

I lost my other reply I think. I wanted to point out that you can use grammars to force the output you want (in llama.cpp at least). So instead of asking to reply json and validating, you can set the grammar so only valid tokens are considered! Very overlooked feature @@Thuvu5

[@xiyangyang1974]

Would there be no disadvantage?

[@marijnstollenga1601]

I think there is a higher chance of repetition, but you have repetition penalties for that. And indeed less 'creativity' but when you classify data in this case you don't want that anyway @@xiyangyang1974

[@vadud3]

Really awesome explanation! I am going to use this. Thank you Thu!!

[@ShivamMiglani]

Thanks Thu, just heard about local LLMs from my boss today and look whose video is on the top to help me out! 😃

[@Thuvu5]

Hey Shivam! Thanks for watching! So happy to see your comment 😍🤗

[@oneallwyn]

Finally the text classification video that I was searching for

[@Echo11days]

Well done I'll try and re-create this. Thank you once again

[@jpcf]

I was looking for THIS! Thanks!!

[@heijd]

A faster and cheaper way to do this is to use the LLM embeddings directly. This is what happens anyway behind the scenes, but it makes the data nicer to handle.

[@pieterjdw]

Could you give some guidance to this approach?

[@raviv5109]

You are awesome! Thanks for making this video.