Paul Iusztin

Mom! I'm on YouTube again. This time it's not my own uploaded 😅 PyCon Lithuania uploaded my session (https://lnkd.in/dY2szb_U you can also check out other talks here: https://lnkd.in/dWwxdcWc) Happy watching! ---- Stay in the loop for the latest insights from Uncle Data! * Click the notification bell icon 🔔 to receive updates on Linkedin * Subscribe to my Newsletter on Substack ✏ * Listen to Uncle Data Podcast ⌨️ Listen. Learn. Connect ---- #dataengineering #patterns #python #spark #iceberg

44

Reasoning models are really interesting: https://lnkd.in/dEbydJHZ Whom it's for: "...and by developers in all fields to build and execute multi-step workflows" Learning to reason with LLMs: https://lnkd.in/dgnBiXfU

1

🚀 Exciting News: Meta Releases Llama 3.2 🚀 Meta has launched its latest iteration of the Llama series – Llama 3.2 ! This update brings even more power and versatility to large language models (LLMs), making them more accessible for developers and businesses alike. 📣 Llama 3.2 Vision (11B & 90B) models can now handle image-in, text-out tasks, excelling in areas such as document understanding, chart interpretation, and image captioning. They’re designed for advanced visual reasoning and making them perfect for tasks like object detection, OCR, and contextual Q&A. 📱 On the other hand, Llama 3.2 lightweight variants (1B & 3B) can run efficiently on most hardware, including smartphones, ensuring low latency and improved privacy by processing data locally. In my latest Medium blog (attached), I explore how to run the Llama 3.2 3B model locally using Ollama. Ollama is a revolutionary tool that allows users to run open-source models on their local machines without relying on external APIs – offering greater control, cost efficiency, and privacy for LLM execution. 💡 If you're curious about how to get started with Llama 3.2 on your local system and unlock its potential for your projects, check out the blog! Learn how to install, pull, and run Llama 3.2, and discover the power of local LLMs. #AI #EdgeAI #MobileAI #Llama32 #PrivacyTech #MetaAI #Innovation #Tech #ArtificialIntelligence #OpenSource #AIModels #LLMsOnLocal

4

I am 𝗾𝘂𝗶𝘁𝘁𝗶𝗻𝗴 𝗰𝗿𝗲𝗮𝘁𝗶𝗻𝗴 𝗰𝗼𝗻𝘁𝗲𝗻𝘁... 𝗝𝗼𝗸𝗶𝗻𝗴, but here is 𝗵𝗼𝘄 to 𝗯𝘂𝗶𝗹𝗱 your 𝗟𝗟𝗠 𝘁𝘄𝗶𝗻 for 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗻𝗴 posts or articles 𝘂𝘀𝗶𝗻𝗴 𝘆𝗼𝘂𝗿 𝘃𝗼𝗶𝗰𝗲 ↓ 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗮𝗻 𝗟𝗟𝗠 𝘁𝘄𝗶𝗻? It's an AI character who writes like you, using your writing style and personality. 𝗪𝗵𝘆 𝗻𝗼𝘁 𝗱𝗶𝗿𝗲𝗰𝘁𝗹𝘆 𝘂𝘀𝗲 𝗖𝗵𝗮𝘁𝗚𝗣𝗧? 𝗬𝗼𝘂 𝗺𝗮𝘆 𝗮𝘀𝗸... When generating content using an LLM, the results tend to: - be very generic and unarticulated, - contain misinformation (due to hallucination), - require tedious prompting to achieve the desired result. 𝗧𝗵𝗮𝘁 𝗶𝘀 𝘄𝗵𝘆, 𝗳𝗼𝗿 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗻𝗴 𝗰𝗼𝗻𝘁𝗲𝗻𝘁, 𝘆𝗼𝘂 𝗻𝗲𝗲𝗱 𝗮 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲𝗱 𝘁𝗼𝗼𝗹 𝘁𝗵𝗮𝘁: → is fine-tuned on your digital content to replicate your persona → has access to a vector DB (with relevant data) to avoid hallucinating and write only about concrete facts . 𝗛𝗲𝗿𝗲 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗺𝗮𝗶𝗻 𝘀𝘁𝗲𝗽𝘀 𝗿𝗲𝗾𝘂𝗶𝗿𝗲𝗱 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱 𝘆𝗼𝘂𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻-𝗿𝗲𝗮𝗱𝘆 𝗟𝗟𝗠 𝘁𝘄𝗶𝗻: 1. A data collection pipeline will gather your digital data from Medium, Substack, LinkedIn and GitHub. It will be normalized and saved to a Mongo DB. 2. Using CDC, you listen to any changes made to the Mongo DB and add them as events to a RabbitMQ queue. 3. A Bytewax streaming ingestion pipeline will listen to the queue to clean, chunk, and embed the data in real-time using Superlinked . 4. The cleaned and embedded data is loaded to a Qdrant vector DB. 5. On the training pipeline side, you use a vector DB retrieval client to build your training dataset, which consists of the cleaned data (augmented using RAG). 6. You fine-tune an open-source Mistral LLM using QLoRA and push all the experiment artifacts to a Comet experiment tracker. 7. Based on the best experiment, you push the LLM candidate to Comet's model registry. You carefully evaluate the LLM candidate using Comet's prompt monitoring dashboard. If the evaluation passes, you tag it as accepted. 8. On the inference pipeline side, you deploy the new LLM model by pulling it from the model registry, loading it, and quantizing it. 9. The inference pipeline is wrapped by a REST API, which allows users to make ChatGPT-like requests. . Do you want to learn how to code such a beast? 🫵 Here is a FREE course, made by Decoding ML, that contains: - 12 written lessons - open-source code 𝗖𝗵𝗲𝗰𝗸 𝗶𝘁 𝗼𝘂𝘁 𝗼𝗻 𝗚𝗶𝘁𝗛𝘂𝗯 𝗮𝗻𝗱 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝘂𝘀 𝘄𝗶𝘁𝗵 𝗮 ⭐️ ↓↓↓ 🔗 𝘓𝘓𝘔 𝘛𝘸𝘪𝘯 𝘊𝘰𝘶𝘳𝘴𝘦 - 𝘎𝘪𝘵𝘏𝘶𝘣 𝘙𝘦𝘱𝘰𝘴𝘪𝘵𝘰𝘳𝘺: https://lnkd.in/dzat6PB6 #machinelearning #mlops #datascience . 💡 Follow me for daily content on production ML and MLOps engineering.

180

12 Comentarii

Enhancing Trust in Large Language Models: Fine-Tuning for Calibrated Uncertainties in High-Stakes Applications https://lnkd.in/diw3uQvM Enhancing Trust in Large Language Models: Fine-Tuning for Calibrated Uncertainties in High-Stakes Applications Large language models (LLMs) often struggle with accurately representing uncertainty in their output, which is crucial for decision-making in fields like healthcare. The dichotomy between black-box and white-box estimation methods presents a challenge for LLM practitioners. To address this, researchers have proposed fine-tuning the models to recognize what they don’t know using a calibration dataset, resulting in faster and more reliable uncertainty estimates with relatively few additional parameters. Practical Solutions and Value: A practical solution involves focusing on black-box techniques for estimating a language model’s uncertainty for open-ended generation. The proposed method includes fine-tuning for better uncertainties, which has shown improved performance compared to commonly used baselines. This approach provides calibrated uncertainties with practical generalization properties, proving to be surprisingly sample-efficient and robust to distribution shifts. For companies looking to leverage AI, it is crucial to identify automation opportunities, define KPIs, select suitable AI solutions, and implement them gradually. To evolve your company with AI and enhance customer engagement, consider the AI Sales Bot from itinai.com/aisalesbot, designed to automate customer interactions 24/7 across all stages of the customer journey. List of Useful Links: AI Lab in Telegram @itinai – free consultation Twitter – @itinaicom #AI #MachineLearning #UncertaintyEstimates #HealthcareTech #AISalesBot#artificialintelligence #ai #machinelearning #technology #datascience #python #deeplearning #programming #tech #robotics #innovation #bigdata #coding #iot #computerscience #data #dataanalytics #business #engineering #robot #datascientist #art #software #automation #analytics #ml #pythonprogramming #programmer #digitaltransformation #developer

1

Errors and warnings in R programming can be challenging, but I've got you covered! I've created a list of common errors and warnings in R, including examples and step-by-step instructions on how to fix them. The list already contains 112 tutorials and is regularly updated: https://lnkd.in/e3pze5T Want to avoid making errors in R? Check out my course, "Introduction to R Programming for Absolute Beginners." This course will help you build a solid foundation in R programming from scratch. Learn more: https://lnkd.in/e3pze5T #analysis #statistics #pythoncode

33

4 Comentarii

Greenmask 0.2.0 - 0.2.5 Releases 🚀 Exciting news for PostgreSQL enthusiasts! The Greenmask team just released versions 0.2.0 to 0.2.5, marking wonderful milestones for our data anonymization and synthetic data generation tool. 🎉 If you've ever wished your database could understand your love for data security, look no further! Greenmask now supports PostgreSQL 17 and introduces features like database subsetting, pgzip support for lightning-fast compression, and restoration in topological order. 📦🔄 Need to anonymize data or simplify your backup process? Greenmask is here to be your database superhero. 🦸‍♂️ It’s stateless, customizable, and won't ask you to change your database schema—no awkward family dinners here. The magic is only a GitHub click away! Explore all the updates and features that make Greenmask your new best friend in backup and anonymization: [Greenmask Releases](https://lnkd.in/ejeHjcM2) Join the community, share your thoughts, or just say hi on Discord or Telegram—we promise not to bite! 🙃 #PostgreSQL #Greenmask #DataSecurity #OpenSource #DatabaseManagement Source: https://lnkd.in/en5uJ8eW

1

I will need to get Lilith to explain how the objects mentioned in this post map to Category Theory. First on Google, in my little filter bubble. https://lnkd.in/dpA5hZKV

1

Harmonics of Learning: A Mathematical Theory for the Rise of Fourier Features in Learning Systems Like Neural Networks https://lnkd.in/dpheiwN6 Harmonics of Learning: A Mathematical Theory for the Rise of Fourier Features in Learning Systems Like Neural Networks Artificial neural networks (ANNs) exhibit consistent patterns in learning natural data, leading to practical insights for machine learning models. Researchers have identified the emergence of Fourier features in vision models, offering valuable applications in tasks such as image processing, efficient coding, and classification. Furthermore, a mathematical explanation has been introduced by experts, demonstrating the relationship between invariance in learning systems and the rise of Fourier features. This theoretical foundation provides a basis for understanding representations in both artificial and biological neural systems. Key Theorems and Applications Two key theorems have been presented, establishing the connection between invariance to a finite group and the weights of machine learning models. Additionally, a practical model has been implemented to support the proposed theory, showcasing the potential to recover the algebraic structure of an unknown group from an invariant model. Implications and Future Work These findings have significant implications for the development of machine learning models and pave the way for future exploration of the proposed theory in real-number analogs. The practical applications of the research hold promise for advancing current practices in the field. AI Solutions for Business Evolution Discover how AI can transform your company and enhance your competitive edge with practical AI solutions. From identifying automation opportunities to implementing AI gradually, our AI experts can guide you through leveraging AI for business success. For AI KPI management advice and continuous insights into leveraging AI, connect with us at [email protected] or stay tuned on our Telegram t.me/itinainews or Twitter @itinaicom. Practical AI Solution Spotlight: AI Sales Bot Explore our AI Sales Bot designed to automate customer engagement 24/7 and manage interactions across all customer journey stages. Discover how AI can redefine your sales processes and customer engagement at itinai.com/aisalesbot. List of Useful Links: AI Lab in Telegram @itinai – free consultation Twitter – @itinaicom #artificialintelligence #ai #machinelearning #technology #datascience #python #deeplearning #programming #tech #robotics #innovation #bigdata #coding #iot #computerscience #data #dataanalytics #business #engineering #robot #datascientist #art #software #automation #analytics #ml #pythonprogramming #programmer #digitaltransformation #developer

2

I came across an excellent GitHub repository packed with hands-on Jupyter notebooks for anyone eager to explore LLMs. It covers: 🔹 Introduction to LLMs – Understanding the basics. 🔹 Tokens & Embeddings – The foundation of language modeling. 🔹 Model Internals – Peeking under the hood of LLMs. 🔹 Text Applications – Classification, clustering, and topic modeling. 🔹 Prompt Engineering – Crafting prompts for better performance. 🔹 Advanced Text Generation – Tools and techniques. 🔹 Semantic Search – Intelligent information retrieval. 🔹 Multimodal Models – Combining text and image understanding. 🔹 Customization – Building embeddings, fine-tuning BERT, and generative models. Big thanks to the repo creator for sharing such a valuable resource. Here is the repo: https://lnkd.in/gUKg7Jp7

77

2 Comentarii

Pay attention to these two guys at the top. They are fresh releases from MistralAI. The first one is Mamba-based (that's Selective State Space Model), which is quite unconventional and is indeed MistralAI's very first model based on that architecture. I haven't tested it yet, but I anticipate that it excels at coding, given its benchmarks. Even more intriguing would be to examine its memory consumption and performance with long inputs, as these are the advantages of the SSM architecture. The second is just another 7B model from the Math domain. Little Santa's Helper for the homework on math ;) By the way, I've just rolled out a huge update for LLM Explorer, which now employs a dynamic scoring and ranking for LLMs. I've named it the "LLM Explorer Rank." As a result, all 34,000 models are adequately ranked, regardless of their presence on the leaderboards. Check it out at https://llm.extractum.io.

9

Anthropic shared some really cool tips on building a RAG with the new idea of "contextual retrieval". Check it out here: https://lnkd.in/dT9cpVKV #rag #claude #openai #context #hallucination #llms #ai

1

1 Comentariu

LlamaIndex released #vdr-2b-multi-v1. vdr-2b-multi-v1: Multilingual embedding model for visual document retrieval without OCR or chunking. Languages: Supports Italian, Spanish, English, French, and German, trained on 500k samples. Cross-Lingual: Search across languages (e.g., German docs with Italian queries). Efficiency: Matryoshka Learning: Reduce vector size by 3x with 98% quality retention. Runs on NVIDIA T4 GPU with ~4.4GB VRAM for 768 image patches (batch size 16). #VLM #llm #multiModel #OCR LangChain

5

There is also an opinion that in its current form, AI will only hinder the development of science. Noam Chomsky argues that if AI like ChatGPT creates texts by compiling the most probable next words in a sentence, which is a generalization of humanity's previous experience on which the model was trained, then this is not how science progresses, but rather the opposite. The development of science is associated with contradiction, i.e., proposing hypotheses or theories that have not yet been found by predecessors, i.e., they are not the most likely result that automatically follows from previous experience. To move forward, one needs to think unconventionally, sometimes challenging universally recognized dogmas.

5

Step by step guide on building Kolmogorov-Arnold in C. Good to understand intervals how it works

6

You've heard of RAG and its different siblings and cousins 👨‍👩‍👧‍👦: GraphRAG, HYDE, LongRAG, SelfRAG, AgenticRAG, and many more... But all of these methods lie in either the database structure, or the retrieval method, but the data we feed into the LLM essentially remains the same. The authors of this paper 📜 thought outside the box: what if the format of the data could affect RAG performance? "HtmlRAG: HTML is Better Than Plain Text for Modeling Retrieved Knowledge in RAG Systems" is a paper authored by Jiejun Tan, Zhicheng Dou, Wen Wang, Mang Wang, Weipeng Chen and Ji-Rong Wen, look into a new approach to improve RAG, and that is html formatting. 💻 Traditionally, when we have data that is in the HTML format, we use an HTML parser to extract the clean text, and completely discard any HTML or CSS code in our data. But this radical approach to text cleaning might make our data hard to interpret and sometimes completely unreadable. For example, when removing HTML from a table, the clean text will be a bunch of randomly floating text that makes no sense 💫. LLMs are trained over an immense amount of data, and amongst this data is HTML formatted web pages. This means that LLMs have the capacity to understand and interpret HTML text. So, should we just feed HTML into our models then? Short answer: sort of. Long answer: not entirely. The paper explains that a single HTML document is on average 80k tokens, which is huge considering LLMs context windows are in the 32k to 128k tokens range, we can't just have one single document take up all of our context window, especially since 90% of these 80k tokens are CSS styles, JavaScript, Comments, or other meaningless tokens. To tackle this challenge, the authors developed Lossless HTML Cleaning and Two-Step Block-Tree-Based HTML Pruning 🌳. Which is a complex process that essentially takes advantage of the tree structure of the data, considerably reducing its size all while retaining the information of the original HTML. I have ommitted a lot of details from this summary, as the paper is so rich in information 🧧, I highly suggest everyone to give it a read on https://lnkd.in/ge5uDxRc . It is very well written with an easy to understand language, it was very nice to read through. For the results 🚀, they showed a significant improvement in performance in various QA benchmarks, such as ASQA, Hotpot-QA, NQ, Trivia-QA, and others 📊. This paper comes with open-source code! https://lnkd.in/gZ2t52Rx #AIResearch #MachineLearning #ML #NaturalLanguageProcessing #NLP #RAG #LLMs #KnowledgeRepresentation #OpenSourceCode #PaperReview

35

2 Comentarii