Armelle Patault

🚀 My First Medium Article is LIVE! 🚀 Dear network 🎉 I’m excited (and nervous!) to share that I’ve just published my very first article on Medium! It's called "The Wizards Behind Data Magic" 🧙‍♂️✨ In this piece, I dive into the different roles in the data world—from Data Analysts to ML Engineers —and how each one plays a crucial part in solving data mysteries. 💡 What's Inside? - A fun, detective-themed breakdown of data roles 🕵️‍♀️ - How data engineers turn big data into "small data" 🛠️🔧 - Why data scientists are the master puzzle solvers 🧩🔍 - My journey toward becoming the "Dumbledore of Data" 🌟 As it's my first post, I’d love your support! 🙏 If you have a moment, please give it a read and let me know what you think. Your encouragement would mean the world to me. 🌱 If you enjoy it, please consider following me on Medium for more insights and stories about data and tech. And I'd love to hear from you—what hat are you wearing in the data world? 🎩 How do you see your role evolving in the AI era? Let's chat in the comments! 💬👥 #FirstMediumPost #DataMagic #DataCommunity #MachineLearning #DataScience #AI #CareerGrowth #DataRoles #YourFeedbackMatters #FollowMeOnMedium

54

6 commentaires

Fine-tuning LLMs: Aligning models with specific customer needs is crucial for real-world success, yet it requires precision to ensure the model adapts effectively without compromising quality. Fine-tuning often introduces risks, such as underfit or overfit layers, which can lead to unpredictable behavior in production. Tools like WeightWatcher are essential for addressing these challenges, offering a data-free diagnostic to evaluate model quality by analyzing the spectral properties of weight matrices. Grounded in the Heavy-Tailed Self-Regularization (HT-SR) theory, WeightWatcher provides insights into layer performance and potential issues, helping developers optimize instruction-tuned models like Llama3.1 and Falcon2 for robust, production-ready use cases. #InstructionFineTuning #LLMEngineering #ModelQuality #CustomerFocusedAI

2

Not long after releasing 2 specialized models, we get another open-source model, Mistral Large 2, from the French 🇫🇷 powerhouse. Closing the gap even more in the debate of close vs. open models. With its 123B parameters, the model focuses on handling detailed and extensive data, allowing it to process a large amount of information on a single node. During training, a key goal was to reduce the model's tendency to produce misleading or incorrect information. By fine-tuning its approach to be more cautious and discerning, the model delivers more reliable and accurate outputs. It's easily accessible through Mistral AI's platform, allowing for straightforward integration into various systems and supporting a range of applications from testing to commercial use. The open-source approach invites a global community of developers, researchers, and enterprises to not only access but also contribute to refining and expanding AI capabilities. Here at Velebit AI, we've been on board with Mistral from its early days, embracing each phase of its evolution. We eagerly anticipate leveraging the latest enhancements brought by its continuous advancements. #opensource #mistralai #machinelearning

25

🚀 Productionizing LLMs: Key Considerations 🚀 : Always remember the key is to make the system measurable and iterative - 1️⃣ Unit Testing: Always start with comprehensive unit tests. As your code evolves, these tests ensure functionality remains intact, providing a solid foundation. 2️⃣ Human Evaluation: Incorporate human checks initially to assess the quality of LLM outputs. Quantify accuracy to set benchmarks for improvement. 3️⃣ Iterative Improvement: Embrace the iterative nature of AI systems. Implement feedback loops and continuously refine the system based on performance metrics. 4️⃣ Logging and Data Gathering: Capture middle steps and interactions for future fine-tuning. This data is crucial for enhancing the accuracy and relevance of your LLM over time. Building AI systems is complex, but by focusing on these principles, we can create robust and adaptable solutions. Let's innovate and iterate! 💡 #AI #SystemDesign #LLM #LLMOps

3

The 2024 Stanford AI Index Report was out a while ago.. a few fun or not-fun revelations/interpretations depending on where you sit : The training costs of state-of-the-art AI models have reached unprecedented levels. OpenAI's GPT-4 used an estimated $78 million worth of compute to train, while Google's Gemini Ultra cost $191 million.. a staggering increase in the computational resources and financial investment required to develop the latest AI models.. Only the beginning, huh.. what say!? Despite AI surpassing human performance on many benchmarks, it still lags behind on more complex tasks like competition-level mathematics, visual commonsense reasoning and planning.. suggests AI still has significant limitations compared to human intelligence.. Only a matter of time, huh.. what say!? The report reveals a significant lack of standardization in responsible AI reporting across leading developers like OpenAI, Google, and Anthropic.. cannot systematically compare the risks and limitations of top AI models, plus fear mongering by top tech companies to keep industry hostage thanks to fearful and old world risk averse leaders calling the shots is at an all time high.. Only job left for human beings then is to call the bluff if they can - call the bluff, huh.. what say!? Notably, the U.S. dominated the development of notable AI models in 2023, producing 61 compared to 21 from the EU and 15 from China.. only metadata here guys, money does what money can huh.. what say!? Industry is dominating frontier AI research, producing 51 notable machine learning models in 2023 compared to only 15 from academia. Industry-academia collaborations also reached a new high. Investment in generative AI technologies surged to $25.2 billion in 2023, despite an overall decline in AI private investment.. Only 25.2B eh, maybe they got the currency wrong on that one huh.. what say!? & closer to heart, AI is transforming FMCG operations and innovation. The report highlights how leading FMCG firms like Unilever, P&G, and Nestlé are leveraging AI to enhance demand forecasting, inventory management, supply chain optimization, and product development. AI is driving significant productivity gains and quality improvements for FMCG workers. Studies cited in the report show AI enables FMCG employees to complete tasks 26-73% faster and improve output quality by up to 40%... only 1 look at the AI Products’ marketplace & that adds up huh, what say Shreyasi Singh Mriganshi Sharma Komal Falodiya Sweta A. Rekha Lath Abhilasha Sinha Quratulain Naseer Darshika Tripathi Krishna Soni S P Ravishankar Vijay Reddy Gopi Krishna M Neelesh Singh Rajpurohit Munna Kumar

25

3 commentaires

These days I’ve been in Montpellier, France, to present our last research results on combinatorial optimization problems over temporal graphs to the 31st symposium on temporal representation and reasoning (TIME2024). Check out our paper at https://lnkd.in/da66ref4 Here we discuss a new heuristic-based approach for minimum timeline cover over temporal graphs. The paper provides theoretical results to a problem largely adopted in social network analysis as interpretable network summarization framework to find important events. Particularly, we used it to analyze job postings, to find the most required skills by companies and understand how the market is evolving. We are already working on evolutions of this approach with Deep Learning. Stay tuned for next publications. #knowledgegraph #combinatorialoptimization #networksummarization

15

1 commentaire

Meta has just released their long-awaited #Llama #3 models. Everyone's talking about it, let me give it a try 😅! Some observations: ☑ #Llama #8B and #70B are available, and the #400B is currently in training ⁉ Yes... well, who cares, right? We see new models every morning? Well, not really, these models are a bit different. Not in terms of innovation – practically the same as #Llama #2 – but rather in the meticulous work done on the quality and quantity of data. ✅ For the first time, we have a model of 8B (and 70B as well) trained with more than 15 trillion tokens! For comparison, #Llama #1 used 1T, and #Llama #2 used 2T. That's 8 times more data. To give you an idea, 15T tokens =~ 69 terabytes. Training an #8B model with 15T far exceeds the norms; Chinchilla Scaling suggests 20 tokens per parameter, which would require 160B tokens for optimal training of a model this size. Here we are talking about 15T, which is 95 larger, that's just astronomical. Is it beneficial? it seems to be the case, at most, it won't hurt. ☑ 5% of the texts are not in English, which is not ideal, but still acceptable because LLMs generalize beyond the training language. ☑ More code in pretraining, as it is proven to improve the reasoning capabilities of LLMs. ☑ What about compute? Training the 8B model required 1.3 million GPU hours (I'm not sure if this includes failures and restarts, or if it's purely 1.3M of training), and the 70B was trained on 6.4 million hours. In total, 7.7 million hours, so at a rate of $2 per hour in a H100 GPU, that amounts to $15 million, not counting the additional costs (engineers, infrastructure, etc.). ☑ Why are these models cool? Because #Llama 1 was the main driver of the proliferation of open-source LLMs (open-weight to be precise). After #Llama 1 (7-65B) and 2 (7-70B), we've seen an explosion of open-source models and research that wouldn't have been possible without these models. ☑ What's next? The next step is #Llama 400B, which is already closing the gap with #GPT4 😲. We will then have all the quantizations, fine-tunings, etc., and we will achieve performances almost close to #GPT4 with smaller models. I thought we'd see this in a year! This will surely motivate OpenAI to release #GPT5 :D. ☑ We will definitely see #Llama with larger context windows than 8k tokens, because 8k is not satisfying. And you? Have you tested these models? Eager to hear your comments on their performance. #LLMs #Llama3 #Llama #Meta

13

2 commentaires

Another open source model, Moshi!! #llm #genai #ai

7

1 commentaire

I encourage you all to try the demo (link at bottom) - mind blown!

1

Presenting Orpomis Prime 7B model series ! These models are a result of trying different merge strategies of the following models: - mistral-orpo-beta from KAIST AI - Hermes-2-Pro-Mistral-7B from NousResearch In my experiments, dare ties merge leads to a significantly better-performing model on the openllm benchmarks In general, the model performance is quite solid in text generation as well as math reasoning. Full result comparison: https://lnkd.in/gb92DXwG Dare ties Model: https://lnkd.in/gMS-H4Sq Model stock Model: https://lnkd.in/giwEJ7Gs

9

Discover What Claude Can Do That ChatGPT Can't! 🤯🚀"imulated 3d physics using WebGL. I have tested Claude Sonne and it is excellent compared to #ChatGPT and #Gemini, among others. It is crucial to identify the differences between each one, especially in terms of context windows, tokenization, and other important variables. For now, Look at what can be done with Claude now! Something that is not yet possible with #ChatGPT: If you want to be part of the AI revolution, share, learn, and follow me, Henry Jiménez.

70

23 commentaires

NVIDIA released a powerful AI model with 70 billion parameters that outperforms GPT-4 and Claude 3.5 Sonnet. The best part? It's open-source, meaning anyone can use it. You can try it : https://lnkd.in/dinKV-As

7

↑Jump starter Insights I 𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐋𝐋𝐌𝐬 Referencing : What We Learned from a Year of Building with LLMs, Part II, by Charles Frye,Eugene Yan, Bryan Bischof,Shreya Shankar & Jason Liu I truly appreciate this series and eagerly anticipate future installments to learn even more from your experiences. While each of us may derive different insights based on our individual perspectives as builders, here are my top five takeaways: ↸𝐎𝐧 𝐌𝐨𝐝𝐞𝐥 𝐋𝐢𝐦𝐢𝐭𝐚𝐭𝐢𝐨𝐧𝐬 𝐚𝐧𝐝 𝐂𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬 Detailed assessment of LLMs' performance highlights their proficiency in generating coherent text and simulating human-like conversation. 𝐻𝑜𝑤𝑒𝑣𝑒𝑟, 𝑖𝑡 𝑎𝑙𝑠𝑜 𝑢𝑛𝑑𝑒𝑟𝑠𝑐𝑜𝑟𝑒𝑠 𝑡ℎ𝑒𝑖𝑟 𝑝𝑟𝑜𝑝𝑒𝑛𝑠𝑖𝑡𝑦 𝑓𝑜𝑟 𝑝𝑟𝑜𝑑𝑢𝑐𝑖𝑛𝑔 𝑖𝑛𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑖𝑒𝑠, ℎ𝑎𝑙𝑙𝑢𝑐𝑖𝑛𝑎𝑡𝑖𝑜𝑛𝑠, 𝑎𝑛𝑑 𝑐𝑜𝑛𝑡𝑒𝑥𝑡𝑢𝑎𝑙𝑙𝑦 𝑖𝑟𝑟𝑒𝑙𝑒𝑣𝑎𝑛𝑡 𝑖𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛. ↸𝐎𝐧 𝐓𝐚𝐫𝐠𝐞𝐭𝐞𝐝 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧 𝐀𝐫𝐞𝐚𝐬 Strategic focus on application domains where LLMs demonstrate high utility, such as automating repetitive tasks, enhancing creative processes, and enabling rapid prototyping for concept validation. ↸𝐎𝐧 𝐇𝐮𝐦𝐚𝐧-𝐀𝐈 𝐒𝐲𝐧𝐞𝐫𝐠𝐲 Emphasis on designing systems that facilitate effective human-AI collaboration, wherein human oversight and contextual understanding are integral to refining AI-generated outputs. ↸𝐎𝐧 𝐄𝐭𝐡𝐢𝐜𝐚𝐥 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤𝐬 Addressing ethical issues such as algorithmic bias, the propagation of misinformation, and ensuring the transparency and accountability of AI systems. This involves implementing robust bias mitigation strategies and fostering a culture of ethical AI development. ↸𝐎𝐧 𝐈𝐭𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐃𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 𝐚𝐧𝐝 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧 Advocating for a continuous cycle of experimentation, feedback, and refinement to iteratively enhance the performance and reliability of LLM deployments. This includes leveraging user feedback and real-world data to fine-tune models. ☞𝐒𝐨𝐮𝐫𝐜𝐞 Read https://lnkd.in/gNHSjkWb #LLMBuild #LLM #AIBuild

7

🚀 𝐔𝐧𝐯𝐞𝐢𝐥𝐢𝐧𝐠 𝐭𝐡𝐞 𝐌𝐚𝐠𝐢𝐜 𝐨𝐟 𝐅𝐢𝐧𝐞-𝐓𝐮𝐧𝐢𝐧𝐠 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 (𝐋𝐋𝐌𝐬) 🎯 🔗 𝐑𝐞𝐚𝐝 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥 𝐬𝐭𝐨𝐫𝐲 𝐨𝐧 𝐌𝐞𝐝𝐢𝐮𝐦: From Generic to Genius: The Art of Fine-Tuning Large Language Models Have you ever wondered how Large Language Models (LLMs) like GPT or BERT transform from being generalists into specialists for tasks like medical diagnostics, sentiment analysis, or legal insights? 𝐈𝐧 𝐦𝐲 𝐥𝐚𝐭𝐞𝐬𝐭 𝐌𝐞𝐝𝐢𝐮𝐦 𝐚𝐫𝐭𝐢𝐜𝐥𝐞, I take you on a storytelling journey through the art of fine-tuning LLMs. Discover: ✅ 𝐖𝐡𝐚𝐭 𝐟𝐢𝐧𝐞-𝐭𝐮𝐧𝐢𝐧𝐠 𝐢𝐬 𝐚𝐧𝐝 𝐰𝐡𝐲 𝐢𝐭’𝐬 𝐚 𝐠𝐚𝐦𝐞-𝐜𝐡𝐚𝐧𝐠𝐞𝐫. ✅ 𝐇𝐨𝐰 𝐭𝐨 𝐩𝐫𝐞𝐩𝐚𝐫𝐞 𝐝𝐚𝐭𝐚𝐬𝐞𝐭𝐬 𝐚𝐧𝐝 𝐢𝐧𝐭𝐞𝐠𝐫𝐚𝐭𝐞 𝐜𝐮𝐬𝐭𝐨𝐦 𝐜𝐥𝐚𝐬𝐬𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧 𝐡𝐞𝐚𝐝𝐬. ✅ 𝐒𝐭𝐞𝐩-𝐛𝐲-𝐬𝐭𝐞𝐩 𝐜𝐨𝐝𝐞 𝐬𝐧𝐢𝐩𝐩𝐞𝐭𝐬 𝐟𝐨𝐫 𝐥𝐨𝐚𝐝𝐢𝐧𝐠 𝐦𝐨𝐝𝐞𝐥𝐬, 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠, 𝐚𝐧𝐝 𝐭𝐞𝐬𝐭𝐢𝐧𝐠. ✅ 𝐏𝐫𝐚𝐜𝐭𝐢𝐜𝐚𝐥 𝐭𝐢𝐩𝐬 𝐭𝐨 𝐭𝐮𝐫𝐧 𝐠𝐞𝐧𝐞𝐫𝐚𝐥-𝐩𝐮𝐫𝐩𝐨𝐬𝐞 𝐋𝐋𝐌𝐬 𝐢𝐧𝐭𝐨 𝐭𝐚𝐬𝐤-𝐬𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐠𝐞𝐧𝐢𝐮𝐬𝐞𝐬. ✨ Fine-tuning isn't just about adjusting models; it's about unlocking their full potential for real-world problem-solving. 👩‍💻 Whether you're a 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡𝐞𝐫, 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐫, 𝐨𝐫 𝐞𝐧𝐭𝐡𝐮𝐬𝐢𝐚𝐬𝐭, this guide simplifies the technical steps and presents them as an engaging story. 🔍 𝐓𝐚𝐤𝐞 𝐚 𝐥𝐨𝐨𝐤 𝐚𝐧𝐝 𝐬𝐡𝐚𝐫𝐞 𝐲𝐨𝐮𝐫 𝐭𝐡𝐨𝐮𝐠𝐡𝐭𝐬: 📖 𝐑𝐞𝐚𝐝 𝐭𝐡𝐞 𝐚𝐫𝐭𝐢𝐜𝐥𝐞: https://lnkd.in/g2wEW2gr 🔄 Repost to inspire others in your network. 💬 Let’s discuss your fine-tuning experiences or challenges in the comments! #LLM #FineTuning #GenerativeAI #MachineLearning #AIForGood #TheLLMInsider

32

Liquid AI recently introduced Liquid Foundational Models (LFM) : A non-GPT based Generative AI Models Is this could be a next frontier in LLMs ? You see LLMs have two major issues : 1. Memory: They require billions of dollars of computational resources 2. Low Customisation for Edge Devices: Inference with LLMs is memory-intensive even for specific use cases like summarising, analysing, etc, that are most use cases for edge devices. Think about it, we don't need billion parameter models for specific use cases to serve edge-devices like mobiles. 100M parameter models customised to our use case can serve the purpose. This is what LFMs have introduced. Liquid AI has introduced 3 Models : 1. LFM-1.3B : Ideal for resource-constrained environments, new state-of-the-art model in 1B category 2. LFM-3.1B : Optimised for edge deployment (like mobile devices, etc), first place among 3B category models, outperforms previous generation of 7B and 13B models as well 3. LFM-40.3B : Mixture of Experts (MoE) model designed for tackling more complex tasks. (this is just cost effective model) Full Comparison - https://lnkd.in/gGG-HSmB

22

Want to test Meta, Google, and Mistral models all on one platform ? **Groq.com** Why Groq? Because speed matters. GroqCloud supports a diverse array of models, enabling you to select the best for your needs: ➡ llama3-8b-8192 by Meta (Context Window: 8,192 tokens) ➡ llama3-70b-8192 by Meta (Context Window: 8,192 tokens) ➡ mixtral-8x7b-32768 by Mistral (Context Window: 32,768 tokens) ➡ gemma-7b-it by Google (Context Window: 8,192 tokens) Groq's commitment to GenAI inference speed is transforming real-time AI applications with its LPU. What is an LPU? ➡ Language Processing Unit, is a revolutionary processing system designed specifically for LANGUAGE applications. Unlike traditional CPUs (Central Processing Units) and GPUs (Graphics Processing Units), which handle a broad range of tasks. LPUs are optimized for the rapid processing demands of LLMs This specialized architecture allows for more efficient handling of sequential data, significantly reducing the time per word processed. ➡ That means you get faster outputs You can run Llama-2 70B at over 300 tokens per second per user!

24

"𝗪𝗵𝗮𝘁 𝘁𝗵𝗲 𝗵𝗲𝗰𝗸 𝗮𝗿𝗲 𝗽𝗲𝗼𝗽𝗹𝗲 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗱𝗼𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝘁𝗵𝗲𝗶𝗿 𝗖𝗹𝗼𝘂𝗱 𝗗𝗮𝘁𝗮 𝗪𝗮𝗿𝗲𝗵𝗼𝘂𝘀𝗲?" is a question we asked ourselves when we joined the Redshift team. We have since found out, and publish as many findings as we can in our VLDB2024 paper "𝗪𝗵𝘆 𝗧𝗣𝗖 𝗶𝘀 𝗻𝗼𝘁 𝗲𝗻𝗼𝘂𝗴𝗵: 𝗔𝗻 𝗮𝗻𝗮𝗹𝘆𝘀𝗶𝘀 𝗼𝗳 𝘁𝗵𝗲 𝗔𝗺𝗮𝘇𝗼𝗻 𝗥𝗲𝗱𝘀𝗵𝗶𝗳𝘁 𝗳𝗹𝗲𝗲𝘁", together with a comparison to the industry-standard TPC benchmarks for analytical systems. https://lnkd.in/dAPys5H4 Next to our analysis of fleet data, we open source 𝙍𝙚𝙙𝙨𝙚𝙩, a dataset of query logs containing statistics over all queries that ran on a sample of Redshift instances over the span of three months. With Redset we want to specifically encourage drawing further conclusions about real world workloads, conducting research on ML-based techniques such as query prediction models and workload forecasting and creating novel benchmarks that closer mimic the observed real-world workloads.  https://lnkd.in/dZMdrkKF Thanks to my co-authors Alexander van Renen, Dominik Horn, Kapil Vaidya, Wenjian Dong, Balakrishnan(Murali) Narayanaswamy, Zhengchun Liu, Gaurav Saxena, Gaurav Saxena, and Tim Kraska, and members of the Redshift team Mohammed Al-Kateb, Matt Abrams, Naresh Chainani, George Erickson, Bruce McGaughy, Davide Pagano, Ippokratis Pandis, and Rahul Pathak for making this possible.

42

1 commentaire

In some areas, the factor of responsibility will remain one of the most important, and AI will not replace standard procedures so quickly. Nevertheless, it is extremely important to conduct such trials and check when technology can already help. It’s nice to read about it! #AI #medicine #biomedical #trials #technology

2

Sometimes I just can't share all my thoughts on LinkedIn. That's why I've launched my own newsletter : https://lnkd.in/eVukhBqB I've just published my first post, where I dive into what TL;DR is and how to craft them effectively : https://lnkd.in/eKTNkeiA Give it a read, and if you enjoy it, be sure to subscribe for more exciting content in the future! #newsletter #datascience

19