Kalyan KS’ Post

📢 Introducing LLM Compressor: SOTA open-sourced framework for compressing LLMs (including Llamas)! 📢 Starting as an internal research project at Neural Magic, I could never imagine that nearly six years later, what I called neuralmagicML would evolve into the incredibly powerful and capable LLM Compressor framework. Thanks to the hard work of our incredible engineering team, this tool can compress LLMs of any size to remarkable levels while recovering full accuracy. (such as our recent Llama 3.1 405B results!) We are thrilled to announce that we have donated this library and its innovative techniques to the vLLM community. By offering efficient, performant, and accurate solutions for large language models, we aim to empower researchers, hackers, and enterprises. Why This Matters: - Cutting-Edge Algorithms: Implement the latest techniques and best practices for top-tier model performance without extensive research. - Superior Flexibility: Enjoy various compression techniques, quantization schemes, and sparsity options for a solution that fits your use cases. - Community-Driven: This tool is open-sourced and seamlessly integrated with vLLM and Hugging Face to ensure compatibility and encourage future contributions. We are excited to witness how the community will leverage this incredible tool! Dive deeper into LLM Compressor by exploring our blog and the repo: - Blog: https://lnkd.in/eHwW5snX - Repo: https://lnkd.in/eTFvDyqF #llms #opensource #ai #generativeAI #llama31

#dailyreport #abtest #multiarmedbandit #mab I have been reading about solving causal inference A/B  test as a Multi-Armed Bandit problem. These is ML  Reinforcement learning algorithms that applied at once  and enhance during time. Popular algorithms: - Upper Confidence Bound (UCB) - deterministic, optimal. - Thompson Sampling - stochastic, optimal. - Epsilon Greedy - stochastic, approximate. Thompson Sampling and UCB have asymptotic regret lower  bound (where N is the number of arms and T is the number  of time steps). : O(√(N*T*log(T))) *regret* is difference between max possible reward and  collected. *optimal* means algoritms able to achive  minimal regret when T → ∞. Follow me at techhub.social/@Anoncheg

[Interpretable RL][Call for papers and reviewers][RLC 2024 workshop] Interpretable and explainable RL is a vastly growing research topic, yet very few venues are dedicated to it. In order to bring the community together, Hector Kohler, Quentil Delfosse, Philippe Preux, and I, are organising a workshop on interpretable reinforcement morning at the coming Reinforcement Learning Conference this summer. Our goal is to foster inclusive discussions on this fascinating topic, here are the main features of the event: 🤖 Part of RLC 2024, the new major conference on feinrorcement learning 🌎 A lineup of international speakers 💬 4h of poster/panel discussions for active exchange 📚 We welcome works at any stage from early results to already published papers (the workshop will not have proceedings) 👁️ We welcome reviews with expertise in RL, get in touch if interested! 📆 Submission deadline: April 26th, 2024 Find out more about the workshop here 👉 https://lnkd.in/enMyBUBt #InterpretableRL #ExplainableAI #ReinforcementLearning #RLC2024 #MachineLearningConference #CallForPapers #AIResearch #ArtificialIntelligence #MachineLearning #DataScienc

One of my favorite quotes from 🔭 Galileo's GenAI Productionize Conference last month was this one by, the astute as always, Craig Wiley : "With GenAI, people tend to forget the 'science' in Data Science". Since the earliest days of applied AI, data scientists have painfully recognized that if you truly want to build an AI powered application that works well, 80% of your time *will* need to be spent iterating, experimenting, A/B comparing and sometimes plain-old human eyeballing. This '80% time' can be massively contracted through powerful evaluation metrics, but it still needs to be done, and it still needs a platform to make it easier. With 'LLM parameter flexing' being the talk of the day across big-tech and 'small-tech-with-big-capital', I often hear AI engineers expecting too much from LLMs, forgetting the need to put in the work into the 'science' of it all. https://lnkd.in/gBj2ycsU?

🧰 First detailed in the “Direct Preference Optimization: Your Language Model is Secretly a Reward Model” paper, DPO presents a simplified yet effective alternative to the intricate processes of Reinforcement Learning from Human Feedback (RLHF) Its stability, performance, and computational lightweight nature makes it an appealing option for projects seeking to align LLMs closely with human judgments. ⚙️ Learn more about DPO, its differences from RLHF, and if it is the best choice for preference tuning LLMs. Read the blog > https://lnkd.in/gnAgjztA #llms #llm #largelanguagemodels #generativeai

PhAI can solve the so-called ‘phase problem’ with lower-quality data than is needed for other methods and quickly arrives at high quality solutions.

With the introduction of #Claude3 by #Anthropic, the ranking of Large Language Models (LLMs) has been updated. Currently, #benchmarks are the most tangible factors for comparing LLMs #LLM https://lnkd.in/dy5BXxex #AI #GenAI https://lnkd.in/dY-jYvfX

𝐒𝐞𝐥𝐞𝐜𝐭𝐋𝐋𝐌 - 𝐂𝐡𝐨𝐨𝐬𝐞 𝐎𝐩𝐭𝐢𝐦𝐚𝐥 𝐋𝐋𝐌𝐬 𝐟𝐨𝐫 𝐐𝐮𝐞𝐫𝐢𝐞𝐬 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐒𝐞𝐥𝐞𝐜𝐭𝐋𝐋𝐌? SELECTLLM is a novel algorithm developed to overcome limitations of individual LLM by choosing appropriate LLMs for a given query. The proposed algorithm helps to harness the diverse capabilities of multiple LLMs to enhance performance on complex tasks. 𝐒𝐞𝐥𝐞𝐜𝐭𝐋𝐋𝐌 𝐌𝐞𝐭𝐡𝐨𝐝𝐨𝐥𝐨𝐠𝐲 SelectLLM utilizes the predictions and confidence scores of a multilabel classifier for selecting the appropriate LLMs. 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐑𝐞𝐬𝐮𝐥𝐭𝐬 SELECTLLM outperforms individual LLMs and achieves competitive results compared to top-performing LLM subsets. It demonstrates significant latency reductions on standard reasoning benchmarks, with 13% lower latency for GSM8K and 70% lower latency for MMLU. #llms #generativeai #deeplearning #nlproc

With high enthusiasm, and strong dedication, I would like to announce the application that we have made from Capstone Project Showcase : SCAN (Smart Classification with Automated Neural Network). Mobile APP-based application that has Machine Learning integration with high accuracy to sort out the types of waste. #LifeAtBangkit #Bangkit24H1 #PMChallenge #GrowWithGoogle

We present our capstone project entitled SCAN (Smart Classification with Automated Neural Network). This project is based on mobile application development, which integrates machine learning technology to automatically identify waste types. #LifeAtBangkit #Bangkit24H1 #PMChallenge #GrowWithGoogle