Ashu Dubey’s Post

Creating high-quality instruction data for aligning large language models (LLMs) is challenging. Traditional methods are costly, time-consuming, and heavily reliant on human effort, making it difficult to scale effectively. This limits the diversity and quality of public alignment datasets. I recently came across a compelling paper by Zhangchen Xu, Fengqing Jiang, and their team, which proposes an innovative solution to this problem called MAGPIE. MAGPIE synthesizes high-quality alignment data for LLMs without relying on predefined queries or extensive prompt engineering. 𝐓𝐡𝐞 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧: MAGPIE leverages the auto-regressive nature of aligned LLMs to generate user queries and responses from pre-query templates. This innovative approach allows for creating large-scale, high-quality instruction datasets with minimal human intervention. MAGPIE generated 4 million instructions and responses, from which 300K high-quality instances were meticulously selected. 𝐇𝐨𝐰 𝐢𝐭 𝐖𝐨𝐫𝐤𝐬:  Pre-query templates are input into Llama-3-Instruct, which generates corresponding user queries. These queries are fed into the LLM to produce high-quality responses, forming complete instruction-response pairs. 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐢𝐧𝐠 𝐚𝐧𝐝 𝐂𝐨𝐦𝐩𝐚𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐧𝐚𝐥𝐲𝐬𝐢𝐬: MAGPIE's data was benchmarked against various public datasets like ShareGPT and WildChat. Models fine-tuned with MAGPIE data performed comparably to the official Llama-3-8B-Instruct, which utilized 10 million data points for supervised fine-tuning (SFT) and subsequent feedback learning. 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐌𝐞𝐭𝐫𝐢𝐜𝐬:  MAGPIE outperformed previous public datasets on alignment benchmarks such as AlpacaEval, ArenaHard, and WildBench. It maintained strong performance on reasoning tasks like MMLU-Redux despite the "alignment tax". 𝐄𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲 𝐚𝐧𝐝 𝐒𝐜𝐚𝐥𝐚𝐛𝐢𝐥𝐢𝐭𝐲:  The data generation process required only 206 GPU hours for MAGPIE-Air and 614 GPU hours for MAGPIE-Pro, highlighting its cost-effectiveness and scalability. MAGPIE represents a significant advancement in creating scalable, high-quality instruction datasets crucial for fine-tuning LLMs. This method reduces reliance on human-generated data and enhances the interpretability and quality of embeddings. Here is the link to the paper: https://lnkd.in/gmV5btzK #AI #MachineLearning #LLM #MAGPIE #DataSynthesis #Research #Innovation

You can easily deploy an ollama endpoint with JarvisLabs.ai & generate Magpie style synthetic datasets with just few lines of code. Here's the example code: https://lnkd.in/gKx99jFc

The most challenging aspect I've encountered in optimizing an information retrieval system is accurately and effectively parsing information from raw sources, regardless of how powerful your LLM may be. In this blog, we describe our approach at C3 AI to multi-modal parsing of unstructured documents with complex layouts. #genai #generativeai #llm #rag

DeepSeek V3: The Ultimate Showdown Against Claude 3.5 Sonnet & OpenAI o1 The AI landscape is heating up, and DeepSeek V3 is making waves as the new contender in the ring! With its 671B parameters, innovative MoE architecture, and jaw-dropping cost efficiency (training at just $5.58M!), this open-source powerhouse is challenging the dominance of Claude 3.5 Sonnet and OpenAI’s o1. From coding to math, logic to creativity, DeepSeek V3 has proven its mettle in head-to-head battles, even outperforming its rivals in specific tasks like programming and mathematical problem-solving. But is it the ultimate AI champion? Or does Claude 3.5 Sonnet still hold the crown in nuanced reasoning and style control? Dive into the full breakdown of this epic showdown to discover: ✅ How DeepSeek V3 stacks up in performance, cost, and versatility. ✅ Where it shines—and where it falls short. ✅ Why this could be a game-changer for developers and businesses alike. Whether you’re an AI enthusiast, a developer, or just curious about the future of tech, this is a must-read. Let’s spark a conversation—what’s your take on the AI race? 👉 Read the full article here: https://lnkd.in/dct8YAhd #AI #DeepSeekV3 #Claude3.5 #OpenAI #TechInnovation #ArtificialIntelligence #FutureOfAI

Interesting article highlighting the potential of domain-adapted foundation GenAI models from LinkedIn Engineering team, who was able to build a domain-adapted Llama 3.1-8B variant which was 6X cheaper compared to GPT 4o while keeping was comparable to State of the Art GPT models. https://lnkd.in/g-GfP9yr #genai#llama

🚀 Hot off the press - Coveo strengthens its leadership in GenAI with the launch of Relevance-Augmented Passage Retrieval API🚀 We're thrilled to announce the launch of our new Relevance-Augmented Passage Retrieval API, which empowers organizations to connect their own Large Language Models (LLM) with Coveo’s relevant and robust AI retrieval infrastructure to seamlessly and securely ground their GenAI experiences in the entirety of their enterprise content. "Coveo’s Relevance-Augmented Passage Retrieval API will play a crucial role in an enterprise's AI and Generative AI architecture, enhancing all digital experiences. It’s a game-changer for organizations building intelligent and high-performing large language model (LLM) applications," said Laurent Simoneau, Co-Founder, President, and CTO at Coveo. To learn more about these groundbreaking advancements, join us at our virtual Relevance 360 event on October 3, 2024, or our exclusive in-person Relevance 360 Breakfast during Dreamforce on September 18, 2024, at 7:30am PDT. 🗞️ Read the full press release: https://lnkd.in/epsQHeCb 🔍 Learn more about our Relevance-Augmented Passage Retrieval API: https://lnkd.in/er5hkZQz 🤓 Dive into our blog from Sebastien Paquet, VP of Machine Learning at Coveo, to see how our new Passage Retrieval API innovation is enhancing user experiences: https://lnkd.in/eZTE6Cjj #CoveoAI #generativeAI #RAG #search #APIretrieval #enterpriseAI #LLMs #digitalexperiences

🚀 Coveo Leads the Way in GenAI with the Launch of Our Relevance-Augmented Passage Retrieval API! 🚀 I’m incredibly proud to share this exciting news! Our latest innovation, the Relevance-Augmented Passage Retrieval API, empowers organizations to seamlessly and securely ground their GenAI solutions in their own enterprise content. We’ve not only listened to our customers but also stayed ahead of the competition by delivering a world-class relevance retriever that enables enterprises to share grounded truth across multiple systems. This release reflects the growing trend of Agentic AI, where systems not only retrieve but actively use relevant information to empower decisions and enhance outcomes. It’s more than just retrieval – it’s about actionable insights. On a personal note, my involvement in this project has been particularly rewarding, as I’ve had the opportunity to work hand-in-hand with our customers, sales, professional services, partners, and customer success teams to ensure a smooth transition for early adopters and prepare for the general availability of this groundbreaking feature. I’m thrilled about what’s next! 🗞️ Full press release: https://lnkd.in/epsQHeCb 🔍 Learn more about our API: https://lnkd.in/er5hkZQz 🤓 Blog post by Sebastien Paquet: https://lnkd.in/eZTE6Cjj #CoveoAI #generativeAI #RAG #AgenticAI #search #enterpriseAI #LLMs

🔍 Did you know that AI can 'hallucinate' too? 🤯 Imagine this: You're relying on your AI assistant to filter out the truth from the noise, but what if it's mistakenly generating fiction instead? Introducing HALVA: Hallucination Attenuated Language and Vision Assistant by Google Research—a groundbreaking innovation designed to minimize those pesky AI hallucinations and deliver more accurate results. Here's the lowdown on HALVA and why it’s a game-changer: 👉 It's all about precision: HALVA aims to sharpen the boundary between reality and AI-generated content. 👉 Enhanced reliability: By reducing hallucinations, HALVA ensures you get trustworthy, actionable insights. 👉 Next-gen AI: Leverage the latest in machine intelligence to stay ahead in your industry. Curious about the nitty-gritty details? Dive into the official blog post and see what HALVA can truly do. Read more: [Full article here](https://lnkd.in/gGwnDXZC) 🌟 — Bringing the frontier of AI closer, one breakthrough at a time. 🚀 P.S. How do you think HALVA will revolutionize the way you use AI assistants? Share your thoughts in the comments below! 👇

Retrieval sets the upper bound of RAG performance. Setting up a retrieval system that satisfies all the compliance, security and performance requirements of today's enterprise data management best practices is hard for any organization. Many orgs want to develop their own customized LLM workflow after retrieval, which precludes using an end-to-end RAG solution. This solves for data management and battle-tested retrieval, letting you focus on developing prompts and LLM flows for your particular use case.

If you're not getting relevant results or answers with your current GenAI, RAG, or LLM applications , (including chatbots & copilots) - now you can leverage best-of-breed, market-leading AI Search from Coveo! This will take your results and answers from subpar to high-performing! Exciting times ahead 🙌