11 Trending LLM Topics Coming to ODSC West 2024
As one of the most rapidly developing fields in AI, the capabilities for and applications of Large Language Models (LLMs) are changing and growing continuously. It can be hard to keep on top of all the advancements. At ODSC West this October 29th-31st, you’ll find a wide range of workshops, tutorials, and talks on LLMs and RAG. Check out a few of them below.
Large Model Quality and Evaluation
Anoop Sinha | Research Director, AI & Future Technologies | Google
Large model development faces many challenges when it comes to ML quality and evaluation, including the coverage, scale, and wide use cases for what LLMs are used for. In this presentation, you’ll explore the challenges of evaluating large models, with case examples from large language models (LLMs) and large multimodal models in particular.
You’ll discuss developing and using benchmarks that are robust to data contamination, creating human-based evaluation criteria, and assessing the responsibility of LLMs and highlight some promising approaches to evaluating LLMs, including using benchmarks, and human input, considering factors like complexity, multilinguality, and responsibility.
Analyzing Unstructured Data at 10B Scale with a Vector Database
Frank Liu | Head of AI & ML | Zilliz
A Popular choice for LLMs, vector databases are broadly applicable for a variety of different types of unstructured data. In this talk, you’ll discuss challenges associated with embedding and analyzing unstructured data at a 10B scale, diving into the vector database features that are important, and the pitfalls to watch out for. We’ll also discuss the myriad of applications for vector search beyond RAG.
How We Trained and Evaluated the First 1M+ Context Length Open Source Model
Mark Kim-Huang | Cofounder, Chief Architect | Gradient
In this presentation, you’ll explore how Gradient’s team applied techniques such as RoPE scaling, ring attention, and specialized network topologies for maximizing memory bandwidth to extend the context length of Llama-3 to 1M tokens.
You’ll go through how they iterated on their empirical experiments to discover how they needed to combine theta scaling (RoPE) in the presence of a limited training token regime to improve the sample efficiency of the model and how we constructed the dataset used for training the long context language understanding capabilities.
You’ll discuss how the field of long context evaluations is still in its infancy and the benchmarks that we considered in reviewing the performance of our model with a particular focus on the Needle in a Haystack evaluation as well as the newer more comprehensive RULER suite of evaluations.
Fine Tuning Strategies for Language Models and Large Language Models
Kevin Noel | AI Lead at Uzabase Speeda | Uzabase Japan-US
Language Models (LM) and Large Language Models (LLM) have proven to have applications across many industries. However, the specialized nature of business requires more precise control and higher accuracy to meet both business cases as well as specific organizational requirements.
In this presentation, you’ll discuss the fine-tuning mechanisms of LM and LLM, with a focus on the fundamental mechanisms behind them as well as the various trade-offs in real-world applications.
The Developers Playbook for Large Language Model Security
Steve Wilson | Chief Product Officer | Exabeam
As Gen AI technologies rapidly advance, the potential risks and vulnerabilities associated with Large Language Models (LLMs) become increasingly significant. This talk provides a comprehensive framework for securing LLM applications. You’ll gain a deep understanding of common vulnerabilities, such as prompt injection, training data poisoning, model theft, and overreliance on LLM outputs.
The session will explore real-world case studies and actionable best practices, illustrating how LLM applications can be safeguarded against these threats. Through examples of past security incidents, both from real-world implementations and speculative scenarios from popular culture, You’ll see the potential consequences of unaddressed vulnerabilities. This talk will also cover the implementation of the RAISE framework, which stands for Responsible AI Security Engineering, designed to provide a step-by-step approach to building secure and resilient AI systems.
RAG in 2024: Advancing to Agents
Laurie Voss | VP, Developer Relations | LlamaIndex
In this session, you’ll learn why retrieval-augmented generation (RAG) is necessary, but not sufficient and you need to add agentic strategies to your system.
You’ll discuss the basic components of an agentic system including routing, memory, planning, reflection, and tool use. You’ll also cover agentic reasoning strategies including sequential, DAG-based, and tree-based, as well as further extensions to agents including observability, controllability, and customizability.
Self-Discover: Large Language Models Self-Compose Reasoning Structures
Steven Zheng, PhD | Senior Staff Software Engineer | Google DeepMind
In this presentation, you’ll explore SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. SELF-DISCOVER substantially improves GPT-4 and PaLM 2’s performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH. It also outperforms inference-intensive methods, such as CoT-Self-Consistency by more than 20%, while requiring 10–40x fewer inference compute. You’ll also see that the self-discovered reasoning structures are universally applicable across model families: from PaLM 2-L to GPT-4, and from GPT-4 to Llama2, and share commonalities with human reasoning patterns.
Bitter Lessons Learned While Building Production-quality RAG Systems for Professional Users of Academic Data
Jeremy Miller | Product Manager, Academic AI Platform | Clarivate
The gap between a RAG Demo and a Production-Quality RAG System remains stubbornly difficult to cross. This talk will cover the critical challenges faced and steps needed when transitioning from a demo to a production-quality RAG system for professional users of academic data, such as researchers, students, librarians, research officers, and others. You’ll explore the criticality of data quality and availability, making data accessible through APIs, and techniques for making data GenAI-ready. You’ll learn solutions for leveraging the strengths of vector databases while avoiding their biggest pitfalls and discuss our experience leveraging data from many disparate databases to form a coherent context for a RAG system.
Develop LLM Powered Applications with LangChain and LangGraph
Eden Marco | LLM Specialist | Google
This engaging and intensive hands-on workshop is designed to unleash the power of LLM agents using the LangGraph library by LangChain. You’ll dive deep into the advanced capabilities of LangGraph, exploring its integration with LangChain to create robust, efficient, and versatile LLM solutions. You’ll also get a comprehensive introduction to key components such as LCEL, multi-agents, reflection agents, Reflexion agents, and more, as well as hands-on experience with advanced RAG architectures.
Streamlining Unstructured Data for Retrieval Augmented Generation
Matt Robinson | Open Source Tech Lead | Unstructured
In this talk, you’ll explore the complexities of handling unstructured data, and offer practical strategies for extracting usable text and metadata from unstructured data. You’ll go over data ingestion from multiple sources, preprocessing unstructured data into a normalized format that facilitates uniform chunking across various file types, and talk about metadata extraction.
By the end of the session, you will have practical insights into organizing your data preprocessing more effectively, enabling you to make better use of unstructured data for RAG applications.
How LLMs Might Help Scale World-Class Healthcare to Everyone
Vivek Natarajan | Research Lead | Google
Transformers and Large Language Models have the potential to act as care multipliers, help improve our understanding of biology, and solve the burden of diseases. In this talk, you’ll learn about recent work done by the speaker’s team at Google AI: Med-PaLM, Med-PaLM 2, Med-PaLM M, AMIE, and Med-Gemini.
Learn about the motivation, principles, and technical innovations underpinning these systems. And explore how these systems might be leveraged to help scale world-class healthcare to everyone and make medicine a humane endeavor again.
Sign me up!
Don’t miss your chance to dive deep into all things LLMs at the leading AI technical conference. Plus you’ll save 40% on your pass when you register by this Friday!
Originally posted on OpenDataScience.com
Read more data science articles on OpenDataScience.com, including tutorials and guides from beginner to advanced levels! Subscribe to our weekly newsletter here and receive the latest news every Thursday. You can also get data science training on-demand wherever you are with our Ai+ Training platform. Interested in attending an ODSC event? Learn more about our upcoming events here.
