Introduction to Large Language Model Customization.pdf

•

0 likes•19 views

Zilliz

An introduction to LLMs, RAG, Fine-tuning, and examples

Technology

1 | © Copyright 2024 Zilliz
1
Yujian Tang | Zilliz
Introduction to LLM
Customization

2 | © Copyright 2024 Zilliz
2
Yujian Tang
Senior Developer Advocate, Zilliz
yujian@zilliz.com
https://www.linkedin.com/in/yujiantang
https://www.twitter.com/yujian_tang
Speaker

3 | © Copyright 2024 Zilliz
3
01 Introduction to LLMs
CONTENTS
03
04 Examples
02 RAG
Fine Tuning

4 | © Copyright 2024 Zilliz
4
01 Introduction to LLMs

5 | © Copyright 2024 Zilliz
5
A Basic Neural Net

6 | © Copyright 2024 Zilliz
6
A Recurrent Neural Network

7 | © Copyright 2024 Zilliz
7
A Transformer Architecture

10 | © Copyright 2024 Zilliz
10
RAG
RAG
Inject your data via a vector
database like Milvus/Zilliz
Query LLM
Milvus
Your Data
Primary Use Case
- Factual Recall
- Forced Data Injection
- Cost Optimization
Embed

11 | © Copyright 2024 Zilliz
11
What Does Vector Data Look Like?

12 | © Copyright 2024 Zilliz
12
Find Semantically Similar Data
Apple made profits of $97 Billion in 2023
I like to eat apple pie for profit in 2023
Apple’s bottom line increased by record numbers in 2023

13 | © Copyright 2024 Zilliz
13
But wait! There’s more!

14 | © Copyright 2024 Zilliz
14
RAG lets us inject data via semantic similarity
provided by vector databases like Milvus

15 | © Copyright 2024 Zilliz
15
03 Fine Tuning

16 | © Copyright 2024 Zilliz
16
RAG vs Fine Tuning
LLM
Fine Tuning
Augment an LLM by training it
on your data
Your Data
“New” LLM
Query
Primary Use Case
- Style transfer
- Domain specific usage

19 | © Copyright 2024 Zilliz
19
Types of Fine Tuning
- Full fine tuning
- LoRA
- QLoRA

20 | © Copyright 2024 Zilliz
20
Fine Tuning Methods
- Supervised Fine Tuning
- Direct Preference Optimization
- 𝚿 (Identity) Preference Optimization
- Odds-Ratio Preference Optimization

21 | © Copyright 2024 Zilliz
21
Fine tuning lets us train LLMs to operate in certain
styles or domains

22 | © Copyright 2024 Zilliz
22
04 Examples
Give Milvus a
Star!

23 | © Copyright 2024 Zilliz
23
RAG without OpenAI project

24 | © Copyright 2024 Zilliz
24
AI Agents Projects

25 | © Copyright 2024 Zilliz
25
Fine Tuning Library

26 | © Copyright 2024 Zilliz
26
Start building
with Zilliz Cloud today!
zilliz.com/cloud

Similar to Introduction to Large Language Model Customization.pdf

Linthicum state of-the-art-cloud-platforms

David Linthicum

IDC datacenter of the future : Oracle point of view

Riccardo Romani

The rise of data and the new economy has led to a paradigm shift that is redefining our world. In today's digital age, information reigns supreme as currency for businesses looking towards an accelerated productivity level with advanced technologies in place; this will allow you to be more competitive by boosting efficiency across all departments at once! The most common data challenges faced by businesses are talked about in detail during this session. You will learn how to overcome them and get practical tips that can help your company succeed. You will gain new insights into the following topics: 1. Avoiding breakdowns in information flows throughout the organization 2. Optimization processes and connecting data silos 3. Making the technology work for your data flow 4. Giving people the right tools to communicate and collaborate 5. Designing effective education of your users to support data sharing across the business You're going to see the technological stack and the strategy I used to develop OpenBOM platform.

PI DX 2020 Atlanta - Data Management Strategy. _ How Do You Establish a Commo...

Oleg Shilovitsky

IBM i Development: Increase Accuracy and Efficiency with SEQUEL's ABSTRACT a...

HelpSystems

Hybrid Cloud Keynote

gcamarda

Hybrid cloud-for-flexible-accelerated-and-sustainable-it16-10-051475673810

Netmagic Solutions Pvt. Ltd.

Exploring Multimodal Embeddings with Milvus

Zilliz

Secure your cloud applications by building solid foundations with enterprise ...

Vladimir Jirasek

Secure Clouds are Happy Clouds

2nd Watch

Neo4j & AWS Bedrock workshop at GraphSummit London 14 Nov 2023.pptx

Neo4j

How to run Real Time processing on Big Data / Ron Zavner (GigaSpaces)

Ontico

7 Things You Need to Know for Your Cloud-First Strategy

Flexera

Achieving digital transformation with Siebel CRM and Oracle Cloud

Sonia Wadhwa

Leading in the Cloud – Oracle Modern Solution

Mohammed Mojibur Raheman

A New Day for Oracle Analytics

Rich Clayton

Organizations everywhere are looking for the best ways to leverage cloud technologies to maximize efficiencies, minimize costs, and increase agility. Research shows that a majority of enterprises are taking a hybrid cloud approach to leverage existing hardware or to meet performance, compliance, and security requirements. In this webinar, we cover key considerations for building and managing a private or hybrid cloud. This includes a discussion of determining appropriate workloads for hybrid clouds, customer examples, and lessons learned. Key topics: 1. Considerations and best practices for designing and implementing a private or hybrid environment. 2. Customer examples of hybrid implementations and key lessons learned. 3. How to easily include virtualized environments into your hybrid cloud plans and implementation. 4. Cost management and analytics for your hybrid cloud. Watch the webinar recording at: http://video.rightscale.com/medias/sej2563yat

RightScale Webinar: Hybrid Cloud Fundamentals and Lessons Learned

RightScale

Final business intelligence in the cloud

Hossam Hassanien

Today the terms "Big Data" and "Internet of Things" draw a lot of attention, but behind the hype there's a simple story. For decades, companies have been making business decisions based on traditional "enterprise data". Beyond that critical data, however, is a potential treasure trove of additional data: weblogs, social media, email, sensors, photographs and much more that can be mined for useful information. More and more organizations are therefore looking to include non-traditional yet potentially very valuable data with their traditional enterprise data in their business intelligence analysis. As the world's most popular open source database, and the leading open source database for Web-based and Cloud-based applications, MySQL is a key component of numerous big data platforms. This presentation explores how you can unlock extremely valuable insights using MySQL with the Hadoop platform.

Unlocking Big Data Insights with MySQL

Matt Lord

Neo4j Keynote: The Art of the Possible with Graph Technology

Neo4j

Decoding Cloud for the Non-IT Executive

Information Services Group (ISG)

Similar to Introduction to Large Language Model Customization.pdf (20)

Linthicum state of-the-art-cloud-platforms

IDC datacenter of the future : Oracle point of view

PI DX 2020 Atlanta - Data Management Strategy. _ How Do You Establish a Commo...

IBM i Development: Increase Accuracy and Efficiency with SEQUEL's ABSTRACT a...

Hybrid Cloud Keynote

Hybrid cloud-for-flexible-accelerated-and-sustainable-it16-10-051475673810

Exploring Multimodal Embeddings with Milvus

Secure your cloud applications by building solid foundations with enterprise ...

Secure Clouds are Happy Clouds

Neo4j & AWS Bedrock workshop at GraphSummit London 14 Nov 2023.pptx

How to run Real Time processing on Big Data / Ron Zavner (GigaSpaces)

7 Things You Need to Know for Your Cloud-First Strategy

Achieving digital transformation with Siebel CRM and Oracle Cloud

Leading in the Cloud – Oracle Modern Solution

A New Day for Oracle Analytics

RightScale Webinar: Hybrid Cloud Fundamentals and Lessons Learned

Final business intelligence in the cloud

Unlocking Big Data Insights with MySQL

Neo4j Keynote: The Art of the Possible with Graph Technology

Decoding Cloud for the Non-IT Executive

More from Zilliz

We present an architecture of embedding models, vector databases, LLMs, and narrow ML for tracking global news narratives across a variety of countries/languages/news sources in https://asknews.app/. As an example, we explore the real-time application of this architecture for tracking the news narrative surrounding the death of Russian opposition leader Alexei Navalny coming from Russian, French, and English sources

Emergent Methods: Multilingual narrative tracking in the news - real-time exp...

Zilliz

"I see eyes in my soup": How Delivery Hero implemented the safety system for ...

Zilliz

Zilliz - Overview of Generative models in ML

Zilliz

Integrating Multimodal AI in Your Apps with Floom

Zilliz

Build streaming LLM with Timeplus and Zilliz

Zilliz

Beyond Retrieval Augmented Generation (RAG): Vector Databases

Zilliz

Voyage AI: cutting-edge embeddings and rerankers for search and RAG

Zilliz

If you are building a RAG application that serves millions of users, you should consider how to scale your system seamlessly and cost-efficiently. The Zilliz Serverless tier represents a significant innovation in the field of vector search, enabling you to rapidly scale to millions of tenants and billions of vectors, while fully leveraging the hot/cold characteristics across tenants to reduce data storage costs. It enables vector storage at costs comparable to S3 and facilitates vector search times in the hundreds of milliseconds for tens of millions of data points! In this talk, we will delve into the implementation details, usage patterns, and performance metrics of Zilliz Serverless. We will discuss how it empowers AI-native applications to achieve rapid business growth by providing a cost-effective and scalable vector storage and search solution.

Leverage Zilliz Serverless - Up to 50X Saving for Your Vector Storage Cost

Zilliz

Embeddings have become a crucial component in contemporary vector search and Retrieval Augmented Generation (RAG) systems. In this talk, I aim to provide a comprehensive overview of training a versatile embedding model, strategies for encoding longer information within such models, along their benefits and limitations. Additionally, I'll delve into various forms of deep learning-powered retrievers.

Training state-of-the-art general text embedding

Zilliz

The rise of Large Language Models has revolutionized the landscape of AI, unlocking huge potential across society. However, it has also introduced the challenge of hallucinations - instances where the model generates rather trippy content in a scarily convincing way. Rest assured, Morena will guide you through an exploration of how we can automatically detect these instances of hallucination to fully unleash the potential of LLMs.

Fact vs. Fiction: Autodetecting Hallucinations in LLMs

Zilliz

VectorDB Schema Design 101 - Considerations for Building a Scalable and Perfo...

Zilliz

Voyage AI Embedding Models for Retrieval Augmented Generation

Zilliz

Chat with your data, privately and locally

Zilliz

Introducing Milvus and new features in 2.4 release

Zilliz

More from Zilliz (14)

Emergent Methods: Multilingual narrative tracking in the news - real-time exp...

"I see eyes in my soup": How Delivery Hero implemented the safety system for ...

Zilliz - Overview of Generative models in ML

Integrating Multimodal AI in Your Apps with Floom

Build streaming LLM with Timeplus and Zilliz

Beyond Retrieval Augmented Generation (RAG): Vector Databases

Voyage AI: cutting-edge embeddings and rerankers for search and RAG

Leverage Zilliz Serverless - Up to 50X Saving for Your Vector Storage Cost

Training state-of-the-art general text embedding

Fact vs. Fiction: Autodetecting Hallucinations in LLMs

VectorDB Schema Design 101 - Considerations for Building a Scalable and Perfo...

Voyage AI Embedding Models for Retrieval Augmented Generation

Chat with your data, privately and locally

Introducing Milvus and new features in 2.4 release

Recently uploaded

Design and Development of a Provenance Capture Platform for Data Science

Paolo Missier

Event-Driven Architecture Masterclass: Challenges in Stream Processing

ScyllaDB

Event-Driven Architecture Masterclass: Engineering a Robust, High-performance...

ScyllaDB

Vector Search @ sw2con for slideshare.pptx

jbellis

Webinar Recording: https://www.panagenda.com/webinars/alles-neu-macht-der-mai-wir-durchleuchten-den-verbesserten-notes-eigenschaftendialog/ Haben Sie sich schon einmal über den zu kleinen Eigenschaftendialog in Notes geärgert? Mussten Sie einen Agenten oder eine Aktion erstellen, um schnell mal ein Feld zu ändern? Haben Sie jedes mal endlos nach dem zu vergleichenden Feld gesucht, nachdem Sie ein neues Dokument ausgewählt haben? Wollten Sie das verdammte Ding einfach nur größer machen? Zum Glück gibt es dafür eine Lösung – und sie ist wahrscheinlich bereits installiert! Mit dem kostenlosen panagenda Document Properties (Pro) erhalten Sie den Eigenschaftendialog, den Sie schon immer haben wollten. Größer, anpassbar, und im Volltext durchsuchbar. Sehen Sie mehrere Dokumente gleichzeitig oder vergleichen Sie mit einem Diff-Viewer. Ändern Sie beliebige Felder und haben Sie endlich eine einfache Möglichkeit, Profildokumente für alle Benutzer zu verwalten. Entdecken Sie mit HCL Ambassador Marc Thomas, wie Document Properties Ihre Arbeit vereinfachen und Sie bei der täglichen Verwendung von Domino-Anwendungen unterstützen kann – im Client oder im Designer. Sie werden es nicht bereuen! Für Sie in diesem Webinar - Was Document Properties ist, welche Editionen es gibt und wo es in Notes und Domino Designer zu finden ist - Wie Sie nach einem beliebigen Feld suchen und es bearbeiten, Dokumente vergleichen oder alle Daten per CSV exportieren können - Suchen, Bearbeiten und auch Löschen von Profildokumenten - Welche Konfigurationseinstellungen verfügbar sind, um Funktionen anzupassen - Wie Ihre Endbenutzer davon profitieren - Sehen Sie alles in einer Live-Demo

Easier, Faster, and More Powerful – Alles Neu macht der Mai -Wir durchleuchte...

panagenda

Working together SRE & Platform Engineering

Marcus Vechiato

Explore the latest trends and insights on JavaScript usage with Pixlogix's informative blog. Discover key statistics and facts about JavaScript's role in web development, its popularity among developers, and its impact on modern websites. Stay updated with the evolving landscape of JavaScript frameworks and libraries, and learn how they're shaping the future of web development. Gain valuable insights to enhance your JavaScript skills and stay ahead in the digital realm.

JavaScript Usage Statistics 2024 - The Ultimate Guide

Pixlogix Infotech

In the ever-evolving landscape of data management, Zero-ETL is an approach that is reshaping how businesses handle and integrate their data. This webinar explores Zero-ETL, a paradigm shift from the traditional Extract, Transform, Load (ETL) process, offering a more streamlined, efficient, and real-time data integration method. We will begin with an introduction to the concept of Zero-ETL, including how it allows direct access to data in its native environment and real-time data transformation, providing up-to-date information with significantly reduced data redundancy. Next, we'll take you through several demonstrations showing how Zero-ETL can deliver real-time data and enable the free movement of data between systems. We will also discuss the various tools that support all aspects of Zero-ETL, providing attendees with an understanding of how they can adopt this innovative approach in their organizations. Lastly, the session will conclude with an interactive Q&A segment, allowing participants to gain deeper insights into how Zero-ETL can be tailored to their specific business needs and how they can get started today. Join us to discover how Zero-ETL can elevate your organization's data strategy.

The Zero-ETL Approach: Enhancing Data Agility and Insight

Safe Software

Ivanti’s Patch Tuesday breakdown goes beyond patching your applications and brings you the intelligence and guidance needed to prioritize where to focus your attention first. Catch early analysis on our Ivanti blog, then join industry expert Chris Goettl for the Patch Tuesday Webinar Event. There we’ll do a deep dive into each of the bulletins and give guidance on the risks associated with the newly-identified vulnerabilities.

2024 May Patch Tuesday

Ivanti

Oauth 2.0 Introduction and Flows with MuleSoft

shyamraj55

WebAssembly is Key to Better LLM Performance

Samy Fodil

Six Myths about Ontologies: The Basics of Formal Ontology

johnbeverley2021

AI mind or machine power point presentation

yogeshlabana357357

In the dynamic field of DevOps, the quest for efficiency and productivity is endless. This talk introduces a revolutionary toolkit: Large Language Models (LLMs), including ChatGPT, Gemini, and Claude, extending far beyond traditional coding assistance. We'll explore how LLMs can automate not just code generation, but also transform day-to-day operations such as crafting compelling cover letters for TPS reports, streamlining client communications, and architecting innovative DevOps solutions. Attendees will learn effective prompting strategies and examine real-life use cases, demonstrating LLMs' potential to redefine productivity in the DevOps landscape. Join us to discover how to harness the power of LLMs for a comprehensive productivity boost across your DevOps activities.

ChatGPT and Beyond - Elevating DevOps Productivity

VictorSzoltysek

Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptx

FIDO Alliance

Discover the top CodeIgniter development companies that can elevate your project to new heights. Our blog explores the best firms known for their expertise in CodeIgniter framework development. From robust web applications to scalable solutions, these companies deliver excellence. Whether you're a startup or an enterprise, find the perfect match for your development needs on Top CSS Gallery's blog.

Top 10 CodeIgniter Development Companies

TopCSSGallery

Introduction to FIDO Authentication and Passkeys.pptx

FIDO Alliance

Google I/O Extended 2024 Warsaw

GDSC PJATK

How to Check GPS Location with a Live Tracker in Pakistan

danishmna97

Continuing Bonds Through AI: A Hermeneutic Reflection on Thanabots

Leah Henrickson

Recently uploaded (20)

Design and Development of a Provenance Capture Platform for Data Science

Event-Driven Architecture Masterclass: Challenges in Stream Processing

Event-Driven Architecture Masterclass: Engineering a Robust, High-performance...

Vector Search @ sw2con for slideshare.pptx

Easier, Faster, and More Powerful – Alles Neu macht der Mai -Wir durchleuchte...

Working together SRE & Platform Engineering

JavaScript Usage Statistics 2024 - The Ultimate Guide

The Zero-ETL Approach: Enhancing Data Agility and Insight

2024 May Patch Tuesday

Oauth 2.0 Introduction and Flows with MuleSoft

WebAssembly is Key to Better LLM Performance

Six Myths about Ontologies: The Basics of Formal Ontology

AI mind or machine power point presentation

ChatGPT and Beyond - Elevating DevOps Productivity

Harnessing Passkeys in the Battle Against AI-Powered Cyber Threats.pptx

Top 10 CodeIgniter Development Companies

Introduction to FIDO Authentication and Passkeys.pptx

Google I/O Extended 2024 Warsaw

How to Check GPS Location with a Live Tracker in Pakistan

Continuing Bonds Through AI: A Hermeneutic Reflection on Thanabots