Nvidia Efficient Large Language Model (LLM) Customization (ELLMC)

Ziele der Schulung

Enterprises need to execute language-related tasks daily, such as text classification, content generation, sentiment analysis, and customer chat support, and they seek to do so in the most cost-effective way. Large language models can automate these tasks, and efficient LLM customization techniques can increase a model’s capabilities and reduce the size of models required for use in enterprise applications. In this course, you'll go beyond prompt engineering LLMs and learn a variety of techniques to efficiently customize pretrained LLMs for your specific use cases—without engaging in the computationally intensive and expensive process of pretraining your own model or fine-tuning a model's internal weights. Using NVIDIA NeMo™ service, you’ll learn various parameter-efficient fine-tuning methods to customize LLM behavior for your organization.

Please note that once a booking has been confirmed, it is non-refundable. This means that after you have confirmed your seat for an event, it cannot be cancelled and no refund will be issued, regardless of attendance.

By the time you complete this course you will be able to:

• Apply parameter-efficient fine-tuning techniques with limited data to accomplish tasks specific to your use cases
• Use LLMs to create synthetic data in the service of fine-tuning smaller LLMs to perform a desired task
• Drive down model size requirements through a virtuous cycle of combining synthetic data generation and model customization.
• Build a generative application composed of multiple customized models you generate data for and create throughout the workshop.

Zielgruppe Seminar

• This course is primarily intended for intermediate level and above Python developers with a solid understanding of LLM fundamentals and some prompt engineering experience.

Voraussetzungen

• Professional experience with the Python programming language.
• Familiarity with fundamental deep learning topics like model architecture, training and inference.
• Familiarity with a modern Python-based deep learning framework (PyTorch preferred).
• Familiarity working with out-of-the-box pretrained LLMs.

Seminarinhalt

Course Introduction

• Orient to the main workshop topics, schedule, and prerequisites.
• Learn about the main tools and technologies that will be utilized in the workshop.
• Discuss parameter-efficient fine-tuning (PEFT) in the context of other LLM customization techniques.
• Learn about the motivation for PEFT, and likely scenarios for when it is most useful.
• Learn about synthetic data generation in the context of PEFT.

Introduction to the Interactive Environment

• Get familiar with the interactive workshop environment including its NVIDIA Language NIM, and Nemo Framework container.
• Review the core LangChain programming patterns that will be utilized throughout the workshop.

LoRA

• Learn and explore the conceptual underpinnings behind LoRA, a key PEFT technique of the workshop.
• Learn the high-level process of training and performing inference with LoRA adapters using Nemo Framework and NVIDIA NIM.
• Perform LoRA fine-tuning for a text summarization task using an open-source dataset.
• Perform inference using LoRA adapters together with NVIDIA NIM.
• Evaluate the LoRA fine-tuned model's performance on summarization against the non-fine-tuned LLM.

Synthetic Data Generation

• Orient to the project use cases you will be tackling with a combination of SDG and PEFT.
• Perform your first synthetic data generation task using the Nemotron-4 340B model.
• Learn how to perform a wide variety of synthetic data generation tasks using Nemo Curator.
• Use Nemo Curator to generate seed synthetic data in preparation for a larger SDG task.
• Generate a diverse and high-quality dataset for use in PEFT.

Project: Automatic Email Categorizer

• Create a naive email categorizer without a fine-tuned model, and observe its performance.
• Prepare synthetic data for LoRA fine-tuning.
• Perform LoRA fine-tuning with synthetic data on the task of email categorization.
• Analyze the performance of the LoRA fine-tuned LLM on the task of email categorization.

Direct Performance Optimization (DPO)

• Learn the key concepts, use cases, and benefits of DPO, including in conjunction with LoRA.
• Create a preference-based dataset well-suited for DPO.
• Use Nemo Aligner to perform DPO using LoRA.
• Evaluate the performance of the DPO using LoRA fine-tuned model.

Final Review

• Review key learnings and answer questions.
• Earn a certificate of competency for the workshop.
• Complete the workshop survey.
• Get recommendations for the next steps to take in your learning journey.

Hinweise

Partner

Dieses Seminar bieten wir in Kooperation mit unserem Nvidia Learning Partner Fast Lane Institute for Knowledge Transfer GmbH an.

Open Badge für dieses Seminar - Ihr digitaler Kompetenznachweis

Durch die erfolgreiche Teilnahme an einem Kurs bei IT-Schulungen.com erhalten Sie zusätzlich zu Ihrem Teilnehmerzertifikat ein digitales Open Badge (Zertifikat) – Ihren modernen Nachweis für erworbene Kompetenzen.

Ihr Open Badge ist jederzeit in Ihrem persönlichen und kostenfreien Mein IT-Schulungen.com-Konto verfügbar. Mit wenigen Klicks können Sie diesen digitalen Nachweis in sozialen Netzwerken teilen, um Ihre Expertise sichtbar zu machen und Ihr berufliches Profil gezielt zu stärken.

Übersicht: NVIDIA Schulungen Portfolio

Mehr zu den Vorteilen von Badges