Train OpenSource Large Language Models from Zero to Hero

How to train Open Source LLMs with LoRA QLoRA, DPO and ORPO.
What you'll learn:
What is language model and how the training pipeline looks like
Fine tuning LLMs with supervised fine-tune (LoRA, QLoRA, DoRA)
Align LLMs to human preference using DPO, KTO and ORPO
Accelerate LLM training with multiple GPUs training and Unsloth library
Requirements:
No prior knowledge is required
Description:
Unlock the full potential of Large Language Models (LLMs) with this comprehensive course designed for developers and data scientists eager to master advanced training and optimization techniques.I'll cover everything from A to Z, helping developers understand how LLMs works and data scientists learn simple and advance training techniques. Starting with the fundamentals of language models and the transformative power of the Transformer architecture, you'll set up your development environment and train your first model from scratch.Dive deep into cutting-edge fine-tuning methods like LoRA, QLoRA, and DoRA to enhance model performance efficiently. Learn how to improve LLM robustness against noisy data using techniques like Flash Attention and NEFTune, and gain practical experience through hands-on coding sessions.The course also explores aligning LLMs to human preferences using advanced methods such as Direct Preference Optimization (DPO), KTO, and ORPO. You'll implement these techniques to ensure your models not only perform well but also align with user expectations and ethical standards.Finally, accelerate your LLM training with multi-GPU setups, model parallelism, Fully Sharded Data Parallel (FSDP) training, and the Unsloth framework to boost speed and reduce VRAM usage. By the end of this course, you'll have a good understanding and practical experience to train, fine-tune, and optimize robust open-source LLMs.
Who this course is for:
Developers, Data scientists, AI enthusiasts
Homepage

https://www.udemy.com/course/train-opensource-large-language-models-from-zero-to-hero/