Training LLMs from Scratch

Interested in this course? Email us at customersupport@analyticsvidhya.com

Course curriculum

1

Introduction to the Course
- Course Overview
- Course Handouts
2

Getting Started with Training LLMs from Scratch
- Introduction to Training LLMs From Scratch
- What do you mean by Training LLMs from Scratch?
- Why train your own LLMs?
- Quiz
3

Important Concepts for Training LLMs from Scratch
- Chinchilla Scaling Laws
- Parallel and Distributed paradigm
- Data Parallelism
- Model Parallelism
- Pipeline Parallelism
- Fully Sharded Data Parellelism/ZeRO 3
- Tensor Parallelism
- 2D and 3D Parallelism
- Quiz
4

Steps involved in Training LLMs from Scratch
- Overview of Steps Involved in Training LLMs from Scratch
- Training Data Curation
- Data preprocessing
- Tokenization
- Model Architecture
- Model Evaluation
- Quiz
5

Training your own LLMs from Scratch
- Estimate Cost of Training LLMs from Scratch
- Hands on - Setting up runpod and intro
- Hands on - Dataset curation and preprocessing
- Hands on - Training BPE Tokenizer
- Hands on - Training LLMs from Scratch using DeepSpeed ZeRO 3/FSDP Distributed Training on 8 GPUs
- Quiz
- Assignment
6

Aligining LLM with Human Preferences
- Recap of Instruction Following LLM
- Introduction to Alignment Methods
- Reinforcement Learning from Human Feedback (RLHF)
- Aligning LLMs to Human Preferences with RLHF
- DPO and IPO
- Preference tuning LLMs with DPO
- Quiz
7

What Next?
- Recent Advancements
- Share your Experience!

We use cookies essential for this site to function well. Please click to help us improve its usefulness with additional cookies. Learn about our use of cookies in our Privacy Policy.