LLM Training & Fine-Tuning — Study Guide

An illustrated, exam-aligned study guide covering the full LLM lifecycle: pretraining objectives and data, scaling laws, distributed training and mixed precision, supervised fine-tuning, RLHF and DPO alignment, and parameter-efficient fine-tuning (LoRA/QLoRA). Diagrams, comparison tables, and worked examples throughout.

Key facts

• Course price: $195 (one-time payment)
• Field: Artificial Intelligence & LLMs
• Prepares for: LLM Training & Fine-Tuning
• 6 modules · 18 lessons · about 4 hours of lessons
• Self-paced, online, illustrated lessons

Course outline

1. Pretraining Objectives & Data — How base models learn: causal vs masked LM, next-token prediction, and the data curation that makes or breaks a corpus.
   • Causal vs Masked Language Modeling
   • Perplexity & the Loss You Optimize
   • Data Curation: Filter, Dedup, Decontaminate
2. Tokenizers & Scaling Laws — Building the vocabulary the model thinks in, and the Chinchilla math that tells you how big and how long to train.
   • Training a Tokenizer (BPE & friends)
   • Chinchilla: Compute, Data & Parameters
   • Reading the Loss Curve
3. Distributed Training at Scale — Splitting a model across many GPUs: data, tensor, and pipeline parallelism, plus ZeRO/FSDP, accumulation, and checkpointing.
   • Data, Tensor & Pipeline Parallelism
   • ZeRO / FSDP: Sharding the Footprint
   • Accumulation & Activation Checkpointing
4. Mixed Precision & Optimization — fp16 vs bf16, loss scaling, fp32 master weights, and the warmup/decay schedules that keep training stable.
   • fp16 vs bf16 & Loss Scaling
   • Why an fp32 Master Copy?
   • Warmup, Decay & AdamW
5. Fine-Tuning & Alignment — From base model to helpful assistant: SFT and chat templates, RLHF with reward models and PPO, and the simpler DPO.
   • The Alignment Pipeline & SFT
   • RLHF: Reward Models & PPO
   • DPO: Alignment Without the RL Loop
6. PEFT, Forgetting & Evaluation — Adapt cheaply with LoRA/QLoRA, avoid catastrophic forgetting and contamination, and evaluate fine-tuned models honestly.
   • LoRA & the Low-Rank Adapter
   • QLoRA & the PEFT Family
   • Forgetting, Contamination & Honest Evaluation