Agenda — Fine-Tuning¶

Session length: 3 hours | Difficulty: Advanced | Prerequisites: Local LLMs, Python dataclasses, basic PyTorch tensor operations

What you will build today¶

A LoRA fine-tuning pipeline for a classification task using PEFT and a Hugging Face base model, compared against a prompt-only baseline.

Schedule¶

Time	Topic	File
0:00–0:20	Overview: full fine-tuning vs PEFT vs RAG vs prompting	01-fine-tuning-overview
0:20–0:55	LoRA and QLoRA: math, parameters, practical configs	02-lora-and-qlora
0:55–1:25	PEFT library: `get_peft_model`, training loop	03-peft
1:25–1:45	Training data: formats, quality, quantity rules	04-training-data
1:45–2:00	Decision framework: when to fine-tune	05-when-to-fine-tune
2:00–2:40	Practice exercises	06-practice-exercises
2:40–3:00	Interview questions review	07-interview-questions

Setup¶

pip install transformers peft trl datasets accelerate bitsandbytes torch

GPU required for full training

QLoRA fine-tuning requires a CUDA GPU with at least 12 GB VRAM for a 7B model. For CPU-only environments, use the smallest models (GPT-2, DistilBERT) or run in Google Colab with a T4 GPU.

← Day 1 Part 2 | Fine-Tuning Overview →