Fine-Tuning & Specialization
When to teach a model versus when to prompt it — and how to do both right
Session Topics
Pre-training vs Fine-tuning
The difference between what a model learned from the internet and what you can teach it about your specific domain.
When to Fine-tune
The decision framework: use RAG first, prompt-tune second, fine-tune only when style/behaviour cannot be achieved otherwise.
LoRA & QLoRA
Low-Rank Adaptation — how to fine-tune massive models efficiently without massive compute. The technique behind most open-source fine-tuned models.
Instruction Optimisation
RLHF, DPO, and instruction-tuning formats — how models learn to follow instructions, not just predict tokens.
Building Training Datasets
The most underestimated part of fine-tuning: data collection, cleaning, formatting, and quality evaluation.
Prepare Your Domain Training Dataset
Create a fine-tuning training dataset for your hiring domain. Format it correctly for instruction fine-tuning. Evaluate quality, estimate costs, and make the decision: is fine-tuning actually the right approach for your use case?
A formatted training dataset with quality evaluation and a written justification for or against fine-tuning in your use case.
Building a Vector Database
Learn how to create and populate a Vector Database — the backbone of any RAG system. This exercise walks you through chunking documents, generating embeddings, and storing them in a vector store so they can be retrieved semantically. Use the accompanying AI Textbook as your reference material throughout the exercise.
Key Takeaways
- Fine-tuning is overused — RAG + prompting solves most problems cheaper
- LoRA makes fine-tuning accessible; QLoRA makes it possible on consumer hardware
- Training data quality matters more than training data quantity
- Always benchmark fine-tuned models against the base model + a strong prompt