What Is Fine-Tuning?
Also called fine tuning, finetuning
Short Answer
Fine-tuning is training an already trained model a little more on a smaller, focused dataset so it gets better at one task, style or format. It changes the model's weights. It suits teaching a behavior, such as a tone or a strict output format, more than adding new facts.
- Taught in lesson 82
Why It Matters
Fine-tuning lets a team turn a general model into a specialist for one job, such as a strict output format or a company's style, or make a small, cheap model perform like a large one on a narrow task. Knowing when it helps, and when prompts or retrieval are enough, saves projects time and money.
How It Works
For new or changing facts, retrieval-augmented generation is usually the better tool, because documents can be updated without retraining. Fine-tuning shines when prompting alone cannot get the behavior right, or when a small fine-tuned model can replace a large expensive one.
Methods such as LoRA train only a small set of extra weights, which makes fine-tuning possible on a single GPU. A good fine-tuning dataset, clean and representative, matters more than the training settings.
A fine-tuning project needs a set of high-quality examples of the behaviour you want, a held-out test set, a baseline to beat (often a well-written prompt) and a plan to deploy and update the tuned model. Methods such as LoRA train only a small number of added weights, which makes fine-tuning possible on modest hardware. Poor or inconsistent examples teach the model the wrong thing, so data quality matters more than quantity.
Fine-Tuning in Practice
Everyday Example
A support team fine-tunes a model on 2,000 of its best past replies so new answers follow the company's tone and structure.
Industry Example: Pharma
A pharmacovigilance team fine-tunes a small model to extract drug names, reactions and dates from adverse event reports in a fixed format for safety databases.
Where You Meet It
- Projects that need a consistent format or tone.
- Teams running open-weight models.
- Machine learning engineer job ads.
- Small models adapted to run cheaply on internal servers.
Key Ideas Around It
- Base model
- The pre-trained model you start from.
- LoRA
- A light method that trains a few added weights.
- Training examples
- Input and output pairs showing the target behaviour.
- Baseline
- The result to beat, often a good prompt.
A Common Mix-Up
Fine-tuning is not the best way to teach a model new facts that change. For knowledge, retrieval is easier to update; fine-tuning is for behaviour.
Check YourselfYour company policy changes every month. Fine-tuning or RAG to keep answers current?
RAG: you update the documents and answers change immediately, without retraining.
Learn It on ai2hired
S7 · #82How LLMs Are Made
Coming in Season 7, Deep Learning and Transformers. Every ai2hired lesson is free on YouTube.
Questions People Ask
Fine-tuning or RAG: which should I use?
Use RAG to give the model facts it must cite and keep current. Use fine-tuning to change how it behaves. Many systems combine both.
How much data does fine-tuning need?
Often hundreds to a few thousand good examples for a focused task. Quality and consistency matter more than volume.
Related Terms
Learn AI From Zero to Hired
Free lessons on YouTube, each one built from what real AI job ads ask for.