Jargon:Fine-tuning
1 min read
Say this at the party
“Fine-tuning is extra training for an AI that already exists, to make it better at one job or talk in one particular way.”
In plain English
Building a big AI from scratch takes huge amounts of data and money. Fine-tuning skips that. You take a model that already knows a lot and give it a smaller round of extra training.
The extra training uses examples of exactly what you want, such as thousands of polite customer emails or medical notes. The model slowly picks up that style or skill.
Most chatbots you use were fine-tuned. A raw model just continues text; fine-tuning is what teaches it to act like a helpful assistant.
Where it came from
Reusing a trained model for a new task was common in image recognition first. In 2018, Jeremy Howard and Sebastian Ruder showed with ULMFiT that the same approach worked well for language, which helped make fine-tuning the normal way to build language tools.
See it in action
A law firm fine-tunes a model on its past contracts. Afterwards, the AI drafts new ones in the firm's house style and uses its usual clauses without being told each time.
What people get wrong
- Fine-tuning isn't the best way to teach an AI new facts. For up-to-date information, looking things up (RAG) usually works better and is easier to change.
- You often don't need it. A good set of instructions in the prompt can get you most of the way for a fraction of the effort.
Go deeper
- ArticleWikipediaFine-tuning (deep learning) - WikipediaSolid overview, including cheaper methods like LoRA.
- PaperHoward and Ruder, 2018Universal Language Model Fine-tuning for Text ClassificationULMFiT, the paper that showed fine-tuning works brilliantly for language.
- GuideAnthropicGlossary - Claude Platform DocsAnthropic's plain definition, and why Claude is already a fine-tuned assistant.
- GuideHugging FaceFine-tuning · Hugging FaceThe hands-on guide, if you want to try fine-tuning a model yourself.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.