Fine-tuning

Fine-tuning means taking a model someone else has already trained and training it further on a smaller set of your own examples. The model keeps its general knowledge from Pre-training and learns the labels, tone or domain vocabulary you care about. It sits between using a model off the shelf and training your own from scratch.

How it works

MLforSEO’s Introduction to Machine Learning for SEO course frames the choice as three options:

• Self-trained: you build a model from scratch with your own data. Best for niche domains, but it needs lots of data, time and skills.
• Pre-trained: a large organisation trains a model on a huge dataset and releases it ready to use. Great for general tasks, weaker in specialised or fast-changing domains such as medical research or cryptocurrency.
• Fine-tuned: a pre-trained model that you retrain on your own data so it performs the same task better in your domain.

In practice, fine-tuning is a form of Supervised Learning. You prepare a few hundred to a few thousand labelled examples (see Data Labelling (Annotation)), split them into training and test sets, and update the model’s weights for a few passes over the data. With Hugging Face Transformers this is a short script; LLM providers such as OpenAI offer fine-tuning as a managed service where you upload example conversations.

How it’s used in SEO and marketing

• Search intent classification tuned to your industry’s vocabulary, where “demo” or “API key” signals a transactional query in SaaS.
• Sentiment or topic classifiers for reviews and support tickets in a specific sector.
• Page-type or product-category classifiers for large sites.
• Teaching an LLM a consistent output format or brand style for repeated generation tasks.

Example

In the Introduction to Machine Learning for SEO intent classification practical, a pre-trained zero-shot transformer labels keywords without any examples, but sometimes misses semantic nuance. The suggested next step is to fine-tune a model such as BERT or DistilBERT on a small sample of your own keywords that are already labelled the way you want, which usually improves accuracy on your data.

Limitations and things to watch

• You need good, consistent labels. A fine-tuned model learns your labelling mistakes too, and can overfit a small dataset.
• Fine-tuning teaches behaviour and format more reliably than it teaches new facts. For up-to-date facts, retrieval (RAG) is usually the better tool.
• Always compare against a cheaper baseline, such as Zero-shot Learning or a few-shot prompt, on the same test set before paying for training.
• Managed tools change: Google Cloud AutoML, mentioned in older course material, now lives inside Vertex AI.

Learn it on MLforSEO

• Introduction to Machine Learning for SEO
• Sentiment Analysis: A Deep Dive Into the Theory, Methods, and Applications
• Search-Intent Classification with Machine Learning: The Approaches Explained
• Transformer-Based Search Intent Classification with BART Zero-Shot & SERP Signals (Colab Notebook)