Fine-tuning continues training a pretrained model on your own labelled data so it specialises in your task.
When Fine-Tuning Helps
- Your domain or style differs from general data (legal documents, medical images).
- You need consistent output formats or behaviour at scale.
- A smaller fine-tuned model could replace a larger, costlier one.
Try These First
For language models, prompting with examples and retrieval-augmented generation often achieve the goal without training. For classification, feature extraction with a simple classifier on top of embeddings is quick and cheap.
The Process
- Collect high-quality labelled examples that reflect real inputs.
- Split into training, validation and test sets.
- Replace or adapt the output layer for your task.
- Train with a low learning rate so you don't destroy pretrained knowledge.
- Evaluate against the base model and simple baselines.
Parameter-Efficient Fine-Tuning
Methods such as LoRA train small additional matrices while freezing the original weights. They need far less memory and produce small adapter files, which makes fine-tuning large models practical.
Risks
- Overfitting small datasets.
- Catastrophic forgetting of general abilities.
- Inheriting and amplifying biases in your data.
- Licences: check the base model's licence permits your use.
Evaluate Properly
Test on held-out data that resembles production, include difficult cases, and compare cost and latency as well as accuracy.