Wiki · AI & Machine Learning · concept
Fine-tuning
Taking a model that already learned a lot and nudging it to master one specific task.
Fine-tuning is how you specialize a model without starting over. You begin with a model that has already been trained on a huge, general pile of data — a foundation model — and give it a little more training on your own, narrower task. The model keeps everything it already learned about the shape of the problem and just adjusts to your specific question. It’s the difference between educating someone from birth and hiring an expert who only needs to learn your particular job.
How it works
You start from the foundation model’s existing weights instead of random ones, then continue training on task-specific data. There’s a spectrum of how much you touch. Full fine-tuning updates every weight in the model — most powerful, but heavy and easy to overfit if your data is small. Parameter-efficient methods like LoRA freeze the big model and train a tiny set of extra weights bolted on top — far cheaper, and you can keep many task-specific adapters for one base model. The lightest option isn’t really fine-tuning at all: freeze the model entirely, pull out its embeddings, and train a small classifier on those. Less data and compute generally push you toward the lighter end.
Why it matters (for bio × AI)
This is the reuse that makes bio-AI practical. A protein language model trained on millions of raw sequences knows nothing about your assay — but fine-tune it on a few thousand measured examples and it becomes a binding-affinity predictor. The same base model becomes a cell-type classifier, a stability estimator, or a localization predictor depending on what you fine-tune it toward. Nobody has the data to train these from scratch; almost everybody has enough to adapt one.
Related
See foundation model for the pretrained starting point, large language model for the text-native models most often fine-tuned, and embeddings for the frozen-feature shortcut.