Training & post-training
Fine-tuning
Additional training that adapts a model to a task, domain or behaviour.
Definition
Fine-tuning updates some or all model parameters using a targeted dataset after the model has completed broad pretraining.
Why it matters
It can improve domain performance but creates new model versions, validation duties and risks of overfitting or lost general capability.
Related concepts
- Supervised fine-tuning
Fine-tuning on curated examples of desired inputs and outputs.
- Low-rank adaptation
A parameter-efficient method for adapting a model without updating every weight.
- Catastrophic forgetting
Loss of previously learned capability when a model is trained on new data.