Fine-tuning continues training a pretrained model on a curated set of examples, changing the model's parameters. It differs from prompting and retrieval, which leave the model unchanged and supply information at request time.
It requires labeled examples and produces a model variant that must be separately maintained.
Alternative Names:
Fine-Tuning, Model Adaptation
Why it Matters?
Fine-tuning is frequently proposed as a solution to accuracy problems that retrieval would address better, and the distinction matters for legal work. Case facts change constantly and are matter-specific, so they belong in retrieval rather than model weights. Fine-tuning suits stable behaviors such as output format and tone. Fine-tuning on client data also raises confidentiality questions that retrieval does not, since the information becomes embedded in the model.
Frequently Confused with
Related terms
Frequently asked questions
When is fine-tuning the right approach?
Does fine-tuning on client data raise confidentiality issues?
