Fine-tuning
Fine-tuning is an additional training process applied to a model that has already been trained. It uses task-relevant data and an objective to adjust behavior. The process can update all model parameters, only some of them, or added trainable components, depending on the method.
How it works
Define the desired behavior, prepare suitable examples, choose an adaptation method, and compare the result with a baseline on separate evaluation data. Training settings and example quality influence both improvements and regressions. The resulting model should be assessed for relevant tasks beyond the narrow training examples.
Fine-tuning data is the material used for adaptation. Keeping the process and dataset distinct helps clarify requirements, responsibilities, and permissions.
Why it matters for licensing
A recipient that wants to fine-tune needs permission covering the actual training activity and relevant data. Agreements may also need to address derived model artifacts, retention, and distribution. A dataset’s commercial usefulness depends on whether it supports the intended adaptation, not merely on its size.
Example
Fictional example: A business adapts a pre-trained classifier to its approved service categories using checked examples. It compares the adapted model with the original on independent cases and investigates categories where accuracy declined.
Limitations and misconceptions
Fine-tuning does not guarantee factual accuracy or improvement across all tasks. Poor examples can teach errors, and narrow adaptation can reduce other capabilities. Retrieval and prompting can serve different needs; the appropriate approach should be evaluated rather than assumed.
Questions to ask
- What behavior should change, and how will improvement be measured?
- Are the examples and rights suitable for this training process?
- Which regressions and disclosure risks will be tested?
Sources
Explore whether your business data could be a fit.
Start with a description of your systems—not a data upload.