Model weights
Model weights are numerical values in a machine-learning model that help determine how inputs become outputs. Training changes these parameters according to an objective and data. A deployed model uses its learned parameters during inference, usually alongside other components such as code, configuration, and tokenization.
How it works
During training, an optimization procedure adjusts parameters to improve the model’s objective on examples. Fine-tuning can modify all or some parameters or train additional parameter sets. The resulting artifact differs from a collection of source records, but that distinction does not mean training data can never influence or be reproduced in outputs.
Research on deduplication and memorization shows why dataset handling and model behavior both matter when assessing exposure.
Why it matters for licensing
A data agreement may need to distinguish access to source data, creation of model artifacts, distribution of weights, and retention after a license ends. These are contractual questions to address explicitly; the technical meaning of “weights” does not decide the rights.
Example
Fictional example: A team fine-tunes a model using a permitted dataset and produces a new parameter file. Its agreement separately addresses the source files and the resulting model artifact, including what may be retained or distributed.
Limitations and misconceptions
Weights are not a simple database of training rows, but it is also unsafe to assume they contain no recoverable information about examples. Deleting source files does not automatically reverse training. Legal treatment and technical risk require separate assessment.
Questions to ask
- Does the agreement address resulting models and distribution of parameter files?
- What memorization or disclosure testing is appropriate?
- Which artifacts may remain after source-data access ends?