Model distillation
Training a smaller model to reproduce the outputs of a larger one. When done without authorization, it becomes unauthorized distillation.
Definition
Distillation itself is a standard machine learning technique for model compression.
The misuse occurs when actors scrape a model's outputs to train their own model without the model owner's permission.
{esc(t(CURRENT_LANG, "not_confused_with"))}
- Not stealing model weights directly.
- Not all distillation is misuse — only unauthorized distillation.