Would you serve a 1.3B model instead of a 7B to reduce...

While scaling laws indicate that smaller models (e.g., 1.3B vs 7B) generally retain less verbatim training data, shrinking model size is an inefficient lever for managing copyright exposure. Downgrading model capacity imposes a permanent performance penalty on every single query the system serves.

A far superior intervention is dataset deduplication prior to training. Thorough deduplication achieves up to a 10x reduction in training data extraction emissions while incurring zero loss in model perplexity or reasoning capability.

Additionally, reducing parameter size fails to fix the primary driver of memorization: heavily duplicated text in the pre-training set remains extractable even in much smaller models. If you are serving a pre-trained model where you cannot modify the training corpus, copyright risk management should be handled at the weights/datastore boundary (e.g., using retrieval filtering) rather than sacrificing general model capability.