A publisher's licence permits ingesting their content i...

The proposal directly violates the publisher’s licensing constraint by passing forbidden text into gradient updates during fine-tuning. Grounding SFT using retrieved pairs derived from the licensed index embeds the underlying passage text directly into the model’s weight updates.

This creates severe compliance and memorization risks. Under memorization principles, data points that appear frequently during training are far more susceptible to verbatim extraction. In a retrieval-augmented fine-tuning pipeline, popular documents are retrieved repeatedly across multiple training queries, appearing many times in the SFT dataset. Consequently, the most frequently retrieved licensed documents become the most extractable from the fine-tuned weights, turning hyperparameter choices like epoch count into legal liability metrics.

To resolve this while preserving model grounding capability, separate the training signal from the copyrighted text. Fine-tune the generator for grounding compliance using synthetic, paraphrased, or permissively licensed passages. Keep the publisher’s licensed dataset strictly restricted to test-time retrieval, where it remains in the datastore and never touches model gradients. Finally, empirically validate compliance by probing the fine-tuned checkpoint against the licensed subset before deployment.