Goldfish Loss modifies language-model training to reduce verbatim memorization. It randomly excludes selected tokens from the loss calculation, making it harder for the model to learn complete sequences exactly. section 3.4
The technique addresses a specific privacy concern, but does not resolve privacy protection on its own.
- Mechanism: selectively omitting tokens from next-token prediction weakens learning of complete sequences. section 3.4
- Limitation: it belongs to a broader set of difficult-to-scale mitigation strategies, while standardized privacy benchmarks and formal guidance remain lacking. section 3.4
