What's this about?
The pipeline rewrites public documents, generates questions that require those documents and filters out questions answerable from memory. The resulting data trains models to use supplied context.
What exactly is memorisation leakage?
If a model saw the source during pretraining, it may answer without reading the new context. Rewriting facts and testing dependency makes that shortcut harder.
Why does this matter?
20 HF upvotes. Roughly 10,000 synthetic samples lifted the reported CL-bench score from 13.7% to 24.6%, though independent replication is still needed.