#synthetic unanswerable math dataset

Researchers developed the SUM dataset to teach AI language models to say 'I don't know,' significantly reducing hallucinations and improving refusal rates without harming accuracy on answerable problems.