#reinforcement finetuning

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.