> Our research is motivated by the observation that previous instruction tuning methods force the model to complete a sentence no matter whether the model knows the knowledge or not.<p>And so... adding the random samples that include "I don't know" to the training datasets brings the accuracy on unknown questions from <10% to over 90%