scone
DoLa: Decoding by Contrating Layers Improves Factuality in Large Language Models