AlpacaEval 2.0

A widely used instruction-following evaluation benchmark containing 805 diverse user instructions. It employs an LLM as an automatic judge to compute win rates and introduces a length-controlled (LC) win rate to mitigate length bias.