Benchmark
AlpacaEval
Estimate instruction-following quality with an automatic judge.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
805 instructions; AlpacaEval 2.0 compares each answer with a GPT-4 Turbo reference answer.
Methodology
A model judge picks the better answer; report win rate and length-controlled win rate.