Benchmark
Adversarial NLI (ANLI)
Evaluate natural-language inference on examples written to expose model weaknesses.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Dataset v1.0 contains three rounds of adversarially collected natural-language inference examples, with separate training, development, and test splits.
Methodology
People and models participate in iterative data collection. Models classify relationships between premises and hypotheses; the original leaderboard separates the three rounds.