Benchmark
BIG-bench
Probe language models on more than 200 crowd-contributed tasks.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
204 tasks contributed by more than 450 authors, spanning linguistics, math, commonsense, social bias, and more.
Methodology
Each task defines its own metric. BIG-Bench Lite and BIG-Bench Hard are commonly used subsets.