LegalBench benchmark tasks
A collaboratively built suite of 160+ legal tasks that measure legal reasoning across statutes, cases, contracts, and procedures—crafted by legal experts and useful to test the performance of models and solutions.
Description
LegalBench is an open benchmark suite spanning six categories of legal reasoning (e.g., rule application, case comparison, statutory interpretation), with tasks and datasets contributed by lawyers and researchers.
It covers multiple task formats (classification, extraction, generation) and document types (opinions, statutes, contracts), making it a practical “starter kit” for comparing models on law-like skills before you run local legal evaluations (e.g., eviction answers, fee-waiver motions, RA letters).
It includes some justice-related tasks like performance at spotting people's legal issues from their short 1-3 paragraph descriptions, through the Learned Hands labeling platform from Stanford and Suffolk LIT Lab.
Use LegalBench to shortlist models, then layer your own actionability/guardrails/language-parity rubrics and jurisdictional materials.