LegalBench is a benchmark designed to evaluate the legal reasoning abilities of large language models. It was built collaboratively by legal experts and researchers, led by Neel Guha of Stanford University, and released in 2023. It consists of more than 160 tasks covering several types of legal reasoning, including issue spotting, rule recall, rule application, rule conclusion, and interpretation.
The tasks are organized around a taxonomy of reasoning types used in legal practice and scholarship. The benchmark is widely used to compare model performance in the legal domain and to measure the effect of changes to models or prompts.