Task organization

LegalBench tasks are organized into six categories based on the type of legal reasoning the task requires. These are:

  • Issue-spotting: tasks in which an LLM must determine if a set of facts raise a particular set of legal questions, implicate an area of the law, or are relevant to a specific party.
  • Rule-recall: tasks which require the LLM to generate the correct legal rule on an issue in a jurisdiction (e.g., the rule for hearsay in US federal court), or answer a question about what the law in a jurisdiction does/does not permit.
  • Rule-application: tasks which evaluate whether an LLM can explain reasoning in a manner which exhibits the correct legal inferences.
  • Rule-conclusion: tasks which require an LLM to determine the legal outcome of a set of facts under a specified rule.
  • Interpretation: tasks which require the LLM to parse and understand a legal text (e.g., classifying contractual clauses).
  • Rhetorical-understanding: tasks which require an LLM to reason about legal argumentation and analysis (e.g., identifying textualist arguments).

A richer description of tasks can be found in the most recent paper. The purpose of organizing tasks is to enable higher-level observations about the types of tasks which different LLMs can perform, using a vocabulary familiar to lawyers.

Train/test splits

Each LegalBench task has a “train split” which contains a handful of labeled samples. These samples can be used as in-context demonstrations in prompts. However, benchmark users are also welcome to resample the train/test splits in order to create larger train sets, or pursue other research goals.

Licenses

LegalBench tasks are subject to different licenses. License information can be found on task pages.

Prompts

Prompts for tasks are available on the Github repository.

Evaluation

Evaluation code for all non-open generation tasks can be found on the Github repository.

LegalBench contains several open-generation tasks. These include rule_qa and the rule-application tasks. For these tasks, we provide a manual gradebook to guide evaluation.


Table of contents