LawBench

Fits Chinese-law question-answering or document-processing assistants. Borrow its three-level split (memorize, understand, apply) and its abstention-rate metric for answers that cannot be parsed.

What it asks the agent to do

• recall a statute article
• classify a case by charge or dispute type
• extract named entities or key facts from a judgment
• predict a sentence or penalty amount
• write a summary or answer a legal consultation

What to borrow from it

taxonomybaseline

Caveats

Chinese civil law only; 20 tasks of 500 examples each, no tools or multi-step work; model rankings date from 2023 (GPT-4 era) and the repo has not been refreshed; built from CAIL and JEC_QA public case data, so judgment text with party names appears.

Details

Links

• Paper
• Github