ArabLegalEval: A Multitask Benchmark for Assessing Arabic Legal Knowledge in Large Language Models

Faris Hijazi (1), Somayah AlHarbi (1), Abdulaziz AlHussein (1), Harethah Abu Shairah (2), Reem AlZahrani (2), Hebah AlShamlan (1), Omar Knio (2), George Turkiyyah (2) ((1) THIQAH, (2) KAUST)

Introduction

ArabLegalEval is a benchmark for LLMs to evaluate their ability to reason in the legal domain in the Arabic language.

Usage

1. Install packages

conda create -n ArabLegalEval python=3.12 -y
conda activate ArabLegalEval
pip install -r requirements.txt

2. Configure LLMs

Edit the llm_config.yaml, these are arguments to populate llama_index LLMs clients, it is pre-populated with examples.

3. Run tasks

Structure

The ArLegalEval dataset is composed of 3 main parent tasks. Evaluation code for each of these tasks can be found in its directory as follows:

• benchmarkQA/ contains the question-answering task

  • task data path: ./benchmarkQA/Najiz_QA_with_context_v2.benchmark.json

  cd benchmarkQA
  python run_benchmark.py
  

• benchmarkArLegalBench/ contains the translated subset of LegalBench

  cd benchmarkArLegalBench
  python run_benchmark.py
  

  • task data path: ./benchmarkArLegalBench/tasks/<task_name>/<split>/*.json

• benchmarkMCQs/ contains the multiple-choice questions task

  • task data path: ./benchmarkMCQs/All_Generated_MCQs_Evaluated_Shuffled.csv

  cd benchmarkMCQs
  python run_benchmark.py
  

---

• the raw data which is scraped from public sources is in data

Tested with Python 3.12

Citing this work

Please include all citations below, which credit all sources ArabLegalEval draws on.

@inproceedings{
      anonymous2024arablegaleval,
      title={ArabLegalEval: A Multitask Benchmark for Assessing Arabic Legal Knowledge in Large Language Models},
      author={Anonymous},
      booktitle={The Second Arabic Natural Language Processing Conference},
      year={2024},
      url={https://openreview.net/forum?id=3EHYXqKKLA}
}