3LM Benchmark
收藏资源简介:
3LM (علم) 是第一个专注于科学推理和编程的阿拉伯语原生基准测试套件。它包含三个子基准测试:1. Native STEM:来自真实阿拉伯教育材料的865道多项选择题(MCQ);2. Synthetic STEM:使用YourBench管道从阿拉伯教育文本生成的1,744道高难度MCQ;3. Arabic Code Benchmarks:通过GPT-4o翻译成阿拉伯语并经过反向翻译和人工验证的HumanEval和MBPP数据集。
3LM (Science) is the first native Arabic benchmark suite focused on scientific reasoning and programming. It encompasses three sub-benchmarks: 1. Native STEM, which includes 865 multiple-choice questions (MCQs) sourced from authentic Arabic educational materials; 2. Synthetic STEM, comprising 1,744 challenging MCQs generated from Arabic educational texts using the YourBench pipeline; 3. Arabic Code Benchmarks, which are HumanEval and MBPP datasets translated into Arabic by GPT-4o, followed by reverse translation and human validation.
3LM Benchmark 数据集概述
数据集简介
- 名称:3LM Benchmark
- 语言:阿拉伯语
- 领域:科学推理与编程
- 目标:评估阿拉伯语大语言模型在结构化推理和正式知识领域的表现
子数据集组成
-
Native STEM
- 类型:865道多项选择题(MCQs)
- 来源:阿拉伯语教育材料(生物、物理、化学、数学、地理)
-
Synthetic STEM
- 类型:1,744道高难度多项选择题
- 生成方式:使用YourBench流水线从阿拉伯语教育文本生成
-
Arabic Code Benchmarks
- 内容:HumanEval和MBPP数据集的阿拉伯语翻译版本
- 翻译方法:通过GPT-4o翻译,并经过反向翻译和人工审核验证
关键结果
- 评估模型数量:超过40个LLMs(包括阿拉伯语中心、多语言和双语模型)
- 性能亮点:
- Gemma-3-27B在STEM完成基准测试中表现最佳
- Qwen2.5-72B在所有领域的MCQ评估中表现优异
- 阿拉伯语代码生成性能与英语代码生成性能高度相关(r≈0.97)
使用方法
-
环境配置: bash git clone https://github.com/tiiuae/3LM-benchmark.git conda create -n 3lm_eval python==3.11 conda activate 3lm_eval pip install -e frameworks/lighteval pip install -e frameworks/evalplus-arabic
-
评估脚本: bash python launch_eval.py -c examples/lighteval_3lm.yaml python launch_eval.py -c examples/lighteval_native.yaml python launch_eval.py -c examples/lighteval_synthetic.yaml python launch_eval.py -c examples/evalplus_arabic_code.yaml
数据集获取
-
HuggingFace数据集:
- SyntheticQA: https://huggingface.co/datasets/tiiuae/SyntheticQA
- NativeQA: https://huggingface.co/datasets/tiiuae/NativeQA
- NativeQA-RDP: https://huggingface.co/datasets/tiiuae/NativeQA-RDP
- Evalplus-Arabic: https://huggingface.co/datasets/tiiuae/evalplus-arabic
-
代码数据集路径:
frameworks/evalplus-arabic/evalplus/data/data_files/
许可与引用
- 许可证:Falcon LLM Licence
- 引用格式: bibtex @article{boussaha2025threeLM, title={3LM: Bridging Arabic, STEM, and Code through Benchmarking}, author={Boussaha, Basma El Amel and AlQadi, Leen and Farooq, Mugariya and Alsuwaidi, Shaikha and Campesan, Giulia and Alzubaidi, Ahmed and Alyafeai, Mohammed and Hacid, Hakim}, journal={arXiv preprint arXiv:2507.15850}, year={2025} }




