3LM
收藏资源简介:
3LM数据集是一套针对阿拉伯语的STEM领域和代码生成而设计的评估基准,旨在填补阿拉伯语大型语言模型在科学和编程领域评估工具的空白。该数据集包含三个部分:STEM相关的问答对、合成的STEM问题和代码生成问题。数据集内容来源于阿拉伯语教科书、教育练习册和在线教育平台,并经过严格的机器翻译和人机协作验证流程,确保了数据的高质量和真实性。3LM数据集旨在支持阿拉伯语大型语言模型在这些重要但代表性不足的领域的研发,为评估模型在正式知识领域的性能提供了一个更加全面和稳健的框架。
The 3LM Dataset is an evaluation benchmark designed for Arabic STEM domains and code generation, which aims to fill the gap in evaluation tools for Arabic large language models (LLMs) in the fields of science and programming. This dataset comprises three components: STEM-related question-answer pairs, synthesized STEM questions, and code generation questions. The content of the dataset is sourced from Arabic textbooks, educational workbooks, and online education platforms, and has undergone rigorous machine translation and human-machine collaborative verification procedures to ensure its high quality and authenticity. The 3LM Dataset aims to support the research and development of Arabic large language models in these important yet underrepresented fields, and provides a more comprehensive and robust framework for evaluating model performance in formal knowledge domains.
3LM Benchmark 数据集概述
数据集简介
3LM (علم) 是首个专注于科学推理和编程的阿拉伯语原生基准测试套件,包含三个子基准:
- Native STEM:865道来自阿拉伯语教育材料(生物、物理、化学、数学和地理)的多项选择题(MCQs)。
- Synthetic STEM:1,744道高难度多项选择题,使用YourBench流水线从阿拉伯语教育文本生成。
- Arabic Code Benchmarks:通过GPT-4o翻译并经过反向翻译和人工验证的HumanEval和MBPP数据集阿拉伯语版本。
关键结果
- 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)。
- 指令调优模型在鲁棒性测试中表现优于基础模型。
使用方法
-
克隆仓库: bash git clone https://github.com/tiiuae/3LM-benchmark.git
-
设置环境: bash conda create -n 3lm_eval python==3.11 conda activate 3lm_eval pip install -e frameworks/lighteval pip install -e frameworks/evalplus-arabic
-
启动评估脚本: 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
数据集下载
- 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
humaneval-ar.jsonmbpp-ar.json
许可证
引用
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} }



