RoMath
收藏资源简介:
RoMath是一个罗马尼亚数学推理基准套件,由三个数据集组成:RoMath-Baccalaureate、RoMath-Competitions和RoMath-Synthetic。这些数据集涵盖了从高中到竞赛级别的多种数学领域和难度级别,旨在提升非英语语言模型的性能并促进多语言AI的发展。数据集通过半自动工作流程收集和整理,使用基础LLM从非结构化的OCR输入中提供结构化输出,并标注相关元数据。RoMath的应用领域包括自动证明生成、验证和从非正式的草图级证明到正式语言的翻译,旨在解决当前AI模型在低资源语言上的局限性。
RoMath is a Romanian mathematical reasoning benchmark suite composed of three datasets: RoMath-Baccalaureate, RoMath-Competitions, and RoMath-Synthetic. These datasets cover a broad spectrum of mathematical domains and difficulty levels ranging from high school to competition-grade levels, aiming to improve the performance of non-English language models and advance the development of multilingual AI. The datasets are collected and curated through a semi-automatic workflow, where base LLMs are utilized to generate structured outputs from unstructured OCR inputs, with relevant metadata annotated accordingly. Application scenarios of RoMath include automated theorem proof generation, verification, and translation from informal sketch-level proofs to formal languages, targeting the current limitations of AI models in low-resource languages.
RoMath: A Mathematical Reasoning Benchmark in Romanian
概述
RoMath是一个罗马尼亚语数学推理基准套件,包含三个数据集:RoMath-Synthetic、RoMath-Baccalaureate和RoMath-Competitions。这些数据集涵盖了不同数学领域和难度级别,旨在改进非英语语言模型并促进多语言AI的发展。
数据集组成
- RoMath-Synthetic: 合成数据集。
- RoMath-Baccalaureate: 针对罗马尼亚高中毕业考试的数据集。
- RoMath-Competitions: 针对罗马尼亚数学竞赛的数据集。
实验与使用
实验代码位于experiments/目录中,每个实验都有单独的脚本。使用./do_sbatch.sh <script.sh> <n_gpus>运行特定bash脚本。
模型训练与预测
- 微调模型:
python fine_tune.py --model <hf_model_name> --dataset [bac|comps|synthetic] --output checkpoints/ - 模型预测:
python predict.py --model <hf_model_name> --dataset [bac|comps|synthetic] --temperature 0.5 --k 3 --shots 5 --output predictions/ - 模型评估:
python evaluate.py --pred_file results/Qwen-Qwen2-1.5B-Instruct_bac_2_0.5.csv --judge_model <hf_model_name> --output results/ - 计算评估指标:
python evaluate/compute_metrics.py --input_dir results/ --output_dir metrics/
翻译与Judge数据集构建
- 翻译: 使用
translate.py和predict_translated.py脚本。 - 构建Judge数据集: 运行
evaluate/make_judge_dataset.py和evaluate_judge.py脚本。
引用
@misc{cosma2024romath, title={RoMath: A Mathematical Reasoning Benchmark in Romanian}, author={Adrian Cosma and Ana-Maria Bucur and Emilian Radoi}, year={2024}, eprint={2409.11074}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2409.11074}, }
许可证




