Non-STEM_TextBook_Arabic
收藏Zenodo2026-01-05 更新2026-05-26 收录
下载链接:
https://zenodo.org/doi/10.5281/zenodo.18152258
下载链接
链接失效反馈官方服务:
资源简介:
The full corpus is curated across multiple STEM disciplines and structured for use in LLM training, evaluation, and instruction tuning (SFT/RLHF). This sample represents the structure and quality of the larger dataset.
Dataset composition (full corpus):
Text corpus: 1.6B+ words of curated STEM and Non-STEM educational content across 22,000+ textbooks in 7 languages (English, Hindi, Arabic, Bahasa, Tamil, Telugu, Kannada)
Question–Answer pairs: 6.5M+ high-quality Q&A pairs of STEM and Non-STEM content in English, Arabic, Hindi, and Indic languages
Video data: 100K+ hours of STEM videos and 30K+ hours of UGC
Audio data: 821K+ hours of podcasts and dual-channel call center data
Medical datasets: 30M+ files including clinical and diagnostic data such as CT scans, MRI, X-ray, pathology, EHRs, USG reports, and echo reports
This repository includes:
A small preview subset of the Non-STEM Arabic TextBook data
Flat, viewer-friendly schema for inspection
Parquet files suitable for benchmarking and evaluation
Purpose of this dataset:
Dataset preview and validation
Model evaluation and experimentation
Schema and format inspection before full-scale access
Note: This repository contains sample data only. Access to the complete dataset is available separately under appropriate licensing or partnership terms.For detailed information and domain-specific requirements, please reach out to vipul.mishra@infobay.ai
本完整语料库覆盖多门科学、技术、工程与数学(Science, Technology, Engineering, Mathematics,简称STEM)学科,经精心遴选构建,适配大语言模型(Large Language Model,LLM)的训练、评估与指令微调,涵盖监督微调(Supervised Fine-Tuning,SFT)与基于人类反馈的强化学习(Reinforcement Learning with Human Feedback,RLHF)两类应用场景。本次发布的样本可体现完整数据集的结构与质量水准。
### 数据集构成(完整语料库)
1. **文本语料库**:收录7种语言(英语、印地语、阿拉伯语、Bahasa、泰米尔语、泰卢固语、卡纳达语)的2.2万余册教材中的160亿+词经过精选的STEM与非STEM教育内容。
2. **问答对数据集**:包含英语、阿拉伯语、印地语及印度语系语言中的650万+条高质量STEM与非STEM内容问答对。
3. **视频数据集**:包含10万+小时的STEM教学视频与3万+小时的用户生成内容(User Generated Content,UGC)。
4. **音频数据集**:包含82.1万+小时的播客内容与双通道呼叫中心语音数据。
5. **医疗数据集**:包含3000万+份文件,涵盖临床与诊断类数据,如CT扫描影像、核磁共振(MRI)影像、X射线影像、病理报告、电子健康档案(EHRs)、超声报告与超声心动图报告。
### 本仓库包含内容
1. 非STEM领域阿拉伯语教材数据的小型预览子集
2. 简洁直观、便于阅览的数据结构,支持数据检视
3. 适配基准测试与模型评估的Parquet格式文件
### 数据集用途
1. 数据集预览与验证
2. 模型评估与实验研究
3. 在获取完整数据集前,对其结构与格式进行检视
**注意**:本仓库仅包含样本数据。完整数据集需通过合规授权或合作协议单独获取。如需获取详细信息或对接特定领域需求,请联系邮箱vipul.mishra@infobay.ai
提供机构:
InfoBay AI创建时间:
2026-01-05



