AIRC (Artificial Intelligence in Real Classroom)
收藏资源简介:
AIRC数据集是一个支持基于视频的教育问题生成(EQG)的数据集,由圣母大学计算机科学系的研究团队创建。该数据集包含两个大学课程的真实课堂讲座录音和屏幕视频录像,以及教育者创建的时间戳相关多项选择题。AIRC数据集旨在解决现有EQG数据集无法反映真实课堂环境的问题,它包含了现实课堂中讲座语音的不规范性和不结构性,以及与教科书相比,信息传递的不简洁性和不结构性。该数据集的创建过程包括收集真实的课堂讲座录音、屏幕视频录像和转录文本,然后由教育者创建时间戳相关的问题。AIRC数据集的应用领域是教育问题生成技术,旨在解决自动生成高质量教育问题的挑战。
The AIRC dataset is a dataset supporting video-based educational question generation (EQG), developed by a research team from the Department of Computer Science at the University of Notre Dame. This dataset includes real classroom lecture audio recordings, screen video recordings from two university courses, as well as timestamp-aligned multiple-choice questions created by educators. The AIRC dataset aims to address the shortcoming that existing EQG datasets fail to reflect real-world classroom environments, as it captures the non-standardized and unstructured characteristics of lecture speech in actual classrooms, as well as the lack of conciseness and structural coherence in information delivery compared to textbooks. The creation process of the AIRC dataset involves collecting real classroom lecture audio recordings, screen video recordings and transcribed texts, followed by the development of timestamp-aligned questions by educators. The AIRC dataset is targeted for applications in educational question generation technology, with the goal of solving the challenge of automatically generating high-quality educational questions.
数据集概述:Question Generation for Real Classrooms
数据集内容
- AIRC数据集:包含教育工作者为真实课堂讲座创建的多项选择题。
- 处理流程:支持基于视频的教育问题生成(QG)的上下文选择和重写流程。
数据集更新
- [2025.04.27]:上传了arXiv论文。
- [2025.04.15]:上传了两门课程的测验生成数据:
LLM-Frontier和DL-Intro。
主要功能
- 主脚本:
run_chatgpt_pipeline.py,通过Bash脚本执行,支持多种上下文和重写策略的组合。 - 输出保存:按配置保存输出,便于分析和比较不同策略的效果。
上下文格式支持
| 格式 | 描述 |
|---|---|
CoTT, DirectT |
基于文本的上下文,Chain-of-Thought (CoT) 或 Direct |
CoTV, DirectV |
视觉增强的上下文,带有CoT或Direct |
CoTMM, DirectMM |
结合文本和视觉的多模态输入 |
Full |
未经修剪或重写的完整原始上下文 |
RuleT3, RuleV3 |
基于规则的上下文格式,V和T变体 |
重写选项
Yes:应用重写策略以增强输入上下文的清晰度和简洁性。No:使用原始提取的上下文。
目录结构
bash . ├── run_pipeline.sh # 运行所有配置组合的Bash脚本 ├── run_chatgpt_pipeline.py # 主流程脚本 ├── ../data/ # 数据目录(在仓库外) │ ├── LLM-Frontier/ │ └── MIT-DL/ ├── ../out/ # 输出目录(自动生成)
输出路径
../out/<out_folder>/<model_name>/<data_split>/<context>_<rewrite>/<run_id>/
配置参数
model_name=gpt-4o-mini# 或其他HuggingFace / OpenAI模型名称max_context_length=4000max_output_length=800temperature=0.1seed=42run_id=0415-1out_folder=ND-LLMdata_split=ND-LLMdevice=cpu# 或cuda(如果可用)api_key=your_api_key# OpenAI/Together APIs所需
依赖项
- Python 3.10+
transformersopenaitogethertqdm
安装依赖
bash conda create -n coser python=3.10 conda activate coser pip install -r requirements.txt
运行流程
bash cd code/ bash run_COSER_pipeline.sh




