2026-10-14T17:04:35Z https://www.scidb.cn/oai

10.57760/sciencedb.43203 2026-07-28T14:38:33Z

2026-07-28 Edu-Eval: A Multidisciplinary Benchmark Dataset for K-12 Interactive Instructional Content Generation doi:10.57760/sciencedb.43203 en Edu Eval is a multidisciplinary benchmark dataset for K-12 interactive teaching content generation tasks, used to evaluate the ability of large language models to generate runnable, visual, and interactive HTML teaching content based on teaching objectives. The dataset contains a total of 226 task instructions, covering four subjects: mathematics, Chinese, English, and physics, including 78 in mathematics, 44 in Chinese, 52 in English, and 52 in physics; Covering three stages of primary school, junior high school, and high school, including 78 in primary school, 96 in junior high school, and 52 in high school. According to the statistics of "discipline stage knowledge point", the dataset contains 9 actual discipline stage combinations and 103 leaf categories; After removing duplicates from the names of knowledge points that are repeated across disciplines, a total of 90 categories of knowledge points are involved. The dataset is organized based on K-12 curriculum standards, teaching syllabus, and structured knowledge point information. The original educational metadata includes subject, stage, grade, first level directory, second level directory, and teaching content information, and is expanded into interactive teaching content generation instructions based on this. Instructions typically require the model to generate a single file interactive teaching page containing HTML, CSS, and JavaScript around specified knowledge points, in order to assess the model's comprehensive abilities in knowledge representation, page rendering, dynamic presentation, parameter control, and interactive feedback. The data is stored in JSON Lines (JSONL) format encoded in UTF-8, with each line corresponding to an independent record. Each record contains three main fields: status, original_data, and generatedprompt. Among them, status represents the instruction generation status; Original_data stores structured educational metadata, including 7 sub fields: "Subject", "Stage 1", "Stage 2", "Grade", "First level Catalog", "Second level Catalog", and "Content"; Generatedprompt is the instruction for generating interactive teaching content for model evaluation. Among the 226 records released this time, the status is all success, and there are no empty values in the main and nested fields. No completely duplicate generated_prompt was found. This dataset belongs to static text benchmark data and does not involve time series, geographic spatial resolution, or personal information. This public content only includes task instructions and corresponding educational metadata, and does not include HTML pages generated by the model, model rating results, or user data. Cognitive depth is used as a principle for data construction and task design, but it is not explicitly labeled item by item. The data can be read using the Python standard library, Pandas, or other software that supports the JSONL format.  Educational large language models; Edu-Eval; K-12; interactive teaching content Zhang Ruohua Liu Tao Lu Ye Song Siyu Liu Wentao Song Yu Zhou Aimin Hao Hao PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank