Dataset Card for EduData

Dataset Summary

EduData is a large-scale Chinese educational question-answering dataset released with our AAAI 2026 paper, "From Diagnosis to Generalization: A Cognitive Approach to Data Selection for Educational LLMs".

The dataset is designed to support the training and evaluation of educational large language models, with a particular focus on data selection, cross-subject transfer, and generalization in exam-style reasoning settings.

According to the accompanying paper, EduData contains 98,000 high-school-level single-choice questions spanning seven subjects:

• Mathematics
• Physics
• Chemistry
• Biology
• History
• Geography
• Politics

The paper-organized version contains 14,000 questions per subject.

Supported Tasks

EduData is primarily intended for:

• Supervised fine-tuning of educational LLMs
• Multiple-choice question answering in Chinese
• Research on data selection for LLM training
• Cross-subject generalization and transfer learning
• Reproduction of the CASS framework experiments

Language

The dataset is in Chinese. Question text is sourced from Chinese mock examinations and college-entrance-exam-style educational materials.

Dataset Structure

Current Release Format

The current release provides a merged JSON file:

• EduData.json: 98,000 examples

Each example is stored in instruction-tuning format with the following fields:

• instruction: the full prompt, including the question stem and answer options
• input: an auxiliary input field; in the current release this is an empty string for all examples
• output: the target answer in natural language form, typically 答案为 A/B/C/D
• id: a unique sample identifier

Data Instance

{
  "instruction": "以下题目为单选题，只有一个正确选项，请根据问题文本和选项给出正确答案 题目文本为： 已知等差数列{$a_{n}},$满足$a_{2}+a_{11}=36,a_{8}=24,$则$a_{5}$等于$\\SIFChoice$ 选项为： (A) $6$ (B) $8$ (C) $10$ (D) $12$",
  "input": "",
  "output": "答案为  B",
  "id": "example-id"
}

Data Fields

• instruction The main textual prompt. In most cases it includes a fixed instruction prefix, the question body, and four answer options.

• input Reserved for optional auxiliary content. It is empty in the released EduData.json.

• output The gold answer label. In the current release, the answer is written as short Chinese text rather than as a bare class label.

• id A unique identifier for the example.

Data Characteristics

From inspection of the released EduData.json:

• Total examples: 98,000
• Unique IDs: 98,000
• All input fields are empty strings
• Answers cover the four options A/B/C/D
• The data is formatted as instruction-following examples rather than as a separately structured question / choices / label schema

Out-of-Scope Use

This dataset is not intended for:

• Commercial use
• High-stakes educational decision-making without human oversight

Curation Rationale

EduData was created to support research on educational LLMs in realistic multi-subject settings. Existing educational datasets are often narrow in subject coverage or insufficient for studying whether a model can generalize beyond a single domain. Our goal was to build a challenging, high-quality benchmark that better reflects practical educational use cases and enables research on cognitively informed data selection.

Source Data

The dataset was curated from Chinese mock examinations and college-entrance-exam-style materials. The released data focuses on single-choice questions and was organized for educational QA and instruction-tuning use.

Processing

The current public release is distributed as instruction-tuning JSON records. During preprocessing, the project also maintained split files that were merged into the final EduData.json.

Biases, Risks, and Limitations

• The dataset is Chinese-only and reflects one educational and cultural context
• It is centered on exam-style single-choice questions rather than open-ended pedagogy
• The merged release does not include explicit subject metadata per row
• Some formatting noise from source documents remains in a small number of examples
• Performance on this dataset should not be treated as a comprehensive measure of educational competence

Licensing

This dataset is released under CC BY-NC 4.0.

Commercial use is prohibited. Users are responsible for ensuring that their use complies with the dataset license and any applicable source-material restrictions.

Citation

If you use EduData or the CASS framework in your research, please cite:

@inproceedings{guo2026cass,
  title     = {From Diagnosis to Generalization: A Cognitive Approach to Data Selection for Educational LLMs},
  author    = {Yuxiang Guo and Yan Zhuang and Qi Liu and Zhenya Huang and Xianquan Wang and Liyang He and Jiatong Li and Rui Li and Shijin Wang},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year      = {2026}
}