A Chinese Elementary Science Question Dataset in Problem-Solving Process Generation

Area of Science:

• Educational Technology
• Artificial Intelligence in Education
• Science Education

Background:

• Large language models (LLMs) show promise for personalized science education.
• A key challenge is adapting LLM-generated problem-solving processes to specific student grade levels.
• Existing datasets may not adequately support the development of grade-level specific science problem-solving skills.

Purpose of the Study:

• To develop and introduce the Chinese Science Question (CSQ) dataset.
• To evaluate and enhance the science problem-solving capabilities of LLMs.
• To enable LLMs to generate grade-level appropriate science problem-solving processes.

Main Methods:

• Created a Chinese Science Question (CSQ) dataset with 12,000 high-quality samples.
• Included diverse question types and discipline properties across four primary school subjects.
• Designed a language model fine-tuned on the CSQ dataset to emulate student thought processes.

Main Results:

• The CSQ dataset, with its extensive annotations, effectively fine-tuned LLMs.
• Both automatic and human evaluations confirmed improvements in LLM performance.
• Generated problem-solving processes were successfully aligned with primary school students' grade levels.

Conclusions:

• The CSQ dataset is a valuable resource for advancing AI in science education.
• LLMs can be effectively trained to provide personalized, grade-appropriate science problem-solving support.
• This work contributes to developing more adaptive and effective educational tools for science learning.