lhpku20010120/Omni-Edu-4B
OmniEdu-4B is a 4.5 billion parameter multimodal foundation model developed by Peking University, University of the Chinese Academy of Sciences, and Zhongguancun Academy, fine-tuned from Qwen3.5-4B-Base. It is specifically designed for K–12 learning and teaching, integrating subject competence, curriculum grounding, diagnostic reasoning, and pedagogical scaffolding. With a 32,768-token context length, it excels at educational tasks like identifying misconceptions, providing targeted feedback, and adapting instructional support.
Loading preview...
OmniEdu-4B: Open Foundation Model for K–12 Education
OmniEdu-4B is a 4.5 billion parameter multimodal model, fine-tuned from Qwen3.5-4B-Base, developed by researchers from Peking University, University of the Chinese Academy of Sciences, and Zhongguancun Academy. It is part of the OmniEdu family, which focuses on creating open foundation models for learning and teaching. This 4B checkpoint serves as an accessible starting point for local experimentation and educational applications, offering a 32,768-token training sequence length.
Key Capabilities
OmniEdu-4B is designed to address the core needs of K–12 education by connecting what is taught, what learners understand, and what instructional steps to take next. Its training is organized around four complementary capabilities:
- Subject Competence: Solves K–12 problems across various subjects and input formats.
- Curriculum Grounding: Links questions, concepts, and solutions to established curriculum standards and prerequisites.
- Diagnostic Reasoning: Identifies learner misconceptions, missing prerequisites, and gaps in understanding.
- Pedagogical Action and Scaffolding: Provides targeted feedback, asks guiding questions, offers explanations, and adapts instructional support.
Training and Performance
The model was trained using full-parameter supervised fine-tuning on a mixture of 69,999 instruction examples, including over 60,000 education-specific examples and 9,048 general-purpose examples, with 15.96M supervised response tokens. Evaluation results demonstrate its strong performance on various K–12 educational benchmarks, including K12-Bench (54.25% EM, 71.75% F1), MathFish (83.19% Acc.), and GAOKAO-Bench (90.46% Full).
Good for
- Educational Research: Ideal for academic studies on AI in education.
- Teacher-Assistance Tools: Developing applications to support educators in lesson planning, assessment, and student interaction.
- Learning-Support Prototypes: Creating innovative tools for personalized learning, tutoring, and diagnostic feedback for K–12 students.
- Multimodal Educational Tasks: Handling questions that involve both text and images, such as analyzing diagrams or problem statements with visual components.