Paper Explained: OmniEdu — Building an Open Education LLM Around Four Skills: Solve, Place, Diagnose, Guide
A from-scratch walkthrough of OmniEdu, an open K-12 LLM family: how ~1.34M candidate examples were curated down to 60,951, how per-example teaching instructions let one model either answer or hold back, and what the results and limitations actually show.
Key points
- OmniEdu is an open model family for K-12 learning and teaching, built by fine-tuning 4B, 9B, and 27B models.
- The education data is organized around four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding.
- Each example gets exactly one primary capability, and within each capability examples are split into fine-grained task buckets by format and supervision type.
- About 1.34M education candidates are narrowed in six stages, and adding 9,048 general instruction examples gives 69,999 examples and 15.96M tokens.
- Fine-tuning 4B, 9B, and 27B models improves all three groups of educational benchmarks at every scale.
OmniEdu: Open Foundation Models for Learning and Teaching
Primary source — what this article is built on
undefined2026-09-19→undefined2026-09-24same month
OmniEdu: Open Foundation Models for Learning and TeachingHao Liang, Qihan Lin, Meiyi Qiang et al. · 2026-09-19 · v1arXiv:2609.23088Paper page·PDFundefined
Educational foundation models must solve problems, understand curriculum structure, diagnose learner difficulties, and provide appropriate instructional support. Existing educational language models often focus on either problem solving or tutoring, with training mixtures organized by source or task rather than capability. We present OmniEdu, an open family of foundation models for K-12 learning and teaching. Its instruction-tuning corpus combines over 100 educational resources and general instruction sources, organized around four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding. Our pipeline integrates deterministic cleaning, semantic auditing and rewriting, task-specific quality scoring, token-budgeted diversity selection, and pedagogical instruction assignment. It yields 69,999 examples and 15.96M supervised response tokens, including 60,951 education-specific examples. We fine-tune 4B, 9B, and 27B models and evaluate curriculum grounding, K-12 problem solving, and pedagogical tutoring, alongside general capability. Education-oriented tuning consistently improves all three educational benchmark groups across model scales. OmniEdu-27B achieves 63.12% EM and 76.69% F1 on K12-Bench, 85.89% on MathFish, 86.95% on EDUMATH, and 78.74% in MathTutorBench's Scaffold setting. It also achieves the highest Teaching average on LongTutor among the evaluated models, at 3.02. These results demonstrate the value of curated, capability-balanced supervision for adapting general language models to educational tasks spanning problem solving, curriculum understanding, and instructional support.
What this paper claims
The paper is "OmniEdu: Open Foundation Models for Learning and Teaching" (arXiv:2609.23088, September 2026).
The abstract, in short: an educational foundation model has to do more than solve problems. It must understand curriculum structure, diagnose where a learner is struggling, and give the right kind of instructional support. Existing education models tend to specialize in either problem solving or tutoring, and their training mixtures are usually organized by source or task rather than by capability. The authors present OmniEdu, an open model family for K-12 learning and teaching. They organize more than 100 educational resources plus general instruction data around four capabilities, then run cleaning, semantic auditing, quality scoring, diversity selection, and teaching-instruction assignment to produce 69,999 examples with 15.96M supervised response tokens. Fine-tuning 4B, 9B, and 27B models improves all three groups of educational benchmarks at every scale.
An analogy: the teacher who answers vs. the teacher who teaches
Picture two math tutors. The first, when asked, writes out the answer and the method right away. The second looks at the student's work, notices "they're stuck on common denominators, a gap from 5th grade," and instead of answering asks, "What happens when you make the denominators match?"
The difference is judgment: where the problem sits in the curriculum, why the student got it wrong, and whether this is a moment to answer or to stay quiet. The paper argues that an educational model has to coordinate what to teach, where the learner is, and how to respond, and that treating education as plain question answering throws that structure away (§1).
Decide what to teach first: four capabilities
The authors frame this partly as a data-design problem. A short answer and a tutoring exchange teach very different behaviors; mix them only by source or subject and the intended response policy gets blurred (§1). So the education data is split into four capabilities (§3.1):
- Subject competence: solve problems and explain the answers
- Curriculum grounding: grade level, knowledge points, prerequisites, difficulty
- Diagnostic reasoning: find the misconception or missing prerequisite in a student's work
- Pedagogical action and scaffolding: pick between asking, hinting, revisiting a prerequisite, or explaining directly, and then do it well
Each example gets exactly one primary capability, and within each capability examples are further split into fine-grained "task buckets" by format and supervision type (§3.1).
From 1.34M to 61K: the data funnel
About 1.34M education candidates are narrowed down in six stages (§3, Figure 1).
| Stage | Education examples remaining |
|---|---|
| Collected | ~1,340,000 |
| Deterministic cleaning + eval decontamination (§3.2) | 870,711 |
| LLM semantic auditing and rewriting (§3.3) | 440,100 |
| Preliminary diversity selection + fine-grained scoring (§3.4) | 121,318 |
| Token-budgeted diversity selection (§3.5) | 60,951 |
Adding 9,048 general instruction examples brings the total to 69,999 examples and 15.96M tokens (§3.7).
Comments
Sign in to comment