OmniEdu: Open Foundation Models for Learning and Teaching

AI summary

Overview

Research area: Natural Language Processing, specifically educational foundation models for K–12 learning and teaching.

Technical level: Advanced. The paper combines a multi-stage, LLM-assisted data curation pipeline, supervised fine-tuning at three model scales, and evaluation across nine-plus educational benchmarks plus three auxiliary general-capability benchmarks.

Scope: One sentence — OmniEdu is an open family of 4B, 9B, and 27B educational foundation models trained on a capability-oriented instruction corpus of 69,999 examples (15.96M supervised response tokens) and evaluated across curriculum grounding, K–12 problem solving, and pedagogical tutoring.

What This Paper Is About

Educational language models are often built to do one thing well: answer subject questions or hold a tutoring dialogue. The authors argue that a genuinely useful teaching model must do both, and must also locate a problem inside a curriculum and diagnose why a learner got it wrong.

The paper's core problem is therefore a data-design problem: educational instruction data are heterogeneous, and mixing them by source or subject alone creates a vague training signal with no clear capability target. OmniEdu's goal is to build an open model family whose training mixture is explicitly organized around four educational capabilities — subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding — and to show that this composition reliably improves all three educational capability groups.

Key Contributions

1. An open family of K–12 foundation models — OmniEdu-4B, OmniEdu-9B, and OmniEdu-27B — organized around four capabilities connecting subject knowledge, curriculum structure, learner diagnosis, and pedagogical action.

2. A reproducible, capability-oriented data construction pipeline combining semantic quality control, task-specific filtering, token-budgeted diversity selection, and explicit pedagogical instructions, producing a 69,999-example training mixture with 15.96M supervised response tokens.

3. A cross-capability evaluation showing consistent gains across curriculum grounding, K–12 problem solving, and pedagogical tutoring at all three model scales, together with auxiliary tests of general capability.

4. Public release entry points for reproducibility: a project repository and a Hugging Face dataset page, plus model pages for OmniEdu-4B, OmniEdu-9B, and OmniEdu-27B.

Main Findings

• Curriculum grounding improves at every scale. OmniEdu-27B raises K12-Bench overall EM from 52.11% to 63.12% and F1 from 73.48% to 76.69%, the best overall performance among evaluated models. On MathFish, OmniEdu-27B improves overall accuracy from 83.54% to 85.89%. On EDUMATH, MaC scores improve by 20.80, 15.40, and 16.35 points for the 4B, 9B, and 27B models, with OmniEdu-27B reaching 86.95%, second only to Kimi-K3's 90.00%.

• K–12 problem solving improves on all three benchmarks. OmniEdu-27B raises the GAOKAO-Bench full-score rate from 91.55% to 94.87%, EXAMS-V overall accuracy to 69.52%, and MDK12-Bench from 46.04% to 57.76%, with open-question correctness rising from 62.08% to 78.77%.

• Pedagogical tutoring shows the largest gains. On MathTutorBench, OmniEdu-27B's Scaffold and Scaffold-hard win rates rise from 57.16% and 55.86% to 78.74% and 83.59%. Scaffold win rate improves by 55.37 points for OmniEdu-4B and 61.26 points for OmniEdu-9B. For LongTutor, the Evidence average rises from 25.67% to 65.88% (4B), 5.81% to 66.63% (9B), and 36.80% to 78.20% (27B), while OmniEdu-27B obtains the best Teaching average of 3.02 among evaluated models. On TutorBench, OmniEdu-27B reaches 59.42%, ranking second behind Kimi-K3.

• Diagnosis remains hard. OmniEdu-27B and OmniEdu-9B rank first and second in LongTutor knowledge-state diagnosis accuracy at 54.04% and 53.55%, but the authors note these accuracies are relatively low and that the task remains comparatively challenging.

• General capability is not sacrificed. MMMU-Pro overall accuracy increases from 50.46% to 52.60% for 4B, 58.38% to 60.75% for 9B, and 64.97% to 67.98% for 27B. Tuning generally improves IFEval metrics and improves both GPQA Main and Diamond accuracy, though changes vary by MMMU-Pro discipline.

• Proprietary comparison. OmniEdu-27B stays competitive with substantially larger proprietary models on several K–12 problem-solving evaluations, though frontier proprietary models retain an advantage on some exam-style evaluations.

Methodology in Plain English

The authors start by sorting educational data by behavior to be taught rather than by subject or dataset. They collect from more than 100 datasets and educational resources, forming an initial education-specific pool of approximately 1.34M examples, and assign each example to one of four capability categories, then to finer task buckets.

Six curation stages follow:

1. Taxonomy and source collection. Two components: 9,048 general-purpose instruction examples from DataFlow-Instruct-10K (7,431), the Tulu-3-SFT-mixture (1,495), and MathV360K (122); plus the education-specific pool.
2. Deterministic cleaning and decontamination. Standardize formats, remove exact duplicates and malformed examples, verify answer consistency, keep only K–12-relevant content from mixed-domain sources, check that images are accessible and aligned, and remove anything overlapping the evaluation set. 870,711 examples remain.
3. Semantic auditing and rewriting. Qwen3.5-122B-A10B-FP8 scores each example 0–100 with a task-specific prompt and labels it keep (85–100), rewrite (50–84), or remove (below 50). Rewritten examples keep the original input and are re-audited. 440,100 examples remain.
4. Preliminary diversity selection and fine-grained filtering. Heavily overrepresented sources are compressed with k-center greedy over BGE-M3 embeddings (RACE from 60,180 to 5,000; AquilaEdu from 23,226 to 2,000; CJEval from 17,178 to 2,000). GPT-5.6-Terra then scores task-specific dimensions on a 1–5 scale, retaining examples only if all applicable dimensions score at least 3, with a threshold of 4 for task-critical dimensions such as correctness, validity, and grounding. 121,318 candidates remain.
5. Token-budgeted diversity selection. Within each task bucket, k-center greedy selection proceeds until a supervised-response-token budget is met, because a tutoring dialogue can contribute an order of magnitude more tokens than a short hint. This yields 60,951 examples and approximately 12.0M supervised response tokens.
6. Pedagogical instruction assignment. Each example is paired with one of 20 task-specific system-instruction templates describing the desired response behavior (for example, withholding the answer during scaffolding versus justifying a final answer in problem solving), inspired by the pedagogical instruction following idea from LearnLM.

Training uses LlamaFactory on Qwen3.5-4B-Base, Qwen3.5-9B-Base, and Qwen3.8-27B, with a learning rate of 5 × 10⁻⁶, a maximum sequence length of 32,768, and 3 epochs. Evaluation spans curriculum-grounding benchmarks (K12-Bench, MathFish, EDUMATH), K–12 problem-solving benchmarks (GAOKAO-Bench, EXAMS-V, MDK12-Bench), tutoring benchmarks (MathTutorBench, TutorBench, LongTutor), and auxiliary diagnostics (IFEval, GPQA, MMMU-Pro). Baselines include open-weight educational models Confucius3-Math, MuduoLLM, and three EduChat variants, plus proprietary models GPT-5.4, GPT-5.6-Sol, Claude-Opus-5, GLM-5.3, and Kimi-K3. Five open-weight educational LLMs and GLM-5.3 do not support image inputs and received only the textual component for image-dependent examples.

Why This Matters

Impact on research. The paper reframes educational model quality as a data-composition question rather than a scale question. It shows that target behaviors can be separated from example content through pedagogical system instructions, and that a 69,999-example corpus with 15.96M supervised response tokens can shift three capability groups at every tested scale. It also supplies a fully open model family and dataset where many comparable systems are closed or depend on product-level orchestration.

Real-world applications.

• K–12 tutoring assistants that scaffold reasoning instead of immediately revealing answers.
• Curriculum-alignment tools that map a problem to knowledge points, grade level, and prerequisite structure.
• Teacher-facing tools that generate standards-aligned practice problems and diagnose student mistakes.
• Personalized tutoring systems that ground responses in a student's long-term learning history.

Industry relevance. The results show that a 27B open model can reach 63.12% EM / 76.69% F1 on K12-Bench, 85.89% on MathFish, and 78.74% Scaffold win rate on MathTutorBench, remaining competitive with substantially larger proprietary systems. Because the corpus, pipeline, and weights are public, the approach is reproducible for organizations that cannot rely on closed APIs.

Future Directions

• Knowledge-state diagnosis. The authors report that OmniEdu-27B's 54.04% and OmniEdu-9B's 53.55% diagnosis accuracies are relatively low, identifying learner-state inference as an unresolved problem.
• Scaling and mixture tuning. Only 4B, 9B, and 27B scales were tested; whether the capability-balanced recipe holds at larger or smaller scales, and how the token budget per task bucket should be set, remain open.
• Broadening capability coverage. The current taxonomy covers four categories; the paper does not report how adding categories such as emotional support or multimodal classroom interaction would affect the mixture.
• Deeper general-capability accounting. The paper reports improvements on MMMU-Pro, IFEval, and GPQA but notes that changes vary by MMMU-Pro discipline, leaving a more granular analysis of specialization costs open. Note that the paper content provided is truncated mid-way through Appendix Table 7, so the detailed GAOKAO-Bench, EXAMS-V, and MDK12-Bench breakdowns and the full Appendix 6.4 general-capability tables are not available here.

Target Audience

Researchers and engineers working on domain-specific LLM post-training, especially those interested in data-centric methods for instruction tuning. Also relevant to teams building education products (tutoring systems, curriculum tooling, teacher-facing platforms), to evaluation researchers designing pedagogy-oriented benchmarks, and to educators or institutions that need open-weight models they can inspect and adapt. A working familiarity with supervised fine-tuning, instruction tuning, and benchmark evaluation is assumed for the experiments section.

Authors’ abstract

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.