PaperlayerView exact revision on arXiv

Source-linked AI summary

OmniEdu: Open Foundation Models for Learning and Teaching

Hao Liang, Qihan Lin, Meiyi Qiang, Linzhuang Sun, Hengyi Feng, Mingrui Chen, Sizhe Qiu, Wentao Zhang

arXiv:2609.23088v1cs.CL

TL;DR

Educational models must combine problem solving with curriculum understanding, learner diagnosis, and instructional support, but prior systems often focus on only parts of this work. OmniEdu organizes an open K–12 model family and its training corpus around four capabilities. Education-oriented tuning improves all three educational benchmark dimensions across model scales, with OmniEdu-27B achieving the highest LongTutor Teaching average among evaluated models.

  • Problem

    Educational usefulness requires more than answer accuracy, yet educational models often focus on problem solving or tutoring and organize data by source or task rather than capability.

    Source: §1 · p. 3

  • Method

    OmniEdu is an open K–12 model family trained on a corpus organized around subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding.

    Source: §1 · pp. 3–4, §3.1 · p. 6

  • Results

    Tuning improves all three educational dimensions across model scales; OmniEdu-27B achieves a Teaching average of 3.02 on LongTutor, the highest among evaluated models.

    Source: §1 · p. 4, §4.4 · p. 11

  • Takeaways & Limitations

    The results extend beyond correct answers to curriculum reasoning, error-sensitive teaching, and use of student history.

    Source: §1 · p. 4

Abstract

from arXiv · show

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.

1 Introduction

Educational models need to coordinate problem solving, curriculum understanding, learner diagnosis, and instructional response, but these capabilities are often developed separately. OmniEdu organizes K–12 training and evaluation around four complementary capabilities and reports gains across educational dimensions and model scales.

1 Introduction

  • Existing educational models often develop or evaluate problem solving, curriculum understanding, and pedagogical interaction separately, leaving no open family explicitly organized around the full learning–teaching loop.A correct answer may still fail to locate the relevant curriculum concept, explain an error, or provide a scaffold that preserves the learner’s opportunity to reason.Source: §1 · p. 3
  • OmniEdu organizes K–12 supervision around subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding.The taxonomy covers solving and explaining problems, mapping curriculum structure, identifying errors and missing prerequisites, and selecting interventions such as hints or prerequisite review.Source: §1 · p. 3
  • Its data-centric pipeline combines deterministic cleaning and decontamination with LLM-assisted semantic auditing, task-specific quality scoring, diversity-aware selection, and pedagogical instructions, yielding 69,999 examples and 15.96M supervised response tokens.This comprises 60,951 education-specific examples and 9,048 general-purpose instruction examples.Source: §1 · p. 3
  • OmniEdu is an open model family whose cross-capability evaluation also measures general instruction following, scientific reasoning, and multimodal understanding.The authors identify data composition and pedagogical behavior as central design axes for open educational foundation models.Source: §1 · pp. 3–4
  • Education-oriented tuning improves all three educational dimensions across 4B, 9B, and 27B scales; OmniEdu-27B reaches 63.12% EM and 76.69% F1 on K12-Bench.It also records 85.89% accuracy on MathFish, 86.95% MaC on EDUMATH, 78.74% Scaffold win rate on MathTutorBench, and a 3.02 Teaching average on LongTutor.Source: §1 · pp. 3–4

2 Related Work

Prior educational models, data methods, and benchmarks address important pieces of learning and teaching, but coverage is often specialized. OmniEdu is positioned as a unified, capability-oriented open model family evaluated across learner, curriculum, and pedagogical dimensions.

2.1 Educational Foundation Models and AI Tutors

  • Open educational models cover broad interactions or curriculum-aligned tasks, while other systems specialize in math or tutoring; many frontier tutoring systems are closed or depend on product orchestration.OmniEdu complements these efforts with an open family targeting learning- and teaching-oriented capabilities jointly.Source: §2.1 · p. 4

2.2 Educational Data and Data-Centric Post-Training

  • Data-centric instruction-tuning work emphasizes quality, complexity, diversity, and capability relevance [37], [2], [18], [30], while educational datasets often target particular subjects or tasks [11], [17], [3].OmniEdu extends educationally structured data toward a unified capability-oriented post-training mixture [3].Source: §2.2 · pp. 4–5

2.3 Evaluation of Educational LLMs

  • Educational benchmarks assess learner competence, curriculum understanding, and pedagogy, extending from final-answer accuracy toward process reasoning, scaffolding, diagnosis, and long-term student modeling.This grouping reflects a shift from whether models solve problems to whether they understand what should be taught and how.Source: §2.3 · p. 5

3 Data Preparation

OmniEdu builds a capability-organized K–12 corpus through staged cleaning, auditing, quality filtering, diversity selection, and task-specific instruction assignment. The pipeline reduces a large heterogeneous pool to a compact mixture while retaining provenance and audit metadata.

3 Data Preparation

  • The six-stage pipeline curates approximately 1.34M education-specific candidates into 60,951 examples, then combines them with 9,048 general examples for a 69,999-example corpus.The final mixture contains 15.96M supervised response tokens.Source: Figure 1 · p. 5, §3 · p. 5, §3.7 · p. 8

3.1 Capability Taxonomy and Source Collection

  • Four behavior-based categories organize educational data: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding.Examples receive one primary capability, with other properties retained as secondary metadata; finer task buckets account for task form, modality, and supervision type.Source: §3.1 · p. 6

3.2 Deterministic Cleaning and Evaluation Decontamination

  • Deterministic cleaning standardizes sources, removes duplicates and malformed or irrelevant examples, checks verifiable answers and image alignment, retains provenance, and decontaminates evaluation data.After this stage, 870,711 education-specific examples remain.Source: §3.2 · p. 6

3.3 LLM-assisted Semantic Auditing and Rewriting

  • Semantic auditing assigns task-specific usability scores and keep, rewrite, or remove actions, checking correctness, curriculum alignment, diagnosis support, and scaffolding quality.Rewritten supervision keeps the original input fixed and must pass the same audit; 440,100 examples remain after this stage.Source: §3.3 · pp. 6–7

3.4 Preliminary Diversity Selection and Fine-grained Quality Filtering

  • K-center greedy selection first compresses highly overrepresented sources, then task-specific scoring retains examples meeting minimum quality thresholds.Examples must score at least 3 on applicable dimensions, with a threshold of 4 for critical dimensions such as correctness, validity, and grounding; the filtered pool contains 121,318 examples.Source: §3.4 · p. 7

3.5 Token-budgeted Diversity Selection

  • Final selection uses k-center greedy within fine-grained task buckets and allocates budgets by supervised response tokens, accounting for large differences in response length.The education-specific pool is reduced to 60,951 examples containing approximately 12.0M supervised response tokens.Source: §3.5 · pp. 7–8

3.6 Pedagogical Instruction Assignment and Final Assembly

  • Each selected example receives one of 20 task-specific system instructions to make its desired response behavior explicit, from reasoned problem solving to Socratic questioning and tutoring.This distinguishes behaviors such as giving and justifying an answer from scaffolding that may withhold it while guiding intermediate reasoning.Source: §3.6 · p. 8

3.7 Final Data Statistics

  • The final corpus contains 60,951 education-specific examples and approximately 12.0M supervised response tokens, plus 9,048 general-purpose examples, for 69,999 examples and 15.96M tokens.Figure 2 presents the capability-category and source-subset distributions alongside dataset quantities.Source: Figure 2 · p. 7, §3.7 · p. 8

4 Experiments

OmniEdu is evaluated across curriculum grounding, K–12 problem solving, tutoring, and general capabilities. Education-oriented tuning improves all three educational dimensions at every model scale, with particularly strong results in curriculum grounding and tutoring.

4.1 Experimental Setup

  • The evaluation tests curriculum grounding, authentic K–12 problem solving, tutoring quality, and general capability retention across three fine-tuned base-model scales.Baselines include open-weight educational and frontier proprietary models; benchmarks span graph-derived curriculum tasks, school exams, tutoring, and auxiliary general-capability tests.Source: §4 · p. 8, §4.1 · pp. 8–9

4.2 Curriculum Grounding

  • OmniEdu-27B raises K12-Bench EM from 52.11% to 63.12% and F1 from 73.48% to 76.69%, achieving the best overall performance.Gains extend across all subtasks, including knowledge grounding and structural curriculum reasoning.Source: §4.2 · p. 9
  • OmniEdu-27B improves MathFish accuracy from 83.54% to 85.89% and reaches 86.95% MaC on EDUMATH, second only to Kimi-K3.EDUMATH MaC improves by 16.35 points for the 27B model.Source: §4.2 · pp. 9–10

4.2 Curriculum Grounding; 4.3 K--12 Problem Solving; 4.4 Pedagogical Tutoring

  • Education-oriented tuning improves all three curriculum and tutoring benchmarks at every scale; tuned models beat their bases on all three K–12 problem-solving benchmarks, with OmniEdu-27B top three overall and competitive with frontier proprietary models.Source: §4.2 · p. 10, §4.3 · p. 10, §4.4 · p. 11

4.3 K--12 Problem Solving

  • All tuned models beat their base models on all three problem-solving benchmarks; OmniEdu-27B reaches 57.76% on MDK12-Bench.It ranks within the top three overall on each benchmark and outperforms evaluated open-weight educational baselines.Source: §4.3 · p. 10

4.4 Pedagogical Tutoring

  • Tuning improves all three tutoring benchmarks at every scale; OmniEdu-27B leads LongTutor’s Teaching average at 3.02.Its MathTutorBench Scaffold win rate rises to 78.74%; TutorBench overall score is 59.42%, second behind Kimi-K3.Source: §4.4 · p. 11

4.5 General Capability Retention

  • Education-oriented tuning generally improves IFEval and GPQA, while MMMU-Pro accuracy rises at all three scales.MMMU-Pro 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; changes vary by discipline.Source: §4.5 · p. 11

5 Conclusion

OmniEdu is an open family of K–12 learning and teaching models trained with capability-oriented supervision. Across 4B, 9B, and 27B scales, it improves educational benchmarks, with strong open-weight results in curriculum grounding and tutoring and competitiveness on problem solving.

5 Conclusion

  • The corpus organizes K–12 supervision around subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding.Its construction combines deterministic cleaning, semantic auditing and revision, task-specific quality assessment, and diversity-aware sampling.Source: §5 · p. 12
  • OmniEdu improves across model scales, with strong open-weight results on curriculum-grounding and tutoring benchmarks and competitive 27B results on K–12 problem solving.The conclusion reports that the 27B model substantially outperforms existing open-weight educational models on problem solving.Source: §5 · p. 12

A Detailed Evaluation Results

The appendix provides complete model-by-model results underlying the main-text summary tables.

A Detailed Evaluation Results

  • Complete model-by-model results retain the original table structure so metrics and baselines can be audited directly.The appendix describes this as the basis for the compact main-text summary tables.Source: A Detailed Evaluation Results · p. 15

A.1 Curriculum Grounding

The curriculum-grounding appendix reports results across three benchmarks and multiple forms of curriculum understanding.

A.1 Curriculum Grounding

  • Detailed results cover curriculum relations, positive and negative alignment cases, and complementary metrics for curriculum-grounded problem generation.The appendix reports that tuning gains extend across these evaluation settings rather than being driven by one setting.Source: §A.1 · p. 15

A.2 K--12 Problem Solving

Education-oriented tuning generally improves K–12 problem-solving results across subjects, modalities, and question formats, though gains vary among subsets.

A.2 K--12 Problem Solving

  • Gains from education-oriented tuning generally extend across subjects, modalities, and problem formats, with improvement magnitude varying across subsets.The section reports breakdowns for GAOKAO-Bench, EXAMS-V, and MDK12-Bench.Source: §A.2 · p. 15
  • GAOKAO-Bench results are broken down by question type and school subject.Source: §A.2 · p. 15
  • EXAMS-V results distinguish modalities, subject groups, and languages, while MDK12-Bench separates objective and open-ended questions.Source: §A.2 · p. 15

A.3 Pedagogical Tutoring

Education-oriented tuning strengthens pedagogically appropriate, context-sensitive tutoring across complementary tutoring settings, which test scaffolding, feedback, and longer-interaction teaching quality.

A.3 Pedagogical Tutoring

  • Across complementary tutoring settings, education-oriented tuning consistently strengthens pedagogically appropriate and context-sensitive responses.Source: §A.3 · p. 15
  • MathTutorBench tests scaffolded and pedagogy-oriented tutoring under standard and harder settings.Source: §A.3 · p. 15
  • TutorBench covers multiple tutoring use cases in text and multimodal settings, while LongTutor evaluates evidence use, knowledge-state diagnosis, and teaching quality over longer histories.Source: §A.3 · p. 15

A.4 General Capability Retention

General-capability diagnostics assess whether education-oriented tuning changes instruction following, scientific reasoning, or multimodal reasoning, with paired base/SFT results enabling direct inspection.

A.4 General Capability Retention

  • The general-capability diagnostics test instruction following, scientific reasoning, and multimodal reasoning for changes associated with education-oriented tuning.Source: §A.4 · p. 15
  • Results are presented in paired base/SFT form to make the effect of tuning directly auditable.Source: §A.4 · p. 15

B Training Configuration

OmniEdu models use full-parameter supervised fine-tuning with shared hyperparameters across the 4B, 9B, and 27B scales; accompanying tables provide detailed benchmark breakdowns and general-capability diagnostics.

B Training Configuration

Related papers

Loading 2609.23088v1…