OmniEdu: Open Foundation Models for Learning and Teaching
Hao Liang, Qihan Lin, Meiyi Qiang, Linzhuang Sun, Hengyi Feng, Mingrui Chen, Sizhe Qiu, Wentao Zhang
cs.CL
Quick verdict
Capability-balanced SFT lifts Qwen 4-27B across K-12 benchmarks; gains plausible, but same-named train/eval sources and missing ablations temper the claims.
Summary
OmniEdu is an open family of K-12 education language models fine-tuned from Qwen bases at 4B, 9B, and 27B scale. The authors build a 69,999-example instruction-tuning corpus (15.96M supervised response tokens) from over 100 sources, organized around four capabilities — subject competence, curriculum grounding, diagnostic reasoning, and pedagogical scaffolding — via a six-stage pipeline combining deterministic cleaning, LLM-based semantic auditing (Qwen3.5-122B), fine-grained quality scoring (GPT-5.6-Terra), token-budgeted k-center diversity selection, and 20 task-specific pedagogical system instructions. Across all three scales, tuned models beat their base checkpoints on curriculum grounding (OmniEdu-27B: 63.12% EM on K12-Bench vs 52.11%), K-12 problem solving (MDK12-Bench 57.76% vs 46.04%), and tutoring (MathTutorBench Scaffold win rate 78.74% vs 57.16%), while IFEval/GPQA/MMMU-Pro show no general degradation; OmniEdu-27B reports the best LongTutor Teaching average (3.02) among evaluated models. Key limitations: several training sources share names with evaluation benchmarks (TutorBench-v6, LongTutor-*, MathFish, K12-KGraph) and the leakage risk is not quantified; no ablation isolates the capability-oriented design versus a source-organized mixture; small margins lack significance testing; and some open-baseline scores are anomalously identical, suggesting harness incompatibilities.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Organizing instruction-tuning supervision around four educational capabilities (rather than by source or task) consistently improves curriculum grounding, K-12 problem solving, and pedagogical tutoring across 4B/9B/27B scales.
- OmniEdu-27B achieves the best open-weight results on curriculum-grounding and tutoring benchmarks, including 63.12% EM on K12-Bench, 86.95% MaC on EDUMATH, 78.74% MathTutorBench Scaffold win rate, and the highest LongTutor Teaching average (3.02) among evaluated models.
- A compact 69,999-example corpus (15.96M supervised response tokens), produced by a six-stage curation pipeline, suffices to turn general Qwen base models into stronger educational systems.
- Education-oriented specialization does not degrade general capability, with IFEval, GPQA, and MMMU-Pro all improving after tuning (e.g., MMMU-Pro 27B: 64.97%→67.98%).
Strengths
Explicit capability taxonomy with provenance
§3.1 assigns each example exactly one of four capabilities (subject competence, curriculum grounding, diagnostic reasoning, scaffolding) and retains source/license/audit metadata throughout; Fig 2 gives per-source counts for all 50+ subsets, which most educational-LLM papers omit.
Consistent paired base-vs-SFT gains at the overall-metric level
Every tuned model beats its base on all three capability groups (Tables 1-3): 27B K12-Bench EM 52.11→63.12%, MDK12-Bench 46.04→57.76%, MathTutorBench Scaffold 57.16→78.74%; general checks (Tables 13-15) show IFEval/GPQA/MMMU-Pro also improve, so the headline claim is internally consistent as stated.
Above-average artifact release
Documented, quantitative curation funnel
Weaknesses
The central causal claim is never ablated
Unquantified train/eval overlap via same-named sources
No statistical rigor anywhere
Anomalous open-baseline scores undermine comparisons
Subset regressions glossed over; base-checkpoint naming inconsistent
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
OmniEdu — Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is OmniEdu?
An open family of K-12 education language models (4B, 9B, 27B parameters) fine-tuned from Qwen base models on a curated 69,999-example instruction corpus organized around four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical scaffolding. Checkpoints, dataset, and code are linked on Hugging Face and GitHub.
How much does OmniEdu-27B improve over its base model?
Per the paper's paired comparisons: K12-Bench exact match 52.11%→63.12%, MDK12-Bench 46.04%→57.76%, MathTutorBench Scaffold win rate 57.16%→78.74%, LongTutor Evidence 36.80%→78.20%. General benchmarks (IFEval, GPQA, MMMU-Pro) also slightly improve, so the authors report no measurable general-capability loss.
Does OmniEdu beat GPT or Claude on education tasks?
Partially. OmniEdu-27B reports the best K12-Bench EM (63.12% vs GPT-5.4's 43.23%) and the highest LongTutor Teaching average (3.02) among evaluated models, but Claude-Opus-5 scores higher on MathTutorBench Scaffold (87.89% vs 78.74%) and LongTutor Evidence (84.17% vs 78.20%), and Kimi-K3 leads on EDUMATH (90.00% vs 86.95%) and TutorBench (63.65% vs 59.42%). Independent verification is advisable since some training data shares names with the evaluation benchmarks.
Is the OmniEdu training data public?
Yes. The paper links a Hugging Face dataset (lhpku20010120/Omni-Edu) containing 69,999 examples with 15.96M supervised response tokens, alongside the three model checkpoints and a project repository with stated release metadata (source inventory, licenses, auditing prompts).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 are available only as embedded text; visual layout and any figure-only details were reconstructed from captions and inline labels.
- Auditing prompts, the full source inventory, licensing details, and benchmark-protocol specifics (judge identity, prompts, per-benchmark item counts) are deferred to 'release metadata' and are not present in the provided text.
- The paper references 2026-era models (Qwen3.5/Qwen3.8, GPT-5.4/5.6-Terra/Sol, Claude-Opus-5, GLM-5.3, Kimi-K3) whose versions and behaviors could not be verified independently from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.