OmniEdu is an open family of K-12 educational LLMs (4B/9B/27B) fine-tuned on a 69,999-example corpus organized around four teaching capabilities (subject skill, curriculum grounding, error diagnosis, and scaffolding), showing that capability-balanced supervision beats source-mixed instruction data for tutoring behavior.
Say you’re building a homework helper. A model that just outputs the correct answer often skips the parts a real tutor does: identifying which curriculum standard the problem hits, spotting the specific misconception behind a wrong answer, and choosing whether to give a hint, ask a Socratic question, or explain directly. General-purpose LLMs and existing educational models tend to be strong at one of these and weak at the others. The authors point out that prior open educational LLMs like EduChat, MuduoLLM, and Confucius3-Math usually specialize (math-only, or Chinese K-12 only, or tutoring-only), and their training data is grouped by source or subject rather than by the teaching behavior each example is meant to induce. Google’s LearnLM proposes conditioning behavior on the learning scenario during post-training, and OmniEdu borrows that framing but delivers open weights and open data.
The core idea is that training examples should be labeled by what capability they teach, not by what subject they cover. The authors define four buckets: subject competence (solve the problem), curriculum grounding (place the problem in a standards hierarchy, identify prerequisites), diagnostic reasoning (given a student’s wrong work, name the misconception), and pedagogical action and scaffolding (pick a hint, question, or explanation and deliver it without leaking the answer).
They then run a six-stage funnel over ~100 source datasets. Roughly 1.34M raw education examples get filtered down to 60,951, plus 9,048 general-instruction examples. The stages: dedupe and decontaminate against the eval sets; use a large LLM auditor to score each example 0-100 and either keep, rewrite, or drop it; apply a stricter per-dimension rubric with a frontier model (e.g. correctness, grounding, scaffolding quality must clear thresholds); use k-center greedy over BGE-M3 embeddings to pick a semantically diverse subset per task bucket; and budget each bucket by supervised response tokens rather than example count, since a tutoring dialogue is worth an order of magnitude more tokens than a short answer.
The last step is the pedagogically interesting one: each example gets one of 20 task-specific system prompts describing the desired response behavior (solve-and-justify vs. Socratic vs. diagnose-and-correct, etc.), so the same base model learns to switch behavior based on the system instruction rather than baking one style in.
for ex in raw_pool: # ~1.34M
if not clean_and_decontaminate(ex): continue
score, action = llm_audit(ex) # 0-100 + keep/rewrite/remove
if action == "remove": continue
if action == "rewrite": ex = repair(ex)
if not fine_grained_rubric(ex): continue # per-dim >=3, critical >=4
bucket[ex.task].add(ex)
selected = [k_center_greedy(b, token_budget=B[b]) for b in bucket]
for ex in selected:
ex.system_prompt = TEMPLATES[ex.task] # 1 of 20
The result is fine-tuned with standard full-parameter SFT on Qwen3.5-Base checkpoints at 4B, 9B, and 27B.
Every scale improves over its own base model on every educational axis, which is the cleanest signal the paper offers because base-vs-tuned isolates the data recipe.
•
Curriculum grounding. On K12-Bench, OmniEdu-27B goes from 52.11% to 63.12% EM over its base. On EDUMATH (generating standards-aligned word problems), the tuned 4B/9B/27B gain 20.8, 15.4, and 16.4 MaC points respectively, with 27B hitting 86.95%, behind only Kimi-K3 among proprietary systems.
•
Tutoring. The largest gains show up here. On MathTutorBench’s Scaffold win rate, the 9B model jumps by 61 points (14% to 75%). On LongTutor’s Evidence score (does the model actually use the student’s prior history?), the 9B goes from 5.81% to 66.63%. The 27B posts the best Teaching average (3.02) among all evaluated models, including Claude-Opus-5 and Kimi-K3.
•
K-12 problem solving. Gains are more modest and don’t unseat frontier proprietary models: 27B improves MDK12-Bench full-score from 46.04% to 57.76%, still trailing Claude-Opus-5 (57.46%) and Kimi-K3 (63.60%) on some subsets.
•
General capability. Auxiliary checks on IFEval, GPQA, and MMMU-Pro actually improve slightly after tuning, so the specialization doesn’t obviously cost general skill on these three benchmarks.
One important nuance: the authors attribute the tutoring jump to the pedagogical-instruction assignment and capability-balanced mixture, but they don’t run a controlled ablation removing just the system-prompt component or just the diversity budget. So the whole-recipe result is strong; the causal decomposition of which stage matters most is not established.
•
If you’re building a tutoring product on open weights, the released training corpus and the 27B checkpoint are directly usable, and the K-12 focus is Chinese-curriculum-heavy so check fit for your audience.
•
If you’re curating your own instruction data, the more transferable lesson is the pipeline shape: label examples by the behavior they should induce, budget by response tokens not counts, and attach a task-specific system prompt rather than a single generic one. This is worth testing even outside education, because the pathology (source-mixed data conflating incompatible response styles) is common.
•
If you’re evaluating tutoring models, note that answer-accuracy benchmarks and tutoring benchmarks disagree sharply here. Kimi-K3 crushes exams but posts a low MathTutorBench Scaffold rate; Claude-Opus-5 is the reverse. Don’t pick a tutor model on exam scores alone.
•
Worth testing but not shown: whether the 20-template approach transfers to non-K-12 domains (code review, medical explanation) where similar behavior-conflation problems exist.
•
No ablation isolates which pipeline stage drives the tutoring gains. The whole recipe works; the decomposition is uncertain.
•
Several benchmarks (K12-Bench, MathFish, EDUMATH) come from prior work by overlapping communities and are Chinese-K-12-centric, so improvements may partly reflect curriculum-locale match rather than a universal tutoring ability.
•
Knowledge-state diagnosis accuracy on LongTutor tops out at 54% even for the best model, so “understanding where the learner is” remains hard.
•
Some open-weight baselines score 21.95% on EXAMS-V (the random-guess floor for the format), suggesting they simply can’t follow the answer format, which inflates OmniEdu’s apparent lead against those specific comparators.
•
Base models are all Qwen3.5/3.8 family; the recipe’s transfer to Llama, Mistral, or Gemma bases is not tested.