OmniEdu: Open Foundation Models for Learning and Teaching

OmniEdu. Open Foundation Models for Learning and Teaching. By Hao Liang and seven others, from Peking University, the University of the Chinese Academy of Sciences, and one other institution. Published on the nineteenth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

OmniEdu is an open family of foundation models for learning and teaching at the K-12 level, released together with the instruction-tuning corpus behind it. The starting point is a list of what an educational model has to do. It must solve problems, understand how a curriculum is structured, diagnose where a learner is struggling, and give the right kind of instructional support. Existing educational language models tend to focus on one side, either problem solving or tutoring, and their training mixtures are organized by source or by task rather than by capability.

The corpus is organized around four capabilities instead. Subject competence. Curriculum grounding. Diagnostic reasoning. And pedagogical action and scaffolding. It contains sixty-nine thousand nine hundred and ninety-nine examples, with fifteen point nine six million supervised response tokens. Of those examples, sixty thousand nine hundred and fifty-one are specific to education. The models are fine-tuned at three sizes: four billion, nine billion, and twenty-seven billion parameters.

The material comes from over one hundred educational resources, combined with general instruction sources. It passes through a pipeline of five stages: deterministic cleaning, then semantic auditing and rewriting, then quality scoring specific to each task, then diversity selection under a token budget, and finally the assignment of pedagogical instructions.

The resource is for anyone adapting general language models to education. The evaluation covers curriculum grounding, K-12 problem solving and pedagogical tutoring, with general capability tracked alongside. Tuning for education improves all three educational benchmark groups, consistently, at every model scale.

The largest model, OmniEdu-27B, reaches sixty-three point one two percent exact match and seventy-six point six nine percent F1 on K12-Bench. It scores eighty-five point eight nine percent on MathFish, and eighty-six point nine five percent on EDUMATH. In the Scaffold setting of MathTutorBench, it reaches seventy-eight point seven four percent. On LongTutor it achieves the highest Teaching average among the evaluated models, three point zero two.

What distinguishes this corpus is its axis of organization. Rather than sorting data by where it came from, the authors sorted it by what it teaches a model to do, and balanced it across those capabilities. They read the consistent gains, across sizes and across problem solving, curriculum and tutoring, as a sign that capability-balanced supervision is what adapts a general model to educational work.

OmniEdu. Open Foundation Models for Learning and Teaching. By Hao Liang and seven others, from Peking University, the University of the Chinese Academy of Sciences, and one other institution. Published on the nineteenth of September. This is the long version: the whole paper rather than the abstract.

A model that answers a school maths question correctly has done only part of a teacher's job. The authors list the rest: connect the problem to the right knowledge point and its prerequisites, find the misconception behind a wrong attempt, and choose an intervention that moves the learner forward. In their words, a model has to coordinate what to teach, where the learner is, and how to respond. Existing open educational models tend to cover one piece. Some are built for exam-style problem solving, others for tutoring dialogue or curriculum alignment, and they are usually evaluated in isolation. The claim here is that the gap is partly a data-design problem. Mixing training data by source or subject hides what response each example is meant to teach, and lets easy or redundant examples crowd the mixture.

OmniEdu organizes its education data around four capabilities rather than around sources. Subject competence is solving problems and explaining answers. Curriculum grounding is knowing grade level, difficulty, prerequisites and where a problem sits in a curriculum. Diagnostic reasoning is spotting errors and misconceptions in a student's work and inferring what prerequisite is missing. Pedagogical action and scaffolding is picking an intervention, such as a hint, a question, a prerequisite review or a direct explanation, and carrying it out. Every example gets exactly one primary capability. The second idea borrows from LearnLM: pedagogical instruction following. A problem-solving answer should give and justify the answer, while a scaffolding reply may deliberately withhold it.

The corpus comes from more than a hundred datasets and educational resources, about one point three four million education examples before filtering. Six stages cut that down. The first stage is the taxonomy itself. The second is deterministic cleaning: standardize formats, drop exact duplicates and malformed records, check answers against references where they can be verified, discard unrelated content such as finance, confirm that images exist and line up with the text, exclude data with unclear licences, and remove anything overlapping the evaluation set. Eight hundred and seventy thousand survive.

The third stage is a semantic audit by a large Qwen model. Each example gets a usability score out of a hundred and one of three actions. High scorers are kept, low scorers are removed, and the middle band is rewritten: the input is held fixed, only the flawed answer or dialogue is repaired, and the result is audited again and kept only if it now passes. Criteria depend on the capability. Tutoring examples, for instance, are checked for premature leakage of the final answer. Four hundred and forty thousand remain.

The fourth stage first shrinks a few bloated sources. Rather than sampling at random, it runs k-center greedy selection over text embeddings to keep a semantically diverse subset. RACE drops from about sixty thousand examples to five thousand. A second model then scores each example on task-specific rubric dimensions from one to five, and keeps it only if every dimension reaches three and critical ones like correctness reach four. The fifth stage runs the same diversity selection inside fine-grained task buckets, but budgets each bucket in supervised response tokens rather than examples, because one long tutoring dialogue can carry an order of magnitude more tokens than a hint. That leaves sixty thousand nine hundred and fifty-one examples. The sixth stage attaches one of twenty system instruction templates, chosen by bucket, leaving the content untouched, and adds nine thousand and forty-eight general examples. Three backbones, at four, nine and twenty-seven billion parameters, are then fully fine-tuned for three epochs.

The comparison that tests the thesis is each tuned model against its own base, and every scale improves on every educational benchmark group. The gains are largest in tutoring. On the MathTutorBench Scaffold setting, the nine billion model's win rate rises from fourteen percent to seventy-five point three. On LongTutor, which asks a tutor to use a student's history, its Evidence score climbs from five point eight to sixty-six point six. Problem solving moves least, because the bases were already strong: the largest model goes from ninety-one point six to ninety-four point nine on GAOKAO-Bench. General ability does not collapse; MMMU-Pro and GPQA both rise slightly at every scale.

The appendix tables show things the averages hide. On MathFish, the curriculum-alignment gain comes from saying yes more readily: the largest model's accuracy on true alignments jumps by about twenty-two points, while its accuracy at rejecting neighbouring standards falls by about nine. Against frontier systems, Kimi-K3 leads EXAMS-V by nearly eighteen points. And knowledge-state diagnosis, which the corpus explicitly targets, tops out near fifty-four percent, which the authors themselves call comparatively challenging.

The paper has no limitations section, so here is what its own evidence does not cover. Each result is a single number with no error bars or repeated runs. The tuned models also get worse in places the text passes over. On GAOKAO-Bench, the four billion model's Chinese score falls from about ninety to fifty-nine percent. The largest model's diagnosis F1 on LongTutor drops below its base. Several open educational baselines score an identical twenty-one point nine five percent across every EXAMS-V column, a pattern the paper does not explain. Five of the open baselines, and GLM-5.3, cannot take images and saw only the text of image-based questions. Finally, the twenty-seven billion backbone comes from a newer Qwen release than the two smaller ones, so comparisons across scale mix size with backbone version.

The idea to carry away is that an educational model's training data should be labelled by the behaviour it teaches, not by where it came from. OmniEdu separates what an example contains from how the model should respond to it, tags each with one capability and one explicit instruction, and budgets the mixture in tokens, because a tutoring dialogue and a hint differ so much in length. The payoff shows up most in scaffolding and in using a student's history, far more than in raw exam scores.