Hugging Face has introduced OmniEdu, an open family of foundation models aimed at K-12 learning and teaching. According to the paper, educational foundation models must not only solve problems but also understand curriculum structure, diagnose learner difficulties, and deliver appropriate instructional support — a combination the authors say existing educational language models rarely cover together.
The project's instruction-tuning corpus draws on more than 100 educational resources alongside general instruction sources, and is organized around four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding. Rather than grouping training data by source or task, the pipeline is built around these capabilities.
The data pipeline combines deterministic cleaning, semantic auditing and rewriting, task-specific quality scoring, token-budgeted diversity selection, and pedagogical instruction assignment. It produces 69,999 examples and 15.96M supervised response tokens, of which 60,951 examples are education-specific.
The team fine-tuned models at 4B, 9B, and 27B scales and evaluated them on curriculum grounding, K-12 problem solving, and pedagogical tutoring, alongside general capability. Education-oriented tuning improved all three educational benchmark groups across model scales, the paper reports.
OmniEdu-27B reached 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 recorded the highest Teaching average on LongTutor among the evaluated models, at 3.02. The authors frame these results as evidence for curated, capability-balanced supervision when adapting general language models to educational tasks.