paper web signal

OmniEdu opens 4B-27B K-12 solver and tutor models on one recipe

TL;DR

  • OmniEdu releases open K-12 educational foundation models at 4B, 9B and 27B parameters, tuned on a single capability-balanced recipe.
  • OmniEdu-27B reports 63.12% EM and 76.69% F1 on K12-Bench, plus 85.89% on MathFish and 86.95% on EDUMATH.
  • The instruction-tuning corpus is 69,999 examples and 15.96M response tokens, of which 60,951 are education-specific.

The OmniEdu paper trains three K-12 educational models at 4B, 9B and 27B parameters on a single recipe, and reports the 27B variant reaching 63.12% exact-match and 76.69% F1 on K12-Bench, 85.89% on MathFish, and 86.95% on EDUMATH.

The framing pitches the split in prior work as the problem. "Existing educational language models often focus on either problem solving or tutoring, with training mixtures organized by source or task rather than capability," the authors write, and they describe OmniEdu as "an open family of foundation models for K-12 learning and teaching."

The instruction-tuning corpus lands at 69,999 examples and 15.96M supervised response tokens, of which 60,951 examples are education-specific, drawn from "over 100 educational resources and general instruction sources." On the tutoring side, OmniEdu-27B posts 78.74% in MathTutorBench's Scaffold setting and 3.02 on LongTutor's Teaching average, which the abstract calls the highest among the evaluated models.

The abstract does not list the license, the languages covered, or the closed-model baselines its numbers should be read against.