Mahal Systems MathTutorBench
Fits math tutoring agents that must teach rather than just answer. Borrow the three-skill split (math expertise, student understanding, pedagogical ability) and the small trained reward model that scores a teacher turn by win rate against the real teacher's turn.
What it asks the agent to do
- solve a math word problem
- ask a Socratic question instead of giving the answer
- judge whether a student's solution is correct and locate the mistake
- write the next teacher turn that scaffolds the student
- follow a stated pedagogy strategy over a dialog
What to borrow from it
grading designtaxonomybaseline
Caveats
Dialog turns are scripted from existing datasets, not a live student; scaffolding scores depend on a 1.5B reward model trained by the authors; Socratic questioning is scored by BLEU; the README gives conflicting licenses (CC-BY-4.0 badge, CC-BY-SA-4.0 text).
Details
| Workloads | education, knowledge qa |
|---|---|
| Interaction | multi turn scripted |
| Environment | static |
| End state checked | No |
| Tools | none |
| Grading | exact match, programmatic check, llm judge |
| Size | Not published |
| Data | mixed |
| Personal data | none |
| License | CC-BY-4.0 |
| Harness | own repo |
| Status | active |