Problem
A learning product that shows a ready solution works like a collection of homework answers: the child copies it and learns nothing. Yandex Uchebnik, a Russian school-education service, wanted an AI math assistant for grades 4-11 that teaches instead. Training data is scarce: textbooks cannot be digitized for copyright reasons, and real children's dialogues are few. Children also write differently from what adults imagine. A child may ask unethical questions or write about a personal problem, so the dialogue needs control.
Example from the article: on the problem "7 chestnuts, 3 times more poplars, how many trees in total", the tutor corrects a student's 24 poplars with guiding questions instead of printing 28.
Solution in brief
The assistant is split into a problem solver (YandexGPT 5 lite), a separate dialogue engine (YandexGPT 5 pro) and a control layer of classifiers and simple agents. The dialogue engine always asks guiding questions in a Socratic style. Its training set mixes real children's dialogues, teacher-rewritten dialogues and moderated synthetic ones.
Questions this case answers
- How to build a language model dialogue that guides a learner to the answer with questions instead of giving the solution
- How to control a conversational model used by minors with separate safety and help-request classifiers and a solver model
- How to build an LLM tutor that guides school students through a math problem instead of handing over the answer
- How to collect training dialogues for a children's tutor when real child dialogues are scarce
Tools and methods
- Problem solver — YandexGPT 5 lite, step-by-step solutions
- Dialogue engine — YandexGPT 5 pro, leads the student with questions
- Ethics classifiers — check the student's input and the model's output
- Call-for-help classifier — spots messages unrelated to the task
- Agents — a basic calculator and a theory-and-terms block
Google or a one-time email code. No password.
