Russian State Exam: Pick the Right Option is a machine learning competition focused on classifying multiple-choice questions from Russia's national school exams, known as ЕГЭ (for university admission at ages 17–18) and ОГЭ (for basic school completion at ages 15–16).
The core challenge lies in the uniqueness of the data. While the exam tasks are drawn from an open bank published by ФИПИ (the state institute that writes the exams), the official answer keys have never been released to the public. Unlike typical benchmarks where answers may leak into training corpora, the labels for this competition were generated specifically for the event by submitting every possible option to the official ФИПИ checker until the correct one was found. This ensures the labels could not have existed in any prior dataset, textbook, or crib sheet.
The Task
Participants must predict the single correct option among 3–6 choices for each task. The problem is a plain multi-class text classification across sixteen subject-exam combinations. These include social studies, biology, geography, physics, mathematics, informatics, and reading comprehension in five foreign languages (English, German, French, Spanish, and Chinese). Some tasks include a shared source passage in the context column.
Data Composition The dataset reflects the actual distribution of the open bank: - ОГЭ social studies is the largest block by a wide margin. - The five foreign languages combined form the second-largest group. - Physics, mathematics, and informatics represent thin tails with only a few dozen tasks each.
Participants can also utilize the public "Russian State Exams" dataset (174,627 tasks with solutions and rubrics) for pre-training or retrieval, though the competition labels themselves come exclusively from the official checker.
Evaluation and Rules Submissions are evaluated based on Categorization Accuracy, which measures the share of tasks where the predicted option matches the correct one. Every task counts equally regardless of subject or the number of options. There is no private leaderboard; the public score reflects performance on the entire test set.
This is a community competition with no prizes and no medals. A strict rule applies: participants should solve the tasks using their models rather than mapping test items back to the bank to query the server directly, as doing so would render the leaderboard meaningless.