README.md

---
license: apache-2.0
task_categories:
- question-answering
language:
- ru
pretty_name: T-math
size_categories:
- n<1K
dataset_info:
  features:
  - name: question
    dtype: string
  - name: verifiable_answer
    dtype: string
  - name: year
    dtype: string
  - name: grade
    dtype: string
  - name: full_answer
    dtype: string
  - name: solutions
    list: string
  - name: task_complexity
    dtype: string
  - name: olympiad
    dtype: string
  splits:
  - name: train
    num_bytes: 510955
    num_examples: 331
  download_size: 228445
  dataset_size: 510955
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---


# 🧮 T-Math
**T-Math** is a dataset of Russian math olympiad problems created to assess the reasoning capabilities of large language models (LLMs) in mathematics.  
It includes 331 problems from the [All-Russian School Olympiad](https://vos.olimpiada.ru/) and the [Moscow Olympiad](https://mos.olimpiada.ru) for high school students, covering the period from 1998 to 2025.  
The tasks and their ground-truth answers were extracted automatically and subsequently verified by human assessors.

Key features:
- Challenging problems that require multi-step reasoning (median completion length for Qwen3-32B is 16K tokens), sourced from top-tier Russian olympiads
- Easily verifiable: answers are numeric-only and checked using the `math_verify` library to compare mathematical expressions
- Not yet saturated, even by frontier reasoning models such as Gemini 2.5 Pro and DeepSeek R1
- Contains 331 samples — the largest Russian math olympiad-level benchmark — making it more statistically robust compared to smaller datasets like the 30-sample AIME benchmark


## 📊 Evaluation Results

|Model|pass@1|
|--|--|
|o4-mini-high|**0.73**|
|DeepSeek-R1-0528|<ins>0.71</ins>|
|Gemini-2.5-Pro|0.70|
|Claude Sonnet 4|0.56|
|T-pro-it-2.0|0.54|
|Qwen3-32B|0.53|


## 🗂️ Filtering procedure

The text was extracted from PDFs using [Qwen/Qwen2.5-VL-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct). Tasks, along with their ground-truth and verifiable (numeric) answers, were extracted via LLM calls.
We filtered out invalid questions using an LLM based on the following criteria:
- Tasks requiring multiple answers
- Tasks without a single correct answer
- Theorem-like tasks where the main goal is proving a statement, making automatic verification non-trivial
- Tasks with non-numeric answers, to simplify answer comparison
- Tasks that cannot be solved without access to an accompanying image

Next, we removed tasks of moderate difficulty where Qwen3-8B achieved a 100% pass@16 rate, as they offer limited value for benchmarking reasoning.
Finally, both the questions and the verifiable answers were manually reviewed by assessors to ensure consistency with the original sources.


## 🛠️ How to use
Add the following system prompt to guide the model to return the final answer in a \boxed{} tag, making it easier to parse:
```
Решите следующую математическую задачу эффективно и ясно. Последняя строка вашего ответа должна иметь следующий формат:
'Таким образом, окончательный ответ: $\boxed{ОТВЕТ}$.' (без кавычек), где ОТВЕТ - это просто окончательное число или выражение, решающее задачу.
Думайте шаг за шагом перед ответом.
```

You can then use the following code snippet with the math_verify library to compare mathematical expressions:
```python
from math_verify import LatexExtractionConfig, parse, verify
from latex2sympy2_extended import NormalizationConfig


def accuracy_reward(completion: str, solution: str) -> float:
    """Reward function that checks if the completion matches the ground truth."""
    # parse the gold solution (assumed to always succeed)
    gold_parsed = parse(solution, extraction_mode="first_match")

    # parse the model’s completion with the same LaTeX extraction settings
    answer_parsed = parse(
        completion,
        extraction_config=[
            LatexExtractionConfig(
                normalization_config=NormalizationConfig(
                    nits=False,
                    malformed_operators=False,
                    basic_latex=True,
                    equations=True,
                    boxed="all",
                    units=True,
                )
            )
        ],
        extraction_mode="first_match",
    )

    # verify and return binary reward; on error, print and give 0.0
    try:
        return float(verify(gold_parsed, answer_parsed))
    except Exception as e:
        print(f"verify failed: {e}, answer: {answer_parsed}, gold: {gold_parsed}")
        return 0.0


```

## 📖 Citation

If you find our work useful in your research, please consider citing the following paper:

```bibtex
@inproceedings{stoianov-etal-2026-pro,
    title = "{T}-pro 2.0: An Efficient {R}ussian Hybrid-Reasoning Model and Playground",
    author = "Stoianov, Dmitrii  and
      Taranets, Danil  and
      Tsymboi, Olga  and
      Latypov, Ramil  and
      Dautov, Almaz  and
      Kruglikov, Vladislav  and
      Nikita, Surkov  and
      Abramov, German  and
      Gein, Pavel  and
      Abulkhanov, Dmitry  and
      Gashkov, Mikhail  and
      Zelenkovskiy, Viktor  and
      Batalov, Artem  and
      Medvedev, Aleksandr  and
      Potapov, Anatolii",
    editor = "Croce, Danilo  and
      Leidner, Jochen  and
      Moosavi, Nafise Sadat",
    booktitle = "Proceedings of the 19th Conference of the {E}uropean Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 3: System Demonstrations)",
    month = mar,
    year = "2026",
    address = "Rabat, Marocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.eacl-demo.22/",
    doi = "10.18653/v1/2026.eacl-demo.22",
    pages = "297--319",
    ISBN = "979-8-89176-382-1"
   }
```