Datasets:

DFKI-SLT
/

kangaroo_math_benchmark

dataset_info:
  features:
    - name: image
      dtype: image
    - name: question
      dtype: string
    - name: ground_truth
      dtype: string
    - name: is_text_only
      dtype: bool
    - name: year
      dtype: int32
    - name: class
      dtype: string
    - name: task_id
      dtype: string
    - name: math_category
      dtype: string
    - name: difficulty
      dtype: string
  splits:
    - name: train
      num_bytes: 158221288
      num_examples: 3557
  download_size: 152844456
  dataset_size: 158221288
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - visual-question-answering
language:
  - de
tags:
  - Känguruh-Wettberwerb
  - math
  - vlm
  - school

Känguruh Wettbewerb Dataset

This dataset is a German benchmark from the official Känguru der Mathematik competition materials covering 1998--2025 (Link). Each instance is a single multiple-choice problem with five options (A - E) from school grade levels 3 - 13. The benchmark targets curriculum-aligned mathematical reasoning in German with explicit visual grounding (diagrams, geometric figures, spatial arrangements) while enabling controlled analyzes by year, grade group, difficulty, math subdomain, and modality.

Processing

Collected PDFs are converted into benchmark-instances via the following four-step pipeline:

1. Task localization and cropping
1. Manual screening
1. Solution alignment
1. Structured fields and derived labels

Schema and Labels

Each instance contains:

image: PNG crop
answer: ground-truth option
year: task year
grade_level: targetet school year of task
task_id: task id
difficulty: diffulty classes A, B, C
math_category: one of arithmetic, geometry, algebra, probability, unknown
is_text_only: bool

Dataset Overview

Statistic	Value
Years	1998 - 2025
Grade levels	3 - 13 (grouped)
Source PDFs / pages	140 / 560
Answer format	5-way MC (A - E)
Total tasks N	3,557
Text-only / visually necessary	61% / 39%
Difficulty (A/B/C)	32.9 / 33.3 / 33.8%

Cite as

@inproceedings{hug2026kangaroobench,
  author    = {Hug, Dennis and Wolf, Ren\'e and Thomas, Philippe},
  title     = {KangarooBench: A Native German Benchmark for Evaluating the Modality Gap in Visual Mathematics},
  booktitle = {Conference and Labs of the Evaluation Forum (CLEF 2026)},
  year      = {2026},
  note      = {To appear}
}