Hugging Face's logo

Datasets:
VAMPSBenchmark
/
VAMPS

Tasks:
Question Answering
Visual Question Answering
Modalities:
Image
Text
Formats:
parquet
Languages:
English
Persian
Size:
1K - 10K
Tags:
benchmark
multimodal
visual-reasoning
mathematical-reasoning
graph-reasoning
tool-use
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
VAMPS / README.md
VAMPSBenchmark's picture
VAMPSBenchmark
Update README.md
0320fd3 verified
preview code
|
Raw
History Blame
5.91 kB
metadata 
dataset_info:
  features:
    - name: sample_id
      dtype: int64
    - name: qnum
      dtype: int64
    - name: question_text
      dtype: string
    - name: question_image
      dtype: image
    - name: option_1_text
      dtype: string
    - name: option_1_image
      dtype: image
    - name: option_2_text
      dtype: string
    - name: option_2_image
      dtype: image
    - name: option_3_text
      dtype: string
    - name: option_3_image
      dtype: image
    - name: option_4_text
      dtype: string
    - name: option_4_image
      dtype: image
    - name: answer
      dtype: string
    - name: subset
      dtype: string
    - name: language
      dtype: string
    - name: source
      dtype: string
  splits:
    - name: Konkour_EN
      num_bytes: 838629
      num_examples: 218
    - name: Konkour_FA
      num_bytes: 983043
      num_examples: 218
    - name: Synth_EN
      num_bytes: 10581059
      num_examples: 366
    - name: Synth_FA
      num_bytes: 10591587
      num_examples: 366
  download_size: 22894723
  dataset_size: 22994318
configs:
  - config_name: default
    data_files:
      - split: Konkour_EN
        path: data/Konkour_EN-*
      - split: Konkour_FA
        path: data/Konkour_FA-*
      - split: Synth_EN
        path: data/Synth_EN-*
      - split: Synth_FA
        path: data/Synth_FA-*
language:
  - en
  - fa
pretty_name: VAMPS
tags:
  - benchmark
  - multimodal
  - visual-reasoning
  - mathematical-reasoning
  - graph-reasoning
  - tool-use
  - desmos
  - multiple-choice
  - persian
  - english
task_categories:
  - question-answering
  - visual-question-answering
size_categories:
  - 1K<n<10K
license: cc-by-nc-4.0

VAMPS: Visual-Assisted Mathematical Problem Solving Benchmark

Code License: CC BY-NC 4.0

VAMPS (Visual-Assisted Mathematical Problem Solving) is a bilingual, multimodal benchmark for evaluating whether vision-language models can benefit from constructing and interpreting graphs while solving mathematical problems.

The benchmark targets a capability gap that is increasingly important for real scientific and engineering workflows: models may solve a problem analytically, but fail when they must externalize the problem through a visualization tool, inspect the resulting graph, and ground the final answer in visual evidence.

Dataset Summary

VAMPS contains 1,168 multiple-choice QA instances across English and Persian. The questions are drawn from Iranian University Entrance Exam (Konkour) algebra/calculus problems and expanded with human-reviewed LLM-generated synthetic variants. Each item is selected so that plotting can provide a natural solution strategy, for example through intersections, extrema, asymptotes, monotonicity, inverse functions, or graph ordering.

The dataset is designed for both benchmarking and diagnosis:

  • Benchmarking: measure accuracy on graph-assisted mathematical multiple-choice questions.
  • Tool-use evaluation: test whether a model can construct useful graphs with a plotting tool such as Desmos.
  • Visual grounding diagnosis: compare direct analytical solving against tool-enabled visual solving.
  • Bilingual evaluation: compare behavior across English and Persian versions of real and synthetic problems.

Dataset Structure

VAMPS is organized into four splits:

Split Language Source Rows
Konkour_EN English Original Konkour seed questions 218
Konkour_FA Persian Original Konkour seed questions 218
Synth_EN English Human-reviewed synthetic variants 366
Synth_FA Persian Human-reviewed synthetic variants 366
Total English + Persian Real + synthetic 1,168

Each row is a multiple-choice problem with optional images for the question and/or answer options.

Fields

Field Type Description
sample_id integer Row-level sample identifier within the split.
qnum integer Question number / problem identifier.
question_text string Main problem statement. May contain LaTeX-style math.
question_image image or null Optional image associated with the problem statement.
option_i_text string Text for answer option i. Empty if the option is image-only.
option_i_image image or null Optional image for answer option i.
answer string / integer Correct answer label in {1, 2, 3, 4}.
subset string Split/subset name.
language string Language code: en or fa.
source string Source/provenance label, e.g. real Konkour or synthetic.

Data Creation and Motivation

The real seed questions come from Iranian University Entrance Exam algebra and calculus problems. The synthetic portion was generated to preserve graph-mediated mathematical structure while increasing coverage and diversity. Synthetic variants were human-reviewed before inclusion.

The benchmark emphasizes problems where a graph can naturally reveal relevant mathematical structure, such as:

  • intersections of functions,
  • extrema and turning points,
  • monotonicity and concavity,
  • asymptotic behavior,
  • inverse functions,
  • ordering or comparison of curves,
  • discontinuities and piecewise behavior.

Evaluation

The companion repository contains a minimal reproduction package for direct no-tool evaluation, Desmos-based tool-agent evaluation, etc.

Links

Licensing

The dataset is released under (CC BY-NC 4.0) License.

VAMPS is intended for evaluation use only. It is not intended and shall not be used for training models, etc.