DeepData/Datasets/Multimodal Math
Focus line

University-level mathematics where the image carries information the text does not: function graphs, geometric figures, diagrams and plots. Each problem ships with its figure and a full solution.

Scale today1M+ problems
Target10M
ModalityImage + text
FormatJSONL + PNG
01Why this data

What makes it worth training on.

Expert-level difficulty

The figure is required to solve the problem, so text-only shortcuts fail.

Graduate & competition math: 1M+ problems →
Scale that tracks compute

1M+ today, scaling to 10M.

Undergraduate STEM: 40M+ today, 100M target →
Training-ready on arrival

Questions reference their figures inline; images ship alongside the JSONL.

100% Markdown/LaTeX · schema-checked JSONL →
02Sample

Inspect the data.

Real items from this line, shown exactly as delivered. Switch to JSON for the raw record, or download the public sample file.

Samples24 of 100
MathematicsUndergraduateShort answerimage

Use the two graphs to find and . Does the two-sided limit exist? The marked open circles are excluded values.
Problem figure

The one-sided limits are and , respectively. The two-sided limit does not exist.

From the inner graph, as and as . The outer graph has left limit and right limit at . The isolated values and do not determine these limits.

03Coverage

How it is sliced.

Visual content

Function graphs, geometric constructions, region diagrams, plots and charts.

Levels

Undergraduate and above.

Alignment

Each question references its image explicitly; solutions reason over what the figure shows.

Also available

1M+ K–12 photo-based grading image-text pairs for answer-evaluation models.

04Schema

Record format.

FieldTypeDescription
idstringStable unique identifier.
subjectstringDiscipline, e.g. Mathematics, Physics, Chemistry.
levelstringUndergraduate · Graduate · Competition.
qtypestringProof · Calculation · Short answer · Single / multiple choice · Fill-in-the-blank · True / false.
questionstringProblem statement in Markdown with LaTeX.
question → <img>image refInline reference to a PNG figure shipped with the JSONL.
optionsarray | nullAnswer options for choice questions.
answerstringFinal answer, verifiable where the question type allows.
solutionstringFull step-by-step reasoning.
has_imageboolTrue for every item in this line.
05Use

Built for.

  • Multimodal reasoningTrain vision-language models to read a figure and reason from it.
  • Image-text alignmentProblems where text and image must be understood together.
  • EvaluationVisual math sets where text-only shortcuts fail.
Contact

Request a pilot slice.

Tell us the multimodal math slice you need. We will send a pilot sample cut to your specification for your own evaluation.