arch-math-multimodal-dataset

A collection of math- and chart-related OCR/vision datasets, combined for training multimodal math models. Seven sources, roughly 25–30GB total, covering handwritten math OCR, printed math OCR, chart QA/instruction data, and chart/graph images.

Repo structure

arch-math-multimodal-dataset/
├── mathwriting/
│   └── png/
│       ├── test/      (individual PNGs, browsable)
│       ├── valid/     (individual PNGs, browsable)
│       └── train.tar  (~6.9GB, packaged — extract to browse)
├── chartqa/
├── chartinstruction/
│   ├── benchmark/
│   ├── filtered_data_138k/
│   └── generated_data/
├── handwritten-formula-ocr/
│   ├── train-00000-of-00001.parquet
│   └── test-00000-of-00001.parquet
├── math-collections-ocr/
├── deepmath-ocr-100000/
│   ├── images/
│   │   ├── train/     (~9,962 files)
│   │   └── test/      (~986 files)
│   ├── train.parquet
│   ├── test.parquet
│   └── *.csv
└── graphs-dataset/
    ├── shard_1/  (4,000 files)
    ├── shard_2/  (4,000 files)
    ├── shard_3/  (4,000 files)
    └── shard_4/  (3,875 files)

Sources

1. mathwriting/

• Origin: MathWriting 2024
• Format: InkML (handwritten stroke/pen-trajectory data) + rendered PNG images
• Structure: test/ and valid/ are individual PNGs, browsable in the HF dataset viewer; train/ is a single png/train.tar (~6.9GB, ~230K files) — not extracted in this repo for size reasons, so pull and extract locally if you need per-file access
• Content: Handwritten math expressions paired with ground-truth LaTeX, plus raw stroke/ink data (InkML) useful for stroke-sequence or online-handwriting models, not just static images
• License: Not specified by source (no license field on the original HF page)

2. chartqa/

• Origin: ChartQA benchmark
• Format: Chart images + question–answer pairs (JSON)
• Content: Question-answering over chart images — bar/line/pie charts with grounded numeric/factual questions
• License: Not specified by source (no license field on the original HF page)

3. chartinstruction/

• Origin: lewy666/ChartInstructionData — English-language version
• Format: 8,538 files across three folders
  • benchmark/ — evaluation set
  • filtered_data_138k/ — ~138K filtered instruction-following examples (largest subfolder, ~2GB)
  • generated_data/ — synthetically generated chart-instruction pairs (~375MB)
• Content: Chart instruction-following / chart-to-text data for training models to describe, summarize, or answer open-ended questions about charts
• License: Not specified by source (no license field on the original HF page)

4. handwritten-formula-ocr/

• Origin: hieudt0803/ocr-math-formula-handwritten
• Format: 2 parquet files — train-00000-of-00001.parquet (343MB) and test-00000-of-00001.parquet (38MB)
• Content: Handwritten math formula images embedded in parquet, paired with OCR/LaTeX labels
• License: Not specified by source (no license field on the original HF page)

5. math-collections-ocr/

• Origin: zrchen03/math-collections-ocr
• Format: Scanned page images, organized by sub-collection
• Content: Three Chinese-language math problem sets at different levels:
  • Intermediate Mathematics 1983–2013 (中等数学1983-2013) — multi-decade problem archive
  • Primary School Math Olympiad (小学奥数) — elementary olympiad problems
  • Classic Works (经典著作) — classic math texts
• License: Not specified by source (no license field on the original HF page)

6. deepmath-ocr-100000/

• Origin: minlik/deepmath-ocr-100000
• Format: 10,951 files (4.7GB) — images/train/ (9,962 images), images/test/ (~986 images), train.parquet, test.parquet, plus CSVs (deepmath_latex_train.csv and variants)
• CSV schema: row_id, field, latex — field values include question (the OCR'd problem) and r1_solution_1 / r1_solution_2 / r1_solution_3 (worked-solution text in LaTeX), so each row_id maps to a question image plus multiple reasoning/solution traces rather than a single flat label
• Content: Math OCR pairs with attached step-by-step solution text — useful for OCR and downstream reasoning tasks
• License: Not specified by source (no license field on the original HF page)

7. graphs-dataset/

• Origin: Kaggle — sunedition/graphs-dataset
• Format: ~15,875 images, sharded into shard_1–shard_4 (4,000 / 4,000 / 4,000 / 3,875 files) for HF dataset-viewer browsability
• Content: ~16K clean images of graphs/plots; source dataset description and any embedded text is Russian-language
• License: CC0-1.0

Notes

• mathwriting/png/train.tar is not extracted, unlike test/ and valid/ which are browsable PNG folders — extract locally (tar -xf train.tar) if you need per-file streaming/viewer access to training images.
• graphs-dataset/ was sharded into 4 subfolders of ≤4,000 files each; chartinstruction/ and deepmath-ocr-100000/ were uploaded as-is since their largest single folders stayed under the ~9–10K file threshold that triggers HF dataset-viewer/upload issues.
• Redundant zip files removed: chartinstruction/benchmark.zip, filtered_data_138k.zip, and generated_data.zip were deleted after confirming the extracted folder versions were already present and correct — only the extracted, browsable folders remain.
• Licenses: only graphs-dataset has a stated license (CC0-1.0). The other six sources have no license field on their original HF pages — none of them ship a dataset card. That means usage terms default to whatever each source's own README/paper implies (MathWriting's paper describes Google-internal ink collection but doesn't state a redistribution license) or, absent that, standard copyright applies. Confirm terms directly with each source's author before any external or commercial release of this combined repo.
• Maintainers: built and maintained by Samrachana-1 and Sid, collecting open-source math/chart datasets for a shared training pipeline.