| # Data provenance |
|
|
| ## Source dataset |
|
|
| The model was trained only on formula crops derived from **The IBEM Dataset: a |
| large printed scientific image dataset for indexing and searching mathematical |
| expressions**, version 1.0: |
|
|
| - Official record: <https://zenodo.org/records/7963703> |
| - DOI: <https://doi.org/10.5281/zenodo.7963703> |
| - License: Creative Commons Attribution 4.0 |
| - Authors: Dan Anitei, Joan Andreu Sánchez, and José Miguel Benedí |
|
|
| IBEM contains 600 scientific documents and 8,272 page images, with annotations |
| for isolated and embedded mathematical expressions. The original images are |
| machine-typeset page images, not handwriting. Their fonts and styles are not |
| uniform because they come from different papers. |
|
|
| ## Preparation used for this model |
|
|
| - Formula images are grayscale bounding-box crops from the original IBEM page |
| JPEGs with a deterministic two-pixel crop margin. |
| - Images were not rerendered into one Typst font. |
| - Labels were derived deterministically from IBEM `latex_norm` annotations and |
| converted to package-free Typst with semantic spacing canonicalization. |
| - Splits are document-disjoint. |
| - Training: 129,303 crops from 480 documents. |
| - Validation: 16,899 crops from 60 different documents. |
| - Dataset lock SHA-256: |
| `10dae9db6eab4de9d36ee28ca99760bc0c6231fab77311004452dee719f64271`. |
| - The calibration and test documents were not opened for this checkpoint. |
|
|
| No Fusion dataset, synthetic formula corpus, teacher-generated labels, |
| external pretrained model, or fused samples were used in this model lineage. |
|
|
| ## What is included |
|
|
| No image or annotation from the raw IBEM archive is redistributed here. The |
| end-to-end smoke-test image at `example/formula.png` is an original |
| project-authored formula rendered with Typst; it is covered by the project MIT |
| License and is not training or evaluation data. |
|
|
