AURORA Data
Specialised font and glyph datasets for document AI
Access to font, glyph and character-fingerprint data. You can review the sample format, labelling and coverage; licensing and access scope are agreed per project.
AURORA ML Training Build better document AI with specialised data and training
We built our own font and glyph data and training pipeline for AURORA OCR. Now we make this expertise available to teams that develop their own document models.
What we offer
AURORA Training
Train or fine-tune a model for your document task
Characters, text lines, fonts or page layout. Hard-example mining, targeted retraining and simulated scanner and camera damage are built in.
AURORA Evaluation
See how your model performs on document data
Our internal test sets and selected public benchmarks, results by category and a report with the method. No score claims we cannot back up.
The data behind the models
380,849unique fontsEvery font in the collection is unique: exact copies are kept only once, while different versions of a font count separately.
24.0 Mcharacter fingerprintsA compact description of one character's shape in one font, used to find fonts and look-alike letters.
≈ 13.8 Mvalidated glyphs11.7 M letters, 1.9 M digits and 0.2 M other signs, each labelled from a trusted font source, checked by a person or confirmed by several independent readers.
1,643characters in our newest modelLatin, Cyrillic and Greek letters with accents, digits and signs; letters of one shape share a class, so the model has 1,593 classes.
Two different counts. ≈ 13.8 M is our whole collection of checked glyphs. The model measured on the OCR page was trained on ≈ 1.1 M glyph samples chosen from our data, each validated by a person.
How a project works
- Define the taskDocument types, scripts, characters and the output you need.
- Select dataThe data scope and coverage that fit the task.
- Train or fine-tuneTargeted training with hard-example mining.
- EvaluateHeld-out test data and relevant public benchmarks.
- Review resultsFindings and next steps together.
The unified API and SDK platform is in development. Early-access requests are open.
Building a document AI model?
Tell us what you are training, which documents you work with and where your model struggles.