Evidence
Benchmarks, with the receipts
A number without its configuration is not evidence. This page states what every figure published here has to carry before it goes up.
No benchmark table is published here yet.
Rather than post figures that cannot be reproduced from this page, the peer-reviewed results below are linked directly, and the disclosure standard a future table has to meet is stated underneath. If you need comparative numbers now, run the evaluation yourself against your own documents — that is the measurement that should decide it anyway.
Standard
What every published figure must carry
- Dataset
- Which documents, at which revision, and where to obtain them.
- Docling version
- The exact release that produced the numbers.
- Configuration
- Pipeline, OCR engine, enrichments and export settings.
- Hardware
- CPU, GPU or accelerator, and memory available to the run.
- Quality metric
- What was measured, and how it was scored.
- Throughput
- Pages or documents per second, and the concurrency used.
- Baselines
- What Docling was compared against, at which versions.
- Limitations
- Where the configuration is known to do badly.
- Script
- A command that reproduces the table from scratch.
Published work
All papers
The peer-reviewed results behind the models
DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation
The layout dataset the detection models are trained and evaluated against.
TableFormer: Table Structure Understanding with Transformers
The table-structure model behind row, column and header recovery.
Docling Technical Report
Pipeline architecture and the measurements published with it.
SmolDocling: An Ultra-Compact Vision-Language Model for End-to-End Multi-Modal Document Conversion
Compact end-to-end conversion model released alongside Docling.