OCRia-1
A 4-billion-parameter vision-language model for document parsing, with a focus on complex tables, charts, and mathematical notation.
- ParseBench overall
- 87.4 /100
- Model parameters
- 4B
Research
Open evaluation
Benchmarks
Document parsing
ParseBench
Tests tables, charts, text, formatting, and visual grounding using human-verified pages and deterministic evaluation rules.
Structured extraction
ExtractBench
Tests schema-guided extraction across 370 enterprise documents, measuring value accuracy, record completeness, source grounding, and cost.
ExtractBench extension
ExtractBench Pro
Extends ExtractBench with six long-tail extraction challenges. A separate synthetic dataset tests conventions, domain rules, drawn relations, and complex charts.
These benchmarks are built and maintained by LlamaIndex. We evaluate our own systems alongside others and publish the datasets and evaluation code for inspection and reproduction.
Publication
OCRia-1
Architecture, training, and evaluation of our 4B document parsing model.
ExtractBench Pro: The Long Tail of Document Extraction
Six challenges in the long tail of extraction, and how we built the data to test them.
ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction
Evaluating value accuracy, completeness, grounding, and measured cost.
ParseBench: A Document Parsing Benchmark for AI Agents
Evaluating document parsing across tables, charts, content, formatting, and grounding.
4 selected publications
Open resources
Code and datasets
Run the evaluations, inspect the data, or build with our open-source document tools.