Carbon Research Labs

Carbon Research Labs builds Carbon-OCR, an OCR engine in C++ and CUDA, and Fullerene, its family of small models trained from scratch. Together they read scans, photos of documents and PDFs into text, markdown, HTML and JSON with boxes, in reading order, with labeled blocks and table cells.

Models

fullerene-3M is the Latin set: Latin, Cyrillic and Greek. It is four networks with 2.77 M parameters in all:

• page orientation;
• line detection, with a second pass for tiny text;
• line recognition;
• reading order, block labels and table cells.

The weights ship as FP32 and FP16 ONNX. On all 2,937 pages of OmniParseBench, fullerene-3M scores a headline of 43.5, where Tesseract 5.5 scores 5.9. On one CPU core it reads 5.4 times as many pages a second.

A set is named by its parameter count, rounded to the nearest million. More script groups follow.

Data

fullerene-data records the training data source by source: where each came from, its license, and every file's checksum. It holds the manifests and the scripts that fetched them, not the third-party data itself.

Engine

Carbon-OCR runs on CPUs, with its own int8 kernels or with ONNX Runtime, one worker per core. On NVIDIA GPUs it runs through TensorRT, with many pages in flight. Its output formats are text, markdown, HTML, JSON, hOCR, ALTO, PAGE XML and searchable PDF.

Licenses

The code is Apache 2.0. The Fullerene weights are under the Carbon-OCR OpenRAIL-AMS.