TeleOCR

TeleOCR, formerly released as NaviDC-OCR, is a vision-language model built specifically for document understanding, unifying two scenarios that earlier approaches usually handle separately: digital documents and photos of documents taken with a camera. It introduces Multi-node Consensus Voting for generating pseudo-labels automatically, geometry-aware document modeling and Curvature-Guided Douglas-Peucker Sampling for camera-captured documents, image-to-image self-verification for refining its own training data, a progressive four-stage training pipeline, and content-structure decoupled learning for tables and formulas.

Its automated data engine combines these consensus-voting, geometry-aware synthesis, and self-verification components to produce training data without heavy manual labeling. The project reports evaluating the model, under its earlier name, against a document-parsing competition, where it scored better than two other named document-parsing models. Community contributions include a GGUF conversion for llama.cpp and reports of deployment on Ascend 910B hardware.