SoMark open-sources OXR-1.0, a 1.3B-parameter document parser

SoMark: SoMark develops AI models specialized in document parsing and understanding. It recently open-sourced OXR-1.0, introducing a modular system that separates layout detection from content recognition and reading order. The organization focuses on enabling flexible, customizable parsing for varied document types without human labeling in post-training. OXR-1.0: OXR-1.0 is a document parsing model that converts PDFs and page images into Markdown or JSON formats with structured outputs. It uses a modular design where a small layout model identifies blocks and the main model handles reading and ordering, allowing teams to integrate custom layout components. The model incorporates a recursive self-improvement loop that automatically generates training data from its own errors using a general vision language model as judge and labeler. Model Design: Layout detection is handled separately by a small model, enabling the main parser to focus on content recognition and reading order while supporting easy swaps of layout components for new document types. Training Method: A recursive self-improvement loop uses a general vision language model to identify model weaknesses, generate targeted new training examples, and cluster errors for focused retraining without human labeling. Output Capabilities: The model identifies and structures 13 element types including tables, mathematical formulas in LaTeX, code blocks, stamps, and chemical structures along with page coordinates.