Logics-MLLM/Logics-Parsing
Logics-Parsing is a 7 billion parameter Vision-Language Model (VLM) developed by Logics-MLLM, specifically fine-tuned for end-to-end document parsing. It excels at accurately analyzing and structuring complex documents, including those with intricate layouts and scientific content like chemical formulas. The model generates rich, structured HTML output, preserving logical structure and identifying content blocks with categories, bounding boxes, and OCR text.
Loading preview...
Overview
Logics-Parsing is a 7 billion parameter Vision-Language Model (VLM) developed by Logics-MLLM, designed for end-to-end document parsing. It processes document images directly into structured output, eliminating the need for multi-stage pipelines. The model is built upon a general VLM through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), making it highly effective for complex document analysis.
Key Capabilities
- End-to-End Processing: A single-model architecture simplifies deployment and inference, directly converting document images to structured output.
- Advanced Content Recognition: Accurately recognizes and structures difficult content, including intricate scientific formulas and chemical structures (represented in SMILES format).
- Rich, Structured HTML Output: Generates clean HTML, preserving logical structure with each content block tagged by category, bounding box coordinates, and OCR text. It automatically filters irrelevant elements like headers and footers.
- State-of-the-Art Performance: Achieves leading performance on an in-house benchmark specifically designed for complex-layout documents and STEM content, outperforming various pipeline tools and general VLMs.
Good for
- Parsing documents with challenging layouts and diverse content.
- Extracting and structuring scientific formulas and chemical structures.
- Generating logically structured HTML from document images.
- Applications requiring robust and accurate document understanding.