metadata
license: mit
base_model: microsoft/Phi-3.5-vision-instruct
tags:
- OCR
pipeline_tag: image-text-to-text
TB-OCR: an end-to-end OCR model handling text, math latex, and markdown formats all at once
Model Summary
TB-OCR-preview (Text Block OCR), created by Yifei Hu, is an end-to-end OCR model handling text, math latex, and markdown formats all at once. The model takes a block of text as the input and returns clean markdown output. Headers are marked with ##. Math expressions are guaranteed to be wrapped in brackets \( inline math \) \[ display math \] for easier parsing. This model does not require line-detection or math formula detection.
Running the model in 4-bit only requires ~2.8GB VRAM to load and exhibits little to none degradation.
This model is recommended to work with TFT-ID-1.0[HF], a text/tale/figure detection model, for full page document parsing.
