Скачать или смотреть GutenOCR: Precise Grounded VLMs for Documents

In this AI Research Roundup episode, Alex discusses the paper: 'GutenOCR: A Grounded Vision-Language Front-End for Documents' GutenOCR introduces a family of grounded vision-language models designed to solve the lack of alignment between text and pixels in current document models. By fine-tuning Qwen2.5-VL backbones, the researchers enable precise line- and paragraph-level geometry for more reliable data extraction. The model unifies four tasks including full-page reading, detection, localized transcription, and answering spatial queries. Trained on diverse datasets, it provides a stable interface for complex retrieval-augmented generation and document understanding workflows. Paper URL: https://arxiv.org/abs/2601.14490 #AI #MachineLearning #DeepLearning #VLM #OCR #DocumentParsing #ComputerVision

Resources:
GitHub: https://github.com/Roots-Automation/G...
Hugging Face model: https://huggingface.co/rootsautomatio...
Hugging Face model 2: https://huggingface.co/rootsautomatio...