SGOCR drops spatially-grounded OCR pipeline and dataset

The move toward "spatially-grounded" OCR is a crucial pivot for VLM reliability, shifting focus from vague reasoning to precise visual localization.

• –Uses a multi-stage distillation stack including Nvidia's Nemotron OCR v2, Gemma 4, and Qwen 3-VL for high-fidelity text extraction and anchoring.
• –Implementation of an agentic optimization loop based on Karpathy's "autoresearch" allows for holistic dataset quality sweeps and automated code changes.
• –Explicit spatial metadata including text polygons and anchor boxes provides the necessary training signal for models to "see" exactly where text resides in complex scenes.
• –The "rescue" loop logic effectively targets weak-coverage images, ensuring high data density without manual intervention.
• –Proves that "less is more" by using Gemini 2.5 Flash as a teacher model, which the author found more effective for grounding tasks than heavier models like Gemini 3.1 Pro or ChatGPT 5.3 Codex.