FireRed-OCR Technical Report

TL;DR

This paper introduces FireRed-OCR, a novel framework designed to transform general-purpose VLMs (based on Qwen3-VL) into pixel-precise structural document parsing experts and proposes a Three-Stage Progressive Training strategy that guides the model from pixel-level perception to logical structure generation.

Abstract

We present FireRed-OCR, a systematic framework to specialize general VLMs into high-performance OCR models. Large Vision-Language Models (VLMs) have demonstrated impressive general capabilities but frequently suffer from ``structural hallucination'' when processing complex documents, limiting their utility in industrial OCR applications. In this paper, we introduce FireRed-OCR, a novel framework designed to transform general-purpose VLMs (based on Qwen3-VL) into pixel-precise structural document parsing experts. To address the scarcity of high-quality structured data, we construct a ``Geometry + Semantics'' Data Factory. Unlike traditional random sampling, our pipeline leverages geometric feature clustering and multi-dimensional tagging to synthesize and curate a highly balanced dataset, effectively handling long-tail layouts and rare document types. Furthermore, we propose a Three-Stage Progressive Training strategy that guides the model from pixel-level perception to logical structure generation. This curriculum includes: (1) Multi-task Pre-alignment to ground the model's understanding of document structure; (2) Specialized SFT for standardizing full-image Markdown output; and (3) Format-Constrained Group Relative Policy Optimization (GRPO), which utilizes reinforcement learning to enforce strict syntactic validity and structural integrity (e.g., table closure, formula syntax). Extensive evaluations on OmniDocBench v1.5 demonstrate that FireRed-OCR achieves state-of-the-art performance with an overall score of 92.94\%, significantly outperforming strong baselines such as DeepSeek-OCR 2 and OCRVerse across text, formula, table, and reading order metrics. We open-source our code and model weights to facilitate the ``General VLM to Specialized Structural Expert'' paradigm.

Figures (9)

• Figure 1: Performance comparison on the OmniDocBench v1.5 benchmark. FireRed-OCR achieves state-of-the-art performance, securing the top rank in the overall evaluation with a score exceeding 90%.
• Figure 2: Comparison with state-of-the-art models on OmniDocBench v1.5 ouyang2025omnidocbench. FireRed-OCR ranks first in the overall metric (top) and demonstrates leading performance across all sub-categories: Text recognition, Formula processing, Table structure analysis, and Reading order inference (bottom).
• Figure 3: Overview of the proposed data processing and optimization pipeline. This pipeline operates through five rigorous stages, including Geometry-Driven Feature Extraction and Dual Indexing, Stratified Sampling and Unified Re-annotation, Synthetic Data Generation for Structural Priors, Automated Quality Control and Hard-Negative Mining and Expert-Level Refinement via Model Distillation.
• Figure 4: Overview of the FireRed-OCR training framework. The pipeline progresses through three stages: (1) Multi-task Pre-alignment, which employs joint training on detection, region OCR, and layout tasks to ground the model’s understanding of document structure; (2) Specialized Supervised Fine-Tuning (SFT), designed to adapt base capabilities for end-to-end structured Markdown generation; and (3) Format-Constrained GRPO, a reinforcement learning stage incorporating rule-based rewards (targeting formula syntax, hierarchical closure, and table integrity) to enforce structural compliance and textual accuracy.
• Figure 5: From Pixels to Formula: The model accurately identifies complex structures such as $\lim_{\Delta x\to0}$ and fraction hierarchies, enabling direct digitization of STEM materials.