MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale

2026-04-06

AI Summary

This work addresses a critical limitation in current document parsing methods, which exhibit consistent failure patterns on challenging samples—suggesting that performance bottlenecks stem from deficiencies in training data rather than model architecture. To overcome this, the authors propose a data-centric parsing paradigm that retains a fixed 1.2B-parameter model while constructing a large-scale training set characterized by high coverage, rich information content, and precise annotations. They introduce a three-stage progressive training strategy: large-scale pretraining, hard-sample fine-tuning, and GRPO-based alignment. Key innovations include diversity- and difficulty-aware sampling, cross-model consistency validation, and an iterative render-and-verify labeling pipeline. Evaluated on OmniDocBench v1.6, the approach achieves a score of 95.69, surpassing the same-architecture baseline by 2.71 points and outperforming the previous state-of-the-art method despite using a model with over 200 times fewer parameters.

Abstract

Current document parsing methods compete primarily on model architecture innovation, while systematic engineering of training data remains underexplored. Yet SOTA models of different architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than architecture itself. Building on this finding, we present \minerupro, which advances the state of the art solely through data engineering and training strategy optimization while keeping the 1.2B-parameter architecture of \mineru completely fixed. At its core is a Data Engine co-designed around coverage, informativeness, and annotation accuracy: Diversity-and-Difficulty-Aware Sampling expands training data from under 10M to 65.5M samples while correcting distribution shift; Cross-Model Consistency Verification leverages output agreement among heterogeneous models to assess sample difficulty and generate reliable annotations; the Judge-and-Refine pipeline improves annotation quality for hard samples through render-then-verify iterative correction. A three-stage progressive training strategy -- large-scale pre-training, hard sample fine-tuning, and GRPO alignment -- sequentially exploits these data at different quality tiers. On the evaluation front, we fix element-matching biases in OmniDocBench~v1.5 and introduce a Hard subset, establishing the more discriminative OmniDocBench~v1.6 protocol. Without any architectural modification, \minerupro achieves 95.69 on OmniDocBench~v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods including models with over 200$\times$ more parameters.

Problem

• document parsing
• training data deficiency
• data-centric learning
• hard samples
• performance bottleneck

Innovation

• data-centric learning
• document parsing
• hard sample mining
• cross-model consistency verification
• progressive training strategy