SandLogicTechnologies/MiniCPM-OCR-Jev-GGUF

MiniCPM-OCR Jev

MiniCPM-OCR Jev is a compact multimodal OCR quality-assessment model developed by OpenBMB for evaluating whether OCR results correctly represent the content of document regions. Instead of generating OCR text itself, the model assesses an existing OCR candidate and produces a score indicating how likely the candidate is to be correct.

This repository provides GGUF quantized variants of MiniCPM-OCR Jev for efficient local deployment. The available quantization formats reduce the storage footprint of the original model while preserving its image-and-text based OCR verification workflow.

MiniCPM-OCR Jev is designed for evaluating text, formula, and table regions. It can be used to automatically accept reliable OCR results, identify regions that require human review, or compare outputs produced by multiple OCR systems and select the strongest candidate.

---

Model Overview

• Model Name: MiniCPM-OCR Jev
• Base Model: openbmb/MiniCPM-OCR-Jev
• Architecture: Multimodal OCR Correctness Decision Model
• Parameter Count: Approximately 2 Billion Parameters
• Input: Region Image, OCR Candidate Text, Rendered Candidate Image
• Output: OCR quality score between 0 and 1
• Base Architecture: Qwen3.5-2B with retained vision and language components
• Decision Head: Binary classification head replacing the language-model output head
• Supported Region Types: Text, Formula, Table
• Developer: OpenBMB
• License: Apache 2.0

MiniCPM-OCR Jev evaluates an OCR candidate by comparing the original document region with the candidate content. The model uses the original region image, candidate OCR text, and a rendered representation of that text to determine whether the candidate is consistent with the source image.

---

Quantization Formats

This repository provides GGUF quantized versions of MiniCPM-OCR Jev for reduced-storage local inference. The available formats provide different trade-offs between model size, memory requirements, and numerical precision.

IQ3_M

• Size reduction of approx 71.13% (1010.37 MB) compared to 16-bit (3.52 GB)
• Aggressive 3-bit quantization providing the smallest representation among the supplied conversions
• Suitable for deployments where minimizing model storage and memory consumption is the primary requirement
• Makes the 2B OCR decision model practical for resource-constrained local verification pipelines
• The reduced precision may have a greater effect on fine-grained OCR quality assessment, particularly for visually complex text, formulas, and tables

IQ4_NL

• Size reduction of approx 67.33% (1.15 GB) compared to 16-bit (3.52 GB)
• Non-linear 4-bit quantization providing a higher-precision representation than IQ3_M
• Suitable for OCR verification workflows involving text, formulas, tables, and document-region quality assessment
• Maintains a substantially smaller model representation while providing additional numerical precision over the 3-bit variant
• Requires more storage than IQ3_M but provides a balanced option for applications where assessment quality is more important than minimum file size

IQ4_XS

• Size reduction of approx 68.18% (1.12 GB) compared to 16-bit (3.52 GB)
• Compact 4-bit representation with a smaller file size than the IQ4_NL conversion
• Suitable for local OCR quality assessment, candidate ranking, and document verification workflows
• Provides another 4-bit deployment option when storage requirements need to remain below the IQ4_NL variant
• Retains the model's multimodal decision architecture while using a more compact stored representation

Q6_K

• Size reduction of approx 58.81% (1.45 GB) compared to 16-bit (3.52 GB)
• Higher-precision 6-bit K-Quant representation among the supplied GGUF variants
• Suitable for applications where additional numerical precision is preferred for OCR correctness scoring
• Provides a larger representation than the IQ3_M and IQ4 variants while remaining considerably smaller than the F16 model
• A strong option for OCR verification workloads where model fidelity is prioritized over minimum storage requirements

---

Training Background (Original Model)

MiniCPM-OCR Jev is trained from Qwen3.5-2B and adapts its multimodal architecture specifically for OCR correctness assessment. Rather than using the standard language-model output head for text generation, the model replaces the lm_head with a binary classification head that learns to determine whether an OCR candidate is correct.

OCR Quality Assessment

• Evaluates individual document regions containing text, formulas, or tables
• Uses the original region image as the visual reference for assessment
• Processes the candidate OCR content together with a rendered representation of that candidate
• Produces a score between 0 and 1 representing the estimated correctness of the OCR result
• Can evaluate results produced by different OCR parsing systems

Candidate Verification

• Supports automatic screening of OCR results based on application-defined thresholds
• Can identify candidates that should be forwarded for human review
• Allows several OCR systems to process the same region independently
• Can compare candidate scores and select the highest-scoring result for a region
• Keeps the assessment focused on OCR correctness rather than generating replacement text

Model Adaptation

• The original vision encoder and language backbone from the base architecture are retained
• The language-model output head is replaced with a binary classification head
• Fine-tuning can be performed for document-specific OCR assessment scenarios
• Training data can contain original region images, candidate OCR content, rendered candidate images, and correctness labels

---

Key Capabilities

• OCR Correctness Assessment
  Evaluates whether an existing OCR result accurately corresponds to the content visible in the original document region.

• Text Region Verification
  Scores OCR candidates for standard textual regions and helps distinguish reliable recognition from potentially incorrect results.

• Formula Verification
  Assesses OCR results containing mathematical formulas and can help identify candidates requiring additional validation.

• Table Verification
  Evaluates OCR candidates for table regions where structural and textual consistency are important.

• Best-of-N Candidate Selection
  Compares multiple OCR candidates generated for the same region and can be used to select the candidate receiving the strongest quality score.

• Human Review Routing
  Supports threshold-based workflows in which high-confidence OCR results are accepted automatically while uncertain regions are sent for manual inspection.

• Cross-Parser Evaluation
  Can assess outputs generated by different OCR systems, allowing a downstream pipeline to compare and rank candidates using a common quality-assessment model.

---

Usage Example

Using llama.cpp

MiniCPM-OCR Jev can be deployed through GGUF-compatible inference infrastructure with multimodal decision-model support.

For Jev scoring, the inference input consists of the original document-region image together with the OCR candidate information. The candidate can be represented using the OCR text and its rendered image. The model then produces a consistency decision or probability indicating how likely the OCR candidate is to correctly represent the original region.

./llama-cli \
  -m SandlogicTechnologies/MiniCPM-OCR-Jev-IQ4_NL.gguf \
  --mmproj mmproj-model-f16.gguf \
  -p "Describe the important visual information in this input."

The model is designed as a decision model rather than a conventional text-generation model. Its primary output is the OCR correctness assessment, not newly generated OCR text.

---

Recommended Usecases

• OCR Quality Control Automatically assess OCR results before passing them into downstream document-processing systems.

• Document Processing Pipelines Add a verification stage after OCR parsing to identify regions where recognition quality may be unreliable.

• Human-in-the-Loop OCR Automatically accept high-confidence results while routing uncertain regions to human reviewers.

• Multi-Model OCR Selection Run multiple OCR systems on the same document region and use Jev scores to select the strongest candidate.

• Formula Recognition Validation Evaluate extracted mathematical expressions and identify formula regions that require additional verification.

• Table Recognition Validation Assess OCR candidates generated from table regions before structured data extraction or downstream processing.

• OCR Benchmarking Compare OCR outputs from different parsing systems using a common correctness-assessment model.

• Resource-Constrained Deployment Use the GGUF quantized variants when the original F16 representation is larger than the available storage or memory budget.

---

Acknowledgments

These GGUF quantized models are derived from the original MiniCPM-OCR Jev release by the OpenBMB development team.

Special thanks to:

• The OpenBMB team for developing MiniCPM-OCR Jev and the broader MiniCPM-OCR document-understanding ecosystem.

• The Qwen team for the Qwen3.5-2B foundation used by MiniCPM-OCR Jev, together with the open-source multimodal and OCR communities contributing to its surrounding tooling and evaluation ecosystem.

• The llama.cpp community for supporting efficient GGUF-based local inference and multimodal decision-model deployment.

---

Contact

For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/.