Nexora OCR v0.1-2B

Nexora OCR v0.1-2B

Status: Experimental Base Model: Qwen3.5 2B Task: OCR / Text Extraction Version: v0.1-2B

Nexora OCR v0.1-2B is an experimental OCR model designed for text extraction from images. It is fine-tuned from Alibaba's Qwen3.5 2B and trained to analyze document and scene-text images while returning accurate transcriptions.

Nexora OCR v0.1-2B is an experimental larger variant of the Nexora OCR v0.1 (0.8B) model, exploring whether increased model capacity improves OCR performance on the same task and dataset family.

The model was trained on a custom OCR instruction dataset prepared for Nexora OCR, analogous to the 0.8B variant.


Table of Contents


Overview

Nexora OCR v0.1-2B is an experimental OCR model designed for text extraction from images.

The model receives an image together with an instruction and generates a textual transcription of the visible text content.

Nexora OCR v0.1-2B was fine-tuned from Qwen3.5 2B using a supervised fine-tuning workflow, following the same general approach as the 0.8B variant.

This release focuses on exploring how a larger foundation model affects OCR metrics such as character accuracy, CER, exact match, and WER on the same benchmark setup.


The Nexora OCR Family

Nexora OCR is an experimental model series focused on practical and efficient OCR models for document understanding, text extraction, and scene-text recognition.

The project explores the use of compact vision-language foundation models for specialized OCR workloads, including document digitization, receipt and form parsing, handwritten text recognition experiments, and lightweight OCR pipelines.

Nexora OCR v0.1 (0.8B) and Nexora OCR v0.1-2B are the initial public releases of the project.

Future releases may improve character accuracy, word error rate, exact-match performance, robustness to low-quality images, and inference efficiency across different model sizes.


What's New

Nexora OCR v0.1-2B

Nexora OCR v0.1-2B introduces a larger, experimental variant of the Nexora OCR model family with a focus on:

  • Using Qwen3.5 2B as the base instead of 0.8B
  • Maintaining the same OCR task and instruction format as v0.1 (0.8B)
  • Providing a direct benchmark comparison between 0.8B and 2B on the same test set

Specific dataset composition, augmentation strategies, and training hyperparameters are not publicly disclosed at this time, consistent with the 0.8B release.


Model Details

Property Details
Model Name Nexora OCR v0.1-2B
Model Family Nexora OCR
Model Type Multimodal Image-Text Model (OCR)
Base Model Qwen/Qwen3.5-2B
Fine-tuning Base Qwen/Qwen3.5-2B
Task Optical Character Recognition (OCR)
Input Image + Text Instruction
Output Transcribed Text
Fine-tuning Method Supervised Fine-Tuning (SFT)
Training Framework Transformers / Custom SFT
Release Status Experimental
Version v0.1-2B

Capabilities

Nexora OCR v0.1-2B is designed to:

  • Extract printed and handwritten text from images
  • Process document pages, forms, receipts, and scene-text images
  • Handle multi-line and multi-word text regions
  • Preserve basic layout and line breaks where possible
  • Support prototype OCR and document-understanding pipelines
  • Work with natural-language instructions (e.g., "Transcribe all text in this image")
  • Run on hardware capable of hosting a 2B-parameter multimodal model

Example output:

INVOICE
Date: 2025-12-10
Total: $42.50

The exact output format may vary depending on the prompt and image content. Applications should normalize, post-process, and validate the transcribed text before using it programmatically.


Limitations

Nexora OCR v0.1-2B is an experimental model and has important limitations.

All limitations described for Nexora OCR v0.1 (0.8B) also apply to the 2B variant, including benchmark interpretation, image quality sensitivity, language and script coverage, layout handling, output consistency, confidence metrics, dataset limitations, generalization, and production-readiness considerations.

In early evaluations, the 2B variant does not uniformly outperform the 0.8B model on all OCR metrics; in particular, word-level error rates can be higher despite similar character-level performance.


Intended Use

Supported Use Cases

Nexora OCR v0.1-2B is intended for:

  • AI and OCR research
  • Document digitization experiments
  • Prototype OCR systems
  • Educational experimentation
  • Benchmarking compact vision-language models for OCR at different scales
  • Development of text-extraction pipelines
  • Research into lightweight OCR inference with larger backbones
  • Testing OCR workflows on common document types
  • Exploring structured and free-form text outputs from multimodal models
  • Researching model robustness against image quality variations

Out-of-Scope Use Cases

The model is not intended for:

  • Fully automated transcription without additional safeguards
  • Safety-critical or high-stakes document processing
  • Applications requiring guaranteed character-level accuracy
  • Systems where incorrect transcriptions could cause significant harm
  • Automatic enforcement actions without human review or secondary validation
  • Medical, legal, financial, or law-enforcement document processing without strong additional controls

Developers should combine model predictions with:

  • Policy rules and validation checks
  • Uncertainty handling and threshold tuning
  • Human review for uncertain or critical cases
  • Periodic evaluation on representative deployment data

Architecture & Training

Nexora OCR v0.1-2B is based on Qwen3.5 2B, a larger multimodal vision-language model designed to process image and text inputs.

Training Configuration

Property Details
Base Model Qwen/Qwen3.5-2B
Fine-tuning Base Qwen/Qwen3.5-2B
Fine-tuning Method Supervised Fine-Tuning
Training Framework Transformers / Custom SFT
Task Optical Character Recognition (OCR)
Input Modality Image + Text Instruction
Output Transcribed Text
Training Dataset Custom Nexora OCR instruction dataset
Model Version v0.1-2B

Nexora OCR v0.1-2B was fine-tuned from Qwen/Qwen3.5-2B.

A supervised fine-tuning workflow was used with a custom multimodal instruction dataset prepared for Nexora OCR, analogous to the dataset used for the 0.8B variant.


Output Format

The intended output is plain text containing the transcribed content, for example:

RECEIPT
Store: Example Mart
Date: 2025-11-03
Items:
- Coffee $3.50
- Sandwich $6.00
Total: $9.50

Because the model generates free-form text, responses may occasionally include additional explanation, formatting differences, or minor errors.

A safer integration should:

  • Extract the generated response
  • Apply text normalization (trimming, whitespace cleanup)
  • Optionally parse into structured fields if needed
  • Apply fallback behavior when output is clearly malformed
  • Route uncertain or critical cases for additional review

Usage

The recommended way to use Nexora OCR v0.1-2B is via the nexora-ai Python package.

Installation

pip install nexora-ai

Basic Example

from nexora.ocr import NexoraOCR

ocr = NexoraOCR(model="nexora-ocr-v0.1-2b")

result = ocr.read("image.png")

print(result)

result is a string containing the transcribed text. You can further process it (e.g., split into lines, parse fields, or store in a database) depending on your application.

Notes

  • Ensure the image path is correct and the file is readable.
  • For best results, use clear, well-lit images with minimal distortion.
  • The model is experimental; validate outputs before using them in critical workflows.
  • The 2B variant requires more GPU memory and compute than the 0.8B variant.

Evaluation

Nexora OCR v0.1-2B should be evaluated using both regression testing and independent held-out evaluation.

Internal Regression Benchmark (0.8B vs 2B)

The following table compares Nexora OCR v0.1 (0.8B) and Nexora OCR v0.1-2B on the same internal benchmark.

Metric 0.8B 2B
Test samples 100 100
Character accuracy 80.77% 79.66%
CER 19.23% 20.34%
Exact match 27.00% 24.00%
WER 43.49% 69.50%

These results are intended for release-to-release comparison and should not be interpreted as independent estimates of real-world performance across all document types and languages.

Observations from this benchmark:

  • Character accuracy and CER are similar between 0.8B and 2B, with the 0.8B variant slightly ahead.
  • Exact match is modest for both, with the 0.8B variant performing slightly better.
  • Word Error Rate (WER) is substantially higher for the 2B variant, indicating more word-level deviations despite comparable character-level performance.

Future releases should be evaluated using:

  • Larger and more diverse held-out datasets
  • Per-document-type metrics (receipts, forms, IDs, scene text, etc.)
  • Additional metrics such as precision/recall for key fields, line-level accuracy, and calibration error where applicable

Risks & Considerations

Developers integrating Nexora OCR v0.1-2B should account for the possibility of character-level transcription errors, word-level misrecognitions, and insertion or deletion errors. Performance may be affected by image quality, compression, font style, text size, layout complexity, lighting, blur, glare, and partially occluded text. The model may also perform differently on unseen image types or scripts, and its output formatting may not always be consistent. Dataset bias toward certain document types or languages and sensitivity to prompt wording may also affect results.

The model should not be used as the sole transcription mechanism in an application. Robust deployments should include appropriate validation and rule-based checks, explicit transcription policies and thresholds, logging and monitoring of error patterns, safe handling of malformed or unexpected outputs, and periodic evaluation on representative deployment data. Human review should be considered for uncertain or critical cases.


Community & Support

Nexora OCR is an experimental project.

Community feedback, testing, and benchmark contributions are welcome.

For information about future releases, visit the Nexora OCR model family on Hugging Face.


License

Nexora OCR v0.1-2B is based on Qwen3.5 2B.

The model is released under the Apache 2.0 license, subject to the applicable terms and conditions of the underlying Qwen model.


Acknowledgements

Nexora OCR v0.1-2B builds upon Qwen3.5 2B by the Qwen team at Alibaba.

We thank the Qwen team for their open foundation models and the broader open-source AI and machine-learning community for the tools, research, and resources that make projects like Nexora OCR possible.


About Nexora

Nexora OCR is an experimental OCR model project focused on developing practical and efficient multimodal AI systems for document understanding and text extraction.

Nexora OCR v0.1 (0.8B) and Nexora OCR v0.1-2B represent the first public releases of the project and establish baselines for future improvements in character accuracy, word error rate, exact-match performance, and robustness across different model scales.


Image Credits

The background image used in the Nexora OCR v0.1-2B model artwork was photographed by J Blueberry and is sourced from Pixabay.

Downloads last month
10
Safetensors
Model size
2B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ArkAiLab-Adl/nexora-ocr-v0.1-2b

Finetuned
Qwen/Qwen3.5-2B
Finetuned
(349)
this model
Quantizations
2 models