Strike-OCR

Strike-OCR is an advanced OCR and document understanding AI model developed by HSR-DeepThink under HSR-Projects.

Built using DeepSeek OCR and optimized for local AI workflows, Strike-OCR is designed for accurate handwritten text recognition, printed document extraction, markdown conversion, and structured OCR tasks.

Features

• Complex handwritten text recognition
• Printed document OCR
• Markdown output formatting
• Multi-language OCR support
• Table and form extraction
• Mathematical text recognition
• Structured document parsing
• Low hallucination OCR pipeline
• Optimized for local deployment

Base Model

Strike-OCR is built using:

• DeepSeek OCR
• Ollama

Platform Support

Strike-OCR supports:

• Linux
• Windows
• macOS

Requirements

• Ollama installed
• GPU recommended for best performance
• Minimum 8GB RAM recommended

Installation

Pull the base OCR model:

ollama pull deepseek-ocr

Create the custom model:

ollama create strike-ocr -f Modelfile

Run Strike-OCR:

ollama run strike-ocr

Example Usage

Basic OCR:

ollama run strike-ocr "Extract text from this image"

Document understanding:

ollama run strike-ocr "Convert this handwritten page into markdown"

Recommended Parameters

Recommended Modelfile settings:

PARAMETER temperature 0.1
PARAMETER top_p 0.9
PARAMETER num_ctx 16384

These settings improve:

• OCR stability
• Handwriting accuracy
• Long document understanding

Example Modelfile

FROM deepseek-ocr

SYSTEM """
You are Strike-OCR, an advanced OCR and handwriting recognition AI developed by HSR-DeepThink under HSR-Projects.

Capabilities:
- Handwritten OCR
- Printed OCR
- Table extraction
- Markdown conversion
- Multi-language support
- Complex document parsing

Rules:
- Never hallucinate missing text
- Preserve formatting
- Preserve equations
- Preserve tables
- Output clean markdown when useful
"""

PARAMETER temperature 0.1
PARAMETER top_p 0.9
PARAMETER num_ctx 16384

Recommended Use Cases

• Dataset generation
• Label Studio workflows
• Document digitization
• Research projects
• Offline OCR systems
• AI annotation pipelines
• Handwritten note extraction

Performance Notes

Strike-OCR performs best when:

• Images are clear and high resolution
• Handwriting is properly visible
• Documents are not heavily compressed

GPU acceleration is highly recommended for faster inference.

License

Apache 2.0 License

Author

Developed by HSR-DeepThink

Under HSR-Projects