cernis-intelligence/cernis-vision-ocr

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Sep 12, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

CernisOCR is a 7 billion parameter vision language model developed by cernis-intelligence, fine-tuned on Qwen2.5-VL-7B-Instruct with a context length of 32768 tokens. This model unifies mathematical LaTeX conversion, handwritten text transcription, and structured document extraction from invoices and receipts into a single, efficient solution. It is optimized for diverse OCR tasks across multiple document domains, handling varied image types, layouts, and text styles.

Loading preview...

Overview

CernisOCR is a 7 billion parameter vision language model (VLM) developed by cernis-intelligence, specifically fine-tuned on Qwen2.5-VL-7B-Instruct. Unlike traditional domain-specific OCR models, CernisOCR integrates three distinct OCR functionalities into a single, efficient model, making it highly versatile for various document processing needs. It is robust to noise and variable image quality, extracting both raw text and structured information.

Key Capabilities

  • Mathematical LaTeX Conversion: Converts handwritten or printed mathematical formulas into LaTeX notation.
  • Handwritten Text Transcription: Accurately transcribes both cursive and printed handwriting.
  • Structured Document Extraction: Extracts structured data from complex documents like invoices and receipts.
  • Multi-domain Processing: Handles diverse image types, layouts, and text styles within a unified framework.

Training Details

CernisOCR was fine-tuned using LoRA (Low-Rank Adaptation) on a dataset of 10,000 samples, comprising nearly 4,000 samples each for LaTeX OCR and handwritten text, and over 2,000 for invoices and receipts. This targeted training reduced the loss from 4.802 to 0.116, representing a 97.6% improvement. The model leverages the Qwen2.5-VL-7B-Instruct base model and was efficiently trained using Unsloth.

Good For

  • Applications requiring unified OCR across different document types.
  • Workflows involving mathematical formula recognition and LaTeX conversion.
  • Transcribing handwritten notes or documents.
  • Automating data extraction from financial documents like invoices and receipts.

Finetunes
1 models