BizOnAI-OCR

A Korean-optimized OCR model for industrial document processing developed by ONTHEIT.

BizOnAI-OCR is fine-tuned specifically for Korean documents — contracts, financial forms, medical records, government paperwork, and more. It handles Korean-specific layouts, mixed Korean/English/Chinese text, decorative spacing, vertical tables, and complex forms that trip up general-purpose OCR models.

We also release KDoc-OCRBench, the first comprehensive Korean document OCR benchmark with 14,738 test cases — designed to fill the gap in standardized Korean OCR evaluation.

Evaluated on KDoc-OCRBench — 14,738 unit-test-style assertions across 804 Korean PDFs in 7 industrial document categories (contracts, financial forms, medical records, logistics, education, government, presentations). Tests check text presence/absence, table cell relationships, and header/footer handling. See the KDoc-OCRBench section below for details.

Evaluated with --table-normalize-spaces --table-normalize-newlines.

Evaluated on KDoc-OCRBench-V2 — 57,062 manually-reviewed unit tests across 850 Korean public-sector PDFs (statistics, reports, manuals, notices). V2 is a separate, complementary benchmark with much heavier table coverage (49,127 table cell tests, ×ばつ v1) and openly-shippable source PDFs (KOGL Type 1). See kdoc_benchmark_v2/ for evaluation instructions.

Evaluated with upstream olmocr.bench.benchmark (plug-compatible).

BizOnAI-OCR is Korean-optimized but also remains competitive on English documents. Evaluated on olmOCR-bench by Allen AI.

Model weights are hosted on HuggingFace: ONTHEIT/BizOnAI-OCR

git clone https://github.com/ONTHEIT-AI/BizOnAI-OCR.git
cd BizOnAI-OCR
uv venv && source .venv/bin/activate
# Recommended: install from lock file (exact tested versions)
uv pip install -r requirements.lock
# Or install from pyproject.toml (with benchmark support)
uv pip install -e ".[bench]"

Tested environment: Python 3.12, vllm==0.11.0, transformers==4.57.6, torch==2.8.0+cu128 (CUDA 12.8)

If model downloads fail with disk-space errors, set export HF_HOME=/path/to/large/disk/.cache/huggingface before running.

# Start vLLM server (uses GPU 0 by default, configurable via VLLM_GPUS)
bizonai-ocr-vllm
# In another terminal, process documents
bizonai-ocr input.pdf ./output

# No separate server needed, but slower for multiple files
bizonai-ocr input.pdf ./output --method hf

The pipeline auto-starts a vLLM server, processes all files, then shuts down.

# Single GPU
CUDA_VISIBLE_DEVICES=0 python -m bizonai_ocr.pipeline ./output \
 --model ONTHEIT/BizOnAI-OCR \
 --input /path/to/pdfs \
 --benchmark \
 --max-num-seqs 32 \
 --max-concurrent 64 \
 -tp 1 \
 --port 8111
# Multi-GPU (data parallel - runs model replicas on each GPU)
CUDA_VISIBLE_DEVICES=0,1 python -m bizonai_ocr.pipeline ./output \
 --model ONTHEIT/BizOnAI-OCR \
 --input /path/to/pdfs \
 --benchmark \
 --max-num-seqs 32 \
 --max-concurrent 64 \
 -tp 1 \
 --data-parallel-size 2 \
 --port 8111
# Post-process raw output to markdown
python -m bizonai_ocr.postprocess ./output --format markdown --output ./output_md

Pipeline options:

Settings via environment variables or local.env file:

# Model
MODEL_CHECKPOINT=ONTHEIT/BizOnAI-OCR # HuggingFace repo or local path
MAX_OUTPUT_TOKENS=12384
# vLLM server
VLLM_API_BASE=http://localhost:8000/v1
VLLM_MODEL_NAME=bizonai-ocr
VLLM_GPUS=0 # Which GPU(s) to use

Existing OCR benchmarks are primarily English-focused, making it difficult to evaluate model performance on Korean documents. KDoc-OCRBench addresses this gap with 14,738 test cases specifically designed for Korean industrial documents.

• 14,738 test cases across 804 Korean PDFs in 7 document categories
• 3 test types: text presence/absence, table structure validation, header/footer detection
• Unit-test-style evaluation based on olmOCR-bench methodology
• Korean documents collected and labeled by ONTHEIT
• Bootstrap confidence intervals for statistical rigor

See the KDoc-OCRBench dataset card for detailed examples of each category.

# 1. Install benchmark deps
uv pip install -e ".[bench]"
# 2. Download PDFs and test cases from HuggingFace
huggingface-cli download ONTHEIT/KDoc-OCRBench --repo-type dataset --local-dir kdoc_benchmark/data
# 3. Run your model (produces raw output)
CUDA_VISIBLE_DEVICES=0 python -m bizonai_ocr.pipeline ./kdoc_benchmark/data/my_model \
 --model ONTHEIT/BizOnAI-OCR \
 --input ./kdoc_benchmark/data/pdfs \
 --benchmark -tp 1
# 4. Post-process to markdown
python -m bizonai_ocr.postprocess ./kdoc_benchmark/data/my_model \
 --format markdown --output ./kdoc_benchmark/data/my_model_md
# 5. Evaluate
python -m kdoc_benchmark --dir ./kdoc_benchmark/data --candidate my_model_md \
 --table-normalize-spaces --table-normalize-newlines

See kdoc_benchmark/README.md for more details.

KDoc-OCRBench-V2 — A larger, manually-reviewed second iteration with 850 public-sector PDFs and 57,062 verified tests (with much heavier table coverage) is available as a separate, complementary benchmark. See kdoc_benchmark_v2/ and the dataset card at ONTHEIT/KDoc-OCRBench-V2.

• Qwen3-VL by Alibaba
• olmocr by Allen AI
• olmOCR-bench by Allen AI
• Chandra by Datalab
• vLLM

This repository is released under the Apache2.0 and OpenRAIL. See the LICENSE file for details.