---
tags:
- ocr
- document-processing
- deepseek
- deepseek-ocr-2
- markdown
- uv-script
- generated
configs:
- config_name: nuextract3
  data_files:
  - split: train
    path: nuextract3/train-*
dataset_info:
  config_name: nuextract3
  features:
  - name: image
    dtype: image
  - name: b_number
    dtype: string
  - name: page_index
    dtype: int64
  - name: source_row
    dtype: int64
  - name: markdown
    dtype: string
  - name: inference_info
    dtype: string
  splits:
  - name: train
    num_bytes: 20538673
    num_examples: 50
  download_size: 20343822
  dataset_size: 20538673
---

# Document OCR using DeepSeek-OCR-2

This dataset contains markdown-formatted OCR results from images in [davanstrien/moh-bench-sample](https://huggingface.co/datasets/davanstrien/moh-bench-sample) using DeepSeek-OCR-2.

## Processing Details

- **Source Dataset**: [davanstrien/moh-bench-sample](https://huggingface.co/datasets/davanstrien/moh-bench-sample)
- **Model**: [deepseek-ai/DeepSeek-OCR-2](https://huggingface.co/deepseek-ai/DeepSeek-OCR-2)
- **Number of Samples**: 50
- **Processing Time**: 4.4 min
- **Processing Date**: 2026-07-08 16:48 UTC

### Configuration

- **Image Column**: `image`
- **Output Column**: `markdown`
- **Dataset Split**: `train`
- **Batch Size**: 8
- **Max Model Length**: 8,192 tokens
- **Max Output Tokens**: 8,192
- **GPU Memory Utilization**: 80.0%

## Model Information

DeepSeek-OCR-2 is a 3B parameter vision-language model featuring Visual Causal
Flow architecture for more human-like visual encoding. Building on DeepSeek-OCR v1,
it offers enhanced document understanding with dynamic resolution up to
(0-6)x768x768 + 1x1024x1024 patches.

### Capabilities
- LaTeX equations - Mathematical formulas preserved in LaTeX format
- Tables - Extracted and formatted as HTML/markdown
- Document structure - Headers, lists, and formatting maintained
- Image grounding - Spatial layout and bounding box information
- Complex layouts - Multi-column and hierarchical structures
- Multilingual - Supports multiple languages

## Dataset Structure

The dataset contains all original columns plus:
- `markdown`: The extracted text in markdown format with preserved structure
- `inference_info`: JSON list tracking all OCR models applied to this dataset

## Usage

```python
from datasets import load_dataset
import json

# Load the dataset
dataset = load_dataset("{{output_dataset_id}}", split="train")

# Access the markdown text
for example in dataset:
    print(example["markdown"])
    break

# View all OCR models applied to this dataset
inference_info = json.loads(dataset[0]["inference_info"])
for info in inference_info:
    print(f"Column: {{info['column_name']}} - Model: {{info['model_id']}}")
```

## Reproduction

This dataset was generated using the [uv-scripts/ocr](https://huggingface.co/datasets/uv-scripts/ocr) DeepSeek-OCR-2 vLLM script:

```bash
uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/deepseek-ocr2-vllm.py \\
    davanstrien/moh-bench-sample \\
    <output-dataset> \\
    --image-column image
```

## Performance

- **Processing Speed**: ~0.2 images/second
- **Processing Method**: Batch processing with vLLM (2-3x speedup over sequential)

Generated with [UV Scripts](https://huggingface.co/uv-scripts)