README.md

---
viewer: false
tags: [uv-script, ocr, vision-language-model, document-processing]
---

# OCR UV Scripts

> Part of [uv-scripts](https://huggingface.co/uv-scripts) - ready-to-run ML tools powered by UV

Ready-to-run OCR scripts that work with `uv run` - no setup required!

## 🚀 Quick Start with HuggingFace Jobs

Run OCR on any dataset without needing your own GPU:

```bash
# Quick test with 10 samples
hf jobs uv run --flavor l4x1 \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py \
    your-input-dataset your-output-dataset \
    --max-samples 10
```

That's it! The script will:

- ✅ Process first 10 images from your dataset
- ✅ Add OCR results as a new `markdown` column
- ✅ Push the results to a new dataset
- 📊 View results at: `https://huggingface.co/datasets/[your-output-dataset]`

## 📋 Available Scripts

### RolmOCR (`rolm-ocr.py`)

Fast general-purpose OCR using [reducto/RolmOCR](https://huggingface.co/reducto/RolmOCR) based on Qwen2.5-VL-7B:

- 🚀 **Fast extraction** - Optimized for speed and efficiency
- 📄 **Plain text output** - Clean, natural text representation
- 💪 **General-purpose** - Works well on various document types
- 🔥 **Large context** - Handles up to 16K tokens
- ⚡ **Batch optimized** - Efficient processing with vLLM

### Nanonets OCR (`nanonets-ocr.py`)

State-of-the-art document OCR using [nanonets/Nanonets-OCR-s](https://huggingface.co/nanonets/Nanonets-OCR-s) that handles:

- 📐 **LaTeX equations** - Mathematical formulas preserved
- 📊 **Tables** - Extracted as HTML format
- 📝 **Document structure** - Headers, lists, formatting maintained
- 🖼️ **Images** - Captions and descriptions included
- ☑️ **Forms** - Checkboxes rendered as ☐/☑

### SmolDocling (`smoldocling-ocr.py`)

Ultra-compact document understanding using [ds4sd/SmolDocling-256M-preview](https://huggingface.co/ds4sd/SmolDocling-256M-preview) with only 256M parameters:

- 🏷️ **DocTags format** - Efficient XML-like representation
- 💻 **Code blocks** - Preserves indentation and syntax
- 🔢 **Formulas** - Mathematical expressions with layout
- 📊 **Tables & charts** - Structured data extraction
- 📐 **Layout preservation** - Bounding boxes and spatial info
- ⚡ **Ultra-fast** - Tiny model size for quick inference

### NuMarkdown (`numarkdown-ocr.py`)

Advanced reasoning-based OCR using [numind/NuMarkdown-8B-Thinking](https://huggingface.co/numind/NuMarkdown-8B-Thinking) that analyzes documents before converting to markdown:

- 🧠 **Reasoning Process** - Thinks through document layout before generation
- 📊 **Complex Tables** - Superior table extraction and formatting
- 📐 **Mathematical Formulas** - Accurate LaTeX/math notation preservation
- 🔍 **Multi-column Layouts** - Handles complex document structures
- ✨ **Thinking Traces** - Optional inclusion of reasoning process with `--include-thinking`


## 🆕 New Features

### Multi-Model Comparison Support

All scripts now include `inference_info` tracking for comparing multiple OCR models:

```bash
# First model
uv run rolm-ocr.py my-dataset my-dataset --max-samples 100

# Second model (appends to same dataset)
uv run nanonets-ocr.py my-dataset my-dataset --max-samples 100

# View all models used
python -c "import json; from datasets import load_dataset; ds = load_dataset('my-dataset'); print(json.loads(ds[0]['inference_info']))"
```

### Random Sampling

Get representative samples with the new `--shuffle` flag:

```bash
# Random 50 samples instead of first 50
uv run rolm-ocr.py ordered-dataset output --max-samples 50 --shuffle

# Reproducible random sampling
uv run nanonets-ocr.py dataset output --max-samples 100 --shuffle --seed 42
```

### Automatic Dataset Cards

Every OCR run now generates comprehensive dataset documentation including:
- Model configuration and parameters
- Processing statistics
- Column descriptions
- Reproduction instructions

## 💻 Usage Examples

### Run on HuggingFace Jobs (Recommended)

No GPU? No problem! Run on HF infrastructure:

```bash
# Basic OCR job
hf jobs uv run --flavor l4x1 \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py \
    your-input-dataset your-output-dataset


# Real example with UFO dataset 🛸
hf jobs uv run \
    --flavor a10g-large \
    --image vllm/vllm-openai:latest \
    -s HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py \
    davanstrien/ufo-ColPali \
    your-username/ufo-ocr \
    --image-column image \
    --max-model-len 16384 \
    --batch-size 128

# NuMarkdown with reasoning traces for complex documents
hf jobs uv run \
    --image vllm/vllm-openai:latest \
    --flavor l4x4 \
    -s HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/numarkdown-ocr.py \
    your-input-dataset your-output-dataset \
    --max-samples 50 \
    --include-thinking \
    --shuffle

# Private dataset with custom settings
hf jobs uv run --flavor l40sx1 \
    -s HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py \
    private-input private-output \
    --private \
    --batch-size 32
```

### Python API

```python
from huggingface_hub import run_uv_job

job = run_uv_job(
    "https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py",
    args=["input-dataset", "output-dataset", "--batch-size", "16"],
    flavor="l4x1"
)
```

### Run Locally (Requires GPU)

```bash
# Clone and run
git clone https://huggingface.co/datasets/uv-scripts/ocr
cd ocr
uv run nanonets-ocr.py input-dataset output-dataset

# Or run directly from URL
uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/nanonets-ocr.py \
    input-dataset output-dataset

# RolmOCR for fast text extraction
uv run rolm-ocr.py documents extracted-text
uv run rolm-ocr.py images texts --shuffle --max-samples 100  # Random sample

```

## 📁 Works With

Any HuggingFace dataset containing images - documents, forms, receipts, books, handwriting.

## 🎛️ Configuration Options

### Common Options (All Scripts)

| Option                     | Default | Description                   |
| -------------------------- | ------- | ----------------------------- |
| `--image-column`           | `image` | Column containing images      |
| `--batch-size`             | `32`/`16`* | Images processed together     |
| `--max-model-len`          | `8192`/`16384`** | Max context length     |
| `--max-tokens`             | `4096`/`8192`** | Max output tokens      |
| `--gpu-memory-utilization` | `0.8`   | GPU memory usage (0.0-1.0)    |
| `--split`                  | `train` | Dataset split to process      |
| `--max-samples`            | None    | Limit samples (for testing)   |
| `--private`                | False   | Make output dataset private   |
| `--shuffle`                | False   | Shuffle dataset before processing |
| `--seed`                   | `42`    | Random seed for shuffling     |

*RolmOCR uses batch size 16
**RolmOCR uses 16384/8192

### RolmOCR Specific

- Output column is auto-generated from model name (e.g., `rolmocr_text`)
- Use `--output-column` to override the default name

💡 **Performance tip**: Increase batch size for faster processing (e.g., `--batch-size 128` for A10G GPUs)

More OCR VLM Scripts coming soon! Stay tuned for updates!