README.md

# dotsocr_markdown_dataset

## Dataset Description

This dataset contains training data for DotsOCR to convert document images directly to markdown format.

## Training Objective

The model learns to:
- Convert document images to clean markdown format
- Preserve document structure and hierarchy
- Extract all text content accurately
- Use appropriate markdown formatting for different content types

## Dataset Structure

- **Training samples**: 798
- **Validation samples**: 200
- **Total samples**: 998

## Files

- `data/dotsocr_markdown_dataset_train.json`: Training data in conversation format
- `data/dotsocr_markdown_dataset_val.json`: Validation data in conversation format
- `images/`: Directory containing all referenced images

## Data Format

Each sample follows this conversation format:

```json
{
  "conversations": [
    {
      "from": "user",
      "value": "<|imgpad|>images/example.jpg<|/imgpad|>\nPlease convert this document image to markdown format..."
    },
    {
      "from": "assistant",
      "value": "# Document Title\n\n## Section Header\n\nContent text here..."
    }
  ]
}
```

## Training Prompt

The training uses this prompt to teach the model markdown conversion:

```
Convert this document image to markdown format. Extract and structure the text content as follows:

- Headers and titles as plain text on separate lines
- Body text as plain text paragraphs
- Lists using * for bullet points
- Tables using HTML table format: <table><thead><tr><th>...</th></tr></thead><tbody><tr><td>...</td></tr></tbody></table>
- Preserve the original reading order and document structure
- Extract all visible text exactly as it appears

Output the content in markdown format with HTML tables for tabular data.
```

## Usage

This dataset is designed to train DotsOCR models for direct image-to-markdown conversion tasks. The model learns to:

1. Analyze document images
2. Extract text content in reading order
3. Apply appropriate markdown formatting
4. Output clean, structured markdown

## Citation

Based on the DotsOCR model by rednote-hilab: https://github.com/rednote-hilab/dots.ocr

Generated on: 2025-09-26 17:35:35