alakxender/dhivehi-image-bbox-ds-fmt

Dhivehi Image Bounding Box Dataset - DeepSeek Format

Dataset Description

This dataset is a transformed version of alakxender/dhivehi-image-bbox-prompt, specifically formatted to align with DeepSeek OCR model requirements for training vision-language models with grounding capabilities.

The original dataset contained Dhivehi (Thaana script) text with bounding box annotations. This version restructures the annotations into DeepSeek's grounding token format, enabling the model to both recognize text and localize it spatially within images.

Dataset Structure

Fields

• image: PIL Image object containing the document/page image
• content: Text string in DeepSeek grounding format with special tokens

Format Specification

The content field follows DeepSeek's grounding format:

<|ref|>{category}<|/ref|><|det|>[[x1, y1, x2, y2]]<|/det|>{text_content}

Where:

• {category}: Element type (Title, Text, Caption, Picture, etc.)
• [x1, y1, x2, y2]: Normalized bounding box coordinates (0-999 scale)
• {text_content}: The actual text content (in Dhivehi/Thaana script)

Example

{
    "image": <PIL.Image>,
    "content": "<|ref|>Title<|/ref|><|det|>[[102, 48, 841, 51]]<|/det|>މާދަމާގެ އެއްވުމަށް ފުލުހުން ޝަރުތުތަކެއް ކަނޑައަޅައި\n<|ref|>Text<|/ref|><|det|>[[67, 177, 876, 208]]<|/det|>ސަރުކާރާ ދެކޮޅަށް އިދިކޮޅު މީހުން މާދަމާ މާލޭގައި..."
}

Transformation Details

Original → DeepSeek Format

Original format:

{
  "bbox": [x, y, width, height],
  "category": "Title",
  "text": "Sample text",
  "width": 800,
  "height": 1200
}

Transformed format:

• Bounding boxes converted from [x, y, w, h] to [x1, y1, x2, y2]
• Coordinates normalized to 0-999 scale (DeepSeek standard)
• Added grounding tokens: <|ref|>, <|/ref|>, <|det|>, <|/det|>
• Combined all annotations into a single content string
• Removed metadata fields, keeping only image and content

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("alakxender/dhivehi-image-bbox-ds-fmt")

# Access splits
train_data = dataset['train']
val_data = dataset['validation']

# Example: Print first sample
sample = train_data[0]
print(f"Content: {sample['content'][:200]}...")

Training with DeepSeek OCR

from transformers import AutoModel, AutoTokenizer

# Load DeepSeek OCR model
model = AutoModel.from_pretrained(
    "your-deepseek-ocr-model",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "your-deepseek-ocr-model",
    trust_remote_code=True
)

# Use the dataset for fine-tuning
# The content field is ready to use as training targets

Inference Format

When using a model trained on this dataset:

# Input prompt
prompt = "<image>\n<|grounding|>Extract all text with their locations. "

# Expected output format (same as content field)
output = model.infer(
    tokenizer=tokenizer,
    prompt=prompt,
    image_file='document.png'
)
# Output: "<|ref|>Title<|/ref|><|det|>[[...]]<|/det|>Text content..."

Dataset Categories

The dataset includes various document element types:

• Title: Document titles and headers
• Text: Body text paragraphs
• Caption: Image captions and labels
• Picture: Image regions (bounding boxes only)

Language

• Primary Language: Dhivehi (Maldivian)
• Script: Thaana (ތާނަ)
• Text Direction: Right-to-left (RTL)

Dataset Splits

• Train: Training set
• Validation: Validation/test set

Intended Use

Primary Use Cases

1. Document OCR with Layout Understanding: Train models to extract text while preserving spatial relationships
2. Grounding-aware Text Recognition: Enable models to localize text regions
3. Dhivehi Document Processing: Specialized for Thaana script recognition
4. Multi-modal Vision-Language Tasks: Train models that understand both visual and textual content

Out-of-Scope Use

• This dataset is specifically formatted for DeepSeek-style grounding models
• Not suitable for classification-only tasks
• Not intended for models that don't support grounding tokens