baconnier/deepsynth-en-legal

DeepSynth - BillSum Legal Document Summarization

Dataset Description

US Congressional bills with human-written summaries. Specialized for legal document summarization with complex structure, formal language, and legislative terminology.

This dataset is part of the DeepSynth project, which uses visual text encoding for multilingual summarization with the DeepSeek-OCR vision-language model. Text documents are converted into images and processed through a frozen 380M parameter visual encoder, enabling 20x token compression while preserving document layout and structure.

Key Features

• Original High-Quality Images: Full-resolution images stored once, augmented on-the-fly during training
• Random Augmentation Pipeline: Rotation, perspective, color jitter, and resize transforms for better generalization
• Visual Text Encoding: 20x compression ratio (1 visual token ≈ 20 text tokens)
• Document Structure Preservation: Layout and formatting maintained through image representation
• Human-Written Summaries: High-quality reference summaries for each document
• Deduplication Tracking: Source dataset and index tracking prevents duplicates

Dataset Statistics

• Total Samples: ~22,000
• Language(s): English
• Domain: US Congressional bills
• Average Document Length: ~3,000 tokens
• Average Summary Length: ~200 tokens

Source Dataset

Based on the BillSum dataset of US Congressional bills.

• Original Authors: Kornilova & Eidelman (2019)
• Paper: BillSum: A Corpus for Automatic Summarization of US Legislation
• License: CC0 1.0 Universal (Public Domain)

Image Augmentation Pipeline

Images are stored at original resolution (up to 1600×2200) and augmented during training for better generalization:

Available Augmentation Transforms

• Random Rotation: ±10° rotation for orientation invariance
• Random Perspective: 0.1-0.2 distortion to simulate viewing angles
• Random Resize: 512-1600px range for multi-scale learning
• Color Jitter: Brightness, contrast, saturation adjustments (±20%)
• Random Horizontal Flip: Optional (use with caution for text)

All transforms preserve aspect ratio with padding to maintain text readability. This approach:

• Reduces storage: 6x less disk space (single image vs 6 resolutions)
• Increases flexibility: Any resolution on-the-fly vs pre-computed fixed sizes
• Improves generalization: Random transforms prevent overfitting to specific resolutions

Dataset Structure

Data Fields

• text (string): Original document text
• summary (string): Human-written summary
• image (PIL.Image): Original full-size rendered document image (up to 1600×2200)
• source_dataset (string): Origin dataset name
• original_split (string): Source split (train/validation/test)
• original_index (int): Original sample index for deduplication

Data Example

{
    'text': 'A BILL to amend the Internal Revenue Code of 1986...',
    'summary': 'This bill amends the Internal Revenue Code to...',
    'image': <PIL.Image>,  # Original resolution (up to 1600×2200)
    'source_dataset': 'billsum',
    'original_split': 'train',
    'original_index': 0
}

Usage

Loading the Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("baconnier/deepsynth-en-legal")

# Streaming for large datasets
dataset = load_dataset("baconnier/deepsynth-en-legal", streaming=True)

Training Example with DeepSeek-OCR and Augmentation

from transformers import AutoProcessor, AutoModelForVision2Seq
from datasets import load_dataset
from deepsynth.data.transforms import create_training_transform

# Load model and processor
model = AutoModelForVision2Seq.from_pretrained("deepseek-ai/DeepSeek-OCR")
processor = AutoProcessor.from_pretrained("deepseek-ai/DeepSeek-OCR")

# Load dataset
dataset = load_dataset("baconnier/deepsynth-en-legal")

# Create augmentation pipeline (random rotation, perspective, resize, color jitter)
transform = create_training_transform(
    target_size_range=(512, 1600),  # Random resize range
    rotation_degrees=10,             # ±10° rotation
    perspective_distortion=0.1,      # Perspective transform
    brightness_factor=0.2,           # ±20% brightness
    contrast_factor=0.2,             # ±20% contrast
)

# Process sample with augmentation
sample = dataset['train'][0]
augmented_image = transform(sample['image'])  # Apply random transforms
inputs = processor(
    images=augmented_image,
    text=sample['text'],
    return_tensors="pt"
)

# Fine-tune decoder only (freeze encoder)
for param in model.encoder.parameters():
    param.requires_grad = False

# Training loop with on-the-fly augmentation...

Training Recommendations

DeepSeek-OCR Fine-Tuning

# Recommended hyperparameters with augmentation
training_args = {
    "learning_rate": 2e-5,
    "batch_size": 4,
    "gradient_accumulation_steps": 4,
    "num_epochs": 3,
    "mixed_precision": "bf16",
    "freeze_encoder": True,  # IMPORTANT: Only fine-tune decoder

    # Augmentation parameters
    "rotation_degrees": 10,           # Random rotation ±10°
    "perspective_distortion": 0.1,    # Perspective transform
    "resize_range": (512, 1600),      # Random resize 512-1600px
    "brightness_factor": 0.2,         # ±20% brightness
    "contrast_factor": 0.2,           # ±20% contrast
}

Expected Performance

• Baseline (text-to-text): ROUGE-1 ~40-42
• DeepSeek-OCR (visual): ROUGE-1 ~44-47 (typical SOTA)
• Training Time: ~6-8 hours on A100 (80GB) for full dataset
• GPU Memory: ~40GB with batch_size=4, mixed_precision=bf16

Dataset Creation

This dataset was created using the DeepSynth pipeline:

1. Source Loading: Original text documents from billsum
2. Text-to-Image Conversion: Documents rendered as PNG images (DejaVu Sans 12pt, Unicode support)
3. Original Resolution Storage: Full-quality images stored once (up to 1600×2200)
4. Incremental Upload: Batches of 5,000 samples uploaded to HuggingFace Hub
5. Deduplication: Source tracking prevents duplicate samples

Note: Images are augmented on-the-fly during training using random transformations (rotation, perspective, resize, color jitter) for better generalization across different resolutions and conditions.

Rendering Specifications

• Font: DejaVu Sans 12pt (full Unicode support for multilingual text)
• Line Wrapping: 100 characters per line
• Margin: 40px
• Background: White (255, 255, 255)
• Text Color: Black (0, 0, 0)
• Format: PNG with lossless compression

Citation

If you use this dataset in your research, please cite:

@misc{deepsynth-en-legal,
    title={{DeepSynth BillSum Legal Document Summarization: Visual Text Encoding with Random Augmentation for Summarization}},
    author={Baconnier},
    year={2025},
    publisher={HuggingFace},
    howpublished={\url{https://huggingface.co/datasets/baconnier/deepsynth-en-legal}}
}

Source Dataset Citation

@inproceedings{kornilova2019billsum,
    title={BillSum: A Corpus for Automatic Summarization of US Legislation},
    author={Kornilova, Anastassia and Eidelman, Vladimir},
    booktitle={Proceedings of the 2nd Workshop on New Frontiers in Summarization},
    year={2019}
}

License

CC0 1.0 Universal (Public Domain Dedication)

Note: This dataset inherits the license from the original source dataset. Please review the source license before commercial use.

Limitations and Bias

• Legal jargon: Heavy use of legislative and legal terminology
• Complex structure: Bills have nested sections, subsections, clauses
• US-specific: United States federal legislation only
• Formal language: Very different from conversational or news text
• Long documents: Bills can be 10,000+ tokens

Additional Information

Dataset Curators

Created by the DeepSynth team as part of multilingual visual summarization research.

Contact

• Repository: DeepSynth GitHub
• Issues: GitHub Issues

Acknowledgments

• DeepSeek-OCR: Visual encoder from DeepSeek AI
• Source Dataset: billsum
• HuggingFace: Dataset hosting and infrastructure